From Cargo Design to a Manufacturing- and Regulatory-Ready Package in One Deterministic Pipeline

An in-silico-first delivery engineering service unifying nucleic-acid and protein cargo design, vehicle design, physics-based structure prediction and a compendial CMC and regulatory layer inside one reproducible engine.

Descriptive Conventions Used in This Document

This whitepaper describes the platform as a single unified service rather than as separate software products. Every internal computational engine is referred to by its GC-series internal codename, and no GC-module is mapped to a specific external component anywhere in this document. Each module is a replaceable backend defined by an input, output and uncertainty contract, so swapping a backend or migrating it to an in-house model has no effect on the customer interface or on deliverable formats. The GC-series names are interfaces, and they are deliberately stable across editions.

Third-party components are named, by licence, in Appendix A. The codenames exist to keep module identity independent of any particular backend - not to conceal what the platform is built from. A customer conducting technical or legal due diligence needs to know the licence position of the stack, so Appendix A lists every external component together with its licence and its commercial-use status, verified against the licence text present on this system. What Appendix A deliberately does *not* do is state which component implements which GC-module; that mapping is what would freeze a replaceable backend into a published dependency.

Public standards - ICH, USP, Ph. Eur., EU GMP Annex 1, ISTA, ASTM - and public scientific theory are named explicitly, because these are the criteria against which the platform asks to be judged. Figures are labelled by source as [internal], [literature] or [verification required].

One convention is new in v4.0 and worth stating at the front, because it governs how every number in this document should be read. Where a count could be either recited from the previous edition or re-measured against the current code, it has been re-measured, and the measurement basis - the constant name and its line number, or the command that produced the figure - is given alongside it. Where the previous edition's counting rule could not be reconstructed, this document reports its own measurement and says that the delta is not computable, rather than presenting a difference between two numbers that were counted differently.

1. Overview - The Problem This Platform Solves

1.1 The Real Bottleneck in RNA Medicine Is the Vehicle, Not the Sequence

In RNA therapeutics, deciding what to express - or which gene to silence - is among the easiest parts of the programme. Once the target is chosen, sequence design takes days. Where programmes actually founder is the vehicle that must carry that sequence intact into the cytosol: the lipid nanoparticle (LNP) and the formulation around it.

The conditions a single ionizable lipid must satisfy simultaneously pull in opposite directions. It must be electrically near-neutral in blood, so that it does not bind serum proteins indiscriminately or lyse the erythrocyte membranes it brushes past, yet cationic inside the endosome, where a positive charge is exactly what is needed to destabilise the limiting membrane and release the payload. It must fully encapsulate a polyanionic payload, which demands strong electrostatic association, yet surrender it at the destination, which demands that the same association be reversible. It must overcome the default tendency of any intravenously administered particle above roughly 50 nm to be filtered into the liver, and it must survive manufacturing, filling, shipping and storage without aggregating, oxidising or hydrolysing.

These are not five independent specifications that can be optimised one at a time. They are five projections of one molecular design, and improving any one of them typically degrades another. Raising the ionizable lipid's apparent pKa improves endosomal protonation and escape, but pushes the particle toward being cationic in plasma, which increases opsonin binding and hepatic clearance. Increasing the PEG-lipid density improves colloidal and storage stability, but forms a steric barrier that suppresses the ApoE adsorption that hepatocyte uptake actually depends on, and raises the anti-PEG liability that governs whether the formulation survives repeat dosing. Adding unsaturation to the helper phospholipid improves fusogenicity and escape, but introduces the oxidisable allylic positions that dominate the stability budget.

The industry currently solves this multi-dimensional trade-off with at least three disconnected toolchains: a cheminformatics and machine-learning stack for lipid design, a molecular-dynamics environment for particle structure, and a patchwork of spreadsheets, LIMS entries and consultant reports for chemistry, manufacturing and controls (CMC). Every hand-off loses context, breaks traceability and re-introduces transcription error. Worse, because the three chains run at different times in different teams, a particle that is biologically excellent but unmanufacturable is discovered only months later - typically in a registration stability batch, where the remedy is a reformulation and the cost is a year.

This platform collapses that chain into one deterministic engine. Nucleic-acid cargo design, ionizable-lipid design, physics-based structure prediction of the assembled particle, in-vivo distribution prediction, and the CMC and regulatory layer actually needed to populate a filing all live inside one pipeline and share one checkpoint contract. The consequence for the scientist is a single ranked list rather than dozens of unreconciled reports - and the molecule that ranks first is also the molecule whose manufacturing and regulatory liabilities have already been enumerated.

1.2 What Has Changed in This Edition (v4.0)

The v3.0 whitepaper described the platform at 17,889 lines of pipeline code with 979 automated tests. The engine described here is 32,967 lines and collects 1,626 tests. That is close to a doubling of both figures in roughly two months, and the character of the growth is as important as its size: the additions are concentrated in the layers that convert a computed number into a defensible one.

Table 1-1. Scale change from v3.0 to v4.0. Every v4.0 figure was re-measured against the current code base; the measurement basis is given so a reviewer can reproduce the count rather than take it on trust.

Three changes dominate the character of this edition.

The registry became the product of the pipeline, not only its input. At v3.0 the platform ranked a catalogue of 62 hand-curated variants. It now also maintains a discovered-variant registry that the discovery loop writes into, currently holding 170 entries. Those 170 are not a wish list: each carries a resolved receptor, a ligand record with its head sequence and topology, an organ assignment and a cargo assignment. The distribution is informative about where the platform has actually been exercised - 128 lung, 30 hepatocyte, 7 brain, 2 liver non-parenchymal, 2 tumour, 1 liver - and that distribution is reported here rather than smoothed, because it is the honest map of which claims in this document rest on many runs and which rest on few.

The evidence layer grew faster than the physics layer. The larger part of the 15,000 new lines is not new physics but new adjudication: schema-versioned caches so that a fix actually reaches the cached consumers that depend on it, three-state gates so that an unmeasured axis is not silently scored as a passing one, and fingerprint checks so that a stale result cannot masquerade as a fresh one. This is unglamorous work and it is the work that decides whether a long computation can be trusted. Chapter 9 is the detail.

Honesty about latent capability. Several capabilities are wired and tested but not yet exercised by any live campaign. The clearest example: all 170 discovered variants carry siRNA cargo, so the circular self-amplifying RNA path, though implemented and unit-tested, has never run on a real registry entry. v3.0 would have described that path in the present tense. This edition marks it latent, and Chapter 10.1 lists every case.

1.3 Five Design Principles

Determinism. A tool whose numbers may enter a regulatory filing must give the same answer twice. Every stochastic step is seeded, and a re-run on identical inputs reproduces every number exactly.

Traceability. Every stage writes a JSON checkpoint to disk. Any value in the final ranked output can be followed back to the exact stage, inputs and module version that produced it. The purpose of this design is to let a reviewer audit a claim rather than trust it. v4.0 strengthens this in one specific way: a checkpoint now carries the schema version of the code that wrote it, so a cached value produced by superseded logic is recognised as stale instead of being re-read as current.

Additivity. No module in the quality layer gates the pipeline. They are additive risk lenses, not pass/fail filters. A particle with a moderate nitrosamine flag is not silently discarded; it remains ranked with that liability quantified, and a human decides the trade-off.

Graceful degradation. Where a heavy physics tool is unavailable, the pipeline does not stop; it marks that axis 'not adjudicated' and proceeds. Not-adjudicated is not the same as failed, and is handled by weight renormalisation in the ranking arithmetic. The distinction is load-bearing: conflating the two is the single most common way a screening pipeline silently converts missing evidence into a passing score.

Honesty. Every module output carries an explicit caveat string. The generated regulatory document flags attributes that only a measurement can settle as open gaps rather than quietly filling them with a computed proxy.

2. Service Scope - Cargo Modalities and the Vehicle Catalogue

The platform's scope is defined by the product of two axes: what is to be delivered (the cargo modality) and what carries it (the vehicle class). Industry practice optimises the two separately, but they do not in fact separate. A heavily chemically modified siRNA changes the lipid-composition requirement, because 2'-O-methyl and phosphorothioate substitution alter both the duplex's charge density and its rigidity, and therefore how tightly it packs against the ionizable lipid. A large self-amplifying RNA demands a completely different particle morphology from a short duplex, because a 10-kilobase single strand cannot be accommodated in the same internal architecture as a 21-mer duplex at any lipid ratio. The platform treats the pair as a joint optimisation problem.

2.1 Cargo Modalities (13 therapeutic + 2 structural = 15 registered types)

Table 2-1. The thirteen therapeutic cargo modalities. Each routes to a dedicated Stage 0 designer, assembly rule, morphology criterion and patent-claim subset. The final column states the biological mechanism behind the difficulty, because the design trade-off the platform computes is only meaningful if the reason for it is understood.

VALID_CARGO_TYPES (L6324) registers 15 entries: the thirteen above plus generic (no modality declared) and custom_rnp, the protein-payload hand-off slot described in 2.1.1 below. CARGO_DEFAULTS (L6519) carries exactly 15 matching keys, so the two registries agree. A sixteenth token, sgRNA, appears only as a sub-element of the CRISPR payload path rather than as an independent top-level modality.

The ssRNAi entry is the one genuinely new modality since v3.0. Its two variants are anchored differently on purpose: SR1 carries a C16 chain, following the unformulated C16-conjugated ss-siRNA dosed in the literature, and SR2 carries the C24 anchor used by the in-house conjugate platform. Holding the oligonucleotide constant and varying only the anchor length is what makes the anchor's contribution attributable rather than confounded.

This modality is, however, under repair. As section 14.2 records, the single-strand RNAi subsystem has 32 failing assertions, the cause is identified as a schema and duplication issue, and the resolution is pending. A reader intending to rely on ssRNAi should read that section first. The twelve other modalities are unaffected by this cluster.

Cargo families are derived, not re-listed. One class of defect is worth describing because the engine now structurally prevents it. The mRNA-like family - the cargo types that express a protein and therefore share storage, innate-immune and MHC-scan treatment - was originally written as four separate literal copies of {mRNA, saRNA, circRNA, multiplex}, all authored before csaRNA existed and none updated when it arrived. The measured consequence on the preceding release: csaRNA was assigned +5 °C storage instead of -20 °C across six storage-stability stages, was left out of the innate-immune double-stranded-RNA argument, and was recorded as "not protein-coding" by the MHC-I scan. The fix was to stop listing the family and derive it - _EXPRESSING_CARGOS (L6602) is computed as every cargo whose expected morphology is Bleb, which yields five members including csaRNA automatically - and to add a registry audit that reports any table naming *some* of a family but not all of it. This is the same failure mode as a stale index: a hand-maintained list of a set that already exists elsewhere will drift, and the drift is silent.

Where the platform has actually been exercised. Of the 170 entries in the discovered-variant registry, all 170 carry siRNA cargo. The other eleven modalities are implemented, unit-tested and reachable, but no live registry entry exercises them. This is stated here rather than in a footnote because it bears directly on how much weight a reader should place on modality-specific claims: the siRNA path is heavily exercised, and the remainder rest on unit tests and literature directionality [verification required].

2.1.1 Protein and Peptide Payloads

Two of the fifteen registered cargo types carry a folded protein, and they are not the same kind of entry. RNP is a fixed specification: SpCas9 at 160 kDa complexed with a guide RNA, its protein physicochemistry held constant in the registry. custom_rnp is a hand-off slot rather than a placeholder. Its protein is supplied per design by an upstream protein-design pipeline over the GeneCrafter.PayloadHandover/1 schema, and the formulation is then recomputed from that specific protein instead of being inherited from SpCas9. That is what makes a designed binder, nuclease or other engineered protein a formulable payload rather than an unsupported one.

What is recomputed, and why the protein moves it. payload_to_cargo_spec (L6654) takes the payload's molecular weight, isoelectric point, net charge at pH 7.4 and net charge per 100 residues, and returns a formulation specification rather than a score. Every quantity below moves because a protein is present, which is the reason a protein payload cannot simply reuse the siRNA defaults.

Recomputed quantity	Why the protein changes it
N/P ratio	A cationic protein condenses part of the nucleic acid itself, so the effective phosphate count falls and less ionizable lipid is needed to reach the same charge balance
Mixing pH	Set one unit below the isoelectric point and clamped to 3.5-5.5, so the protein stays net positive during assembly without being taken to a pH that unfolds it
Ionizable-lipid pKa	Mapped from net charge per 100 residues across 6.2-6.8; a cationic payload needs less help from the lipid to hold the complex together
Payload hydrodynamic radius	Estimated from molecular weight for a globular fold, which sets the encapsulation volume the particle has to accommodate
Minimum core diameter	Derived from the protein volume plus the guide-RNA volume, because a particle cannot be specified smaller than its own contents
Ionic strength	Scaled with absolute net charge across 50-300 mM, since more surface charge needs more counter-ion screening to stay colloidally stable

Table 2-1b. The protein-payload specification recomputed per design for the custom_rnp route. These values are a formulation recipe, not a ranking: each one is an instruction to the downstream build stages.

The target hydrodynamic diameter is then taken as the larger of the registry default and the computed minimum core diameter plus 10 nm, so a large designed protein widens the particle specification instead of silently violating it. apply_payload_handover merges the result into CARGO_DEFAULTS['custom_rnp'] and records it per payload identifier, so the downstream coarse-grained build and the N:P stoichiometry see the per-protein values rather than the SpCas9 ones. A missing or unreadable hand-off file is a graceful no-op rather than a silent default: the run continues on the registry specification and records that it did.

Peptide is deliberately not a cargo modality, and the distinction is worth stating rather than leaving to inference. Peptides enter the platform one layer further out, as the targeting head on the particle surface. The peptide entries in the 29-ligand catalogue are built as coarse-grained Martini 3 topologies by a rule-based generator, de-novo sequences have their fold predicted because they have no homologs, and a peptide head is co-folded against its receptor binding domain to yield an interface confidence before it is accepted. A peptide therapeutic would require its own Stage 0 designer, assembly rule, morphology criterion and patent-claim subset, and the registry does not claim one. A reader who sees peptide handling in the platform should therefore read it as targeting, not as a peptide drug modality.

2.2 Vehicle Classes (5 classes / 17 distinct subtypes)

Table 2-2. The vehicle catalogue. VEHICLE_SUBCLASS (L6290) holds 20 variant-to-subtype mappings resolving to 17 distinct subtypes; the collisions are deliberate, since two variants may share a subtype while differing in composition or cargo.

Three of the five classes - exosome, polymer and VLP - are grouped as _NON_LNP_VEHICLE_CLASSES (L6428) precisely because they have no lipid-nanoparticle molecular-dynamics path. That grouping exists because of a measured failure rather than as a taxonomy exercise: an exosome variant carrying mRNA was routed to the mRNA-LNP coarse-grained builder, which then asked it for an ionizable lipid it does not contain. Naming the set explicitly is what makes the builder dispatch checkable.

The variant roster is larger than any static list. ALL_VARIANTS (L6313) concatenates thirteen family lists, giving 73 catalogued variants: 56 cargo-and-vehicle variants plus 17 organ-targeting composition variants. On top of that the pipeline loads a runtime registry from discovered_variants.json, currently 170 entries, and a tissue-combination generator. The live roster is therefore not statically determinable - a property worth stating plainly, because a reader comparing this document's counts against a running instance will see a larger number and should know why.

Organ targeting by composition, not by ligand. The 17 organ-targeting variants implement selective organ targeting, in which a fifth charged lipid redirects tropism with no targeting ligand at all. The charge sign is the mechanism, and the registry records it: the lung set is permanently cationic, the spleen set is anionic (phosphatidic acid, phosphatidylserine, phosphatidylglycerol), and the central-nervous-system set is again permanently cationic. The biology behind this is corona composition rather than receptor recognition - a cationic surface recruits a different plasma-protein layer than an anionic one, and it is that adsorbed layer, not the lipid, that the endothelium reads.

One detail in this area is a good illustration of how easily a screen can fool itself. Several organ-targeting variants are chemically identical to each other - two lung variants share one cationic lipid, two spleen variants share one anionic lipid, two CNS variants share one permanent cation - and stages 1, 2 and 8 key on the lipid structure alone. Left alone, those pairs would receive byte-identical scores and appear as independent confirmations of the same result. _CHEM_DUPLICATE_GROUPS (L6473) records the groupings so that a duplicate is recognised as a duplicate rather than counted twice.

2.3 The Targeting Layer (29 ligands)

LIGAND_CATALOG (L7249) holds 29 entries, each carrying a receptor, an organ, a reference dissociation constant, a linker and a literature citation; roughly twenty carry an explicit structure, while several antibody and Fab heads use a peptide-proxy placeholder structure, which is marked rather than hidden.

The catalogue's design reveals a quantitative trap worth stating, because getting it wrong inflates predicted potency by orders of magnitude. The canonical hepatocyte ligand is a tri-antennary sugar cluster binding the asialoglycoprotein receptor at roughly 2.5 nM. That figure belongs to the trivalent cluster; a single sugar unit binds the same receptor at roughly 1 mM - about six orders of magnitude weaker. The avidity model must therefore construct the trivalent number *from* the monovalent one, because a model that starts from the trivalent constant and then applies a multivalency enhancement counts the cluster effect twice. The catalogue stores both constants side by side so the derivation cannot silently start from the wrong end. An earlier revision of this entry additionally carried the wrong sugar - the C4 epimer, which mammalian receptors do not bind at all, the binding lectins being avian - so structure and affinity parameter disagreed by six orders of magnitude while looking individually plausible.

Ligand or composition - the platform ranks the choice rather than assuming it. For a hepatocyte target both routes exist: passive uptake, where a conventional ionizable lipid adsorbs apolipoprotein E and enters through the low-density-lipoprotein receptor, and active targeting through the sugar-receptor route. The mouse-hepatocyte variants deliberately take the passive route and explicitly record the sugar decoration as an intentional None rather than by omission. The reason is attributability: decorating the same particle with both mechanisms would give two independent hepatocyte uptake routes in one experiment and make neither one measurable. This is the same discipline as the single-anchor-variable in the ssRNAi pair.

That variant family also exists because of a coverage hole the audit found rather than because someone requested it. The catalogue contained 35 hepatocyte variants and not one of them was the combination (LNP carrier, siRNA cargo): every liver siRNA entry was a conjugate or polymer or exosome, and every liver LNP entry carried mRNA, ASO, saRNA or circRNA. The single most obvious question in the field - a conventional siRNA LNP for the liver - had no variant that could express it, so there was nothing to run. A catalogue can be large and still have a hole exactly where the common case sits.

2.4 Species and Route Coverage

Route selection is not cosmetic: it changes the target window of the central physicochemical parameter. The apparent-pKa optimum is stored per route (L10656) - intravenous 6.2 to 6.5, intramuscular 6.6 to 6.9, subcutaneous 6.4 to 6.8, inhalation 6.6 to 7.0, and intrathecal, intraocular and intratumoral 6.4 to 6.9. Those windows differ because the pH and protein environment the particle first meets differ, and because the balance between "stay neutral in circulation" and "protonate on arrival" shifts when circulation is short or absent. A lipid optimal for an intravenous therapeutic is therefore measurably suboptimal for an intramuscular vaccine, and the platform scores it against the window for the route actually declared.

Cross-species translation is handled explicitly in the pharmacokinetic layer rather than assumed. This matters because the composition-based tropism model is trained predominantly on rodent in-vivo observation, and carriage to primate and human is a translation by route-dependent literature factors rather than a retrained human model [literature].

3. Platform Architecture - Ten Parent Stages and 76 Half-Stages

3.1 Three Dispatch Tables, Not One

The pipeline's control flow is worth describing precisely, because its shape is what makes the quality layer additive rather than gating.

Stage 9, the roster-level ranking stage, is deliberately not in STAGE_FNS. It is special-cased in the orchestrator because it is the only stage that operates across variants rather than within one, and therefore cannot be scheduled by the same per-variant loop.

Checkpoint identifiers follow a stated convention: a half-stage's id is its parent's id times ten, plus five - so a hook on stage 3 is 135, a hook on stage 7 is 175. The offset exists so a half-stage id can never collide with a parent's integer key. Later hooks were numbered by extending the same block rather than by strictly reapplying the formula, so the ids run 105 through 186 with gaps.

3.2 The Ten Parent Stages

Table 3-1. The ten parent stages. Only stages 2 and 8 gate; every other stage contributes additively.

3.3 Where the 76 Half-Stages Sit

The distribution is the single most informative fact about what this platform actually is.

Table 3-2. Half-stage distribution (76 total; counted from POST_STAGE_HOOKS per parent). That 42 of them hang off the intellectual-property stage - and only 4 off the three physics stages combined - is the honest shape of the platform: it is a decision and quality engine with a physics layer attached, not a molecular-dynamics package with paperwork bolted on.

3.4 The Checkpoint Contract and Resume Semantics

Every stage writes one JSON file per variant, named for the stage and stored under that variant's directory. The write stamps a UTC timestamp, and one detail of that write is deliberate: when a checkpoint is annotated in place rather than recomputed, the original timestamp is preserved, because advancing it would make the annotated checkpoint look newer than the results derived from it and thereby invert the provenance order.

Resume eligibility is a whitelist, not a blacklist: only ok and skipped permit reuse. Everything else - partial, stub, pending, error, and any status the code does not recognise - forces re-execution. A whitelist is the correct choice here because an unrecognised status is far more likely to mean a new failure mode than a new success.

One consequence of getting this wrong is recorded in the code. A dry-run mode once wrote its own checkpoints, and 2,030 of those dry_run records both polluted the status census and destroyed real results when the resume branch read them back as legitimate. Dry runs now return their result without writing.

3.5 Cache Invalidation - The Layer That Decides Whether a Fix Actually Runs

This is the part of the architecture that most directly determines whether the numbers in a report were produced by the code that claims to have produced them, and it deserves its own treatment because two consecutive releases got it wrong in instructive ways.

A stage's cache key is a fingerprint composed of a fingerprint-format version, the stage and cargo identity, a content hash of each external wrapper script the stage actually executes, a per-variant hash of that variant's own registry rows, a hash of the input sequence rather than the gene symbol, and a per-stage reference-data version. Scoping the registry hash per variant is deliberate: correcting one variant's specification should not invalidate every other variant's cached work.

The first failure. Five stages repaired in one release carried no fingerprint at all. With no fingerprint, the freshness check cannot fail, so 860 checkpoints written by the previous release were re-read verbatim - and the release's repaired code would never have executed on a single variant. Nothing errored; the run simply reported the old numbers.

The second failure, in the very next release, is the more interesting one. The fix reached five of its six target stages. It missed the sixth because that stage's changed logic was an in-module function rather than an external wrapper script, and the fingerprint only hashed wrappers. The evidence was unambiguous once looked for: all six target variants held a checkpoint marked ok, with no fingerprint key, carrying the superseded value where the current code computes a different one.

The generalisation the code now states is worth quoting as a principle, because it applies to any staged pipeline with a cache: the question "which cache must I invalidate" is answered by tracing the changed symbols to their cached consumers, not by listing the stages whose bodies were edited. A shared constant or an in-module helper has no stage of its own, and counting edited stages will miss it every time.

The current release extends this again. An activity model reaches one stage through four hops of ordinary Python imports, none of which any fingerprint covered; a hash of that model is now part of the key. The consequence is stated in the code rather than hidden: 184 checkpoints re-run and their numbers will move. Publishing that a fix changes results is the opposite of reassuring, and it is the only honest way to report it.

3.6 A Structural Invariant Enforced at Import

One safety property is enforced by making the program refuse to start. The set of stages that carry a fingerprint and the set of stages classified as expensive must be disjoint, and a non-empty intersection raises an error at import time.

The reason is specific. When a fingerprint mismatch invalidates a cache, the invalidation branch deletes and recomputes without checking whether the work being discarded is cheap or expensive. That is safe only because no expensive stage carries a fingerprint. If a fingerprint were ever added to the trajectory-analysis stage, any edit to the engine file would begin silently discarding multi-day simulation results. Rather than documenting that hazard in a comment, the code asserts the invariant and fails loudly if a future change breaks it.

3.7 The Parallelism Model

Three execution modes coexist. A thread pool handles concurrency within the production-dynamics stage. A process pool handles variant-level concurrency, but only for the six stages explicitly marked parallel-safe - payload design, lipid screen, physicochemistry, trajectory analysis, corona, and the intellectual-property filter. Everything else runs sequentially. Box packing runs once globally after the build stage rather than per variant, because packing decisions interact.

The whitelist matters more than it appears to. A stage is parallel-safe only if it neither mutates shared state nor contends for a single accelerator, and the default for an unlisted stage is sequential execution. Defaulting to safe and opting in explicitly is the right polarity for a pipeline whose expensive stages write multi-gigabyte trajectories.

4. Internal Algorithm Module Reference (GC-Series, 37 modules)

This chapter enumerates the platform's computational engines as functional contracts. Each entry is defined by what it consumes and what it produces; the implementing backend is not stated. This is an architectural fact rather than marketing concealment - modules are designed to be replaceable so long as the contract holds, and several have in fact changed backend during development without any change to upstream or downstream code.

The GC-series vocabulary is unchanged from v3.0, deliberately. A customer-facing module name is an interface, and renaming interfaces between editions would break every document, protocol and report that cites them. What has changed is the half-stage machinery behind several modules; the "backing" column states where each now runs.

4.1 Cargo and Payload Design Family (5)

New since v3.0 in this family: an expression-control hook (172) and a joint shadow optimisation hook (171), the latter being the stage whose numbers move in this release because its activity model is now covered by the cache key.

4.2 Lipid and Formulation Design Family (10)

New backing in this family: a lipid-conformer hook (115), a lipid phase-behaviour hook (157) and a metal-catalysed degradation hook (159).

4.3 Structure and Physics Family (7)

4.4 Biology and Pharmacokinetics Family (5)

New backing: a delivery-barriers hook (186) that makes the physical barriers between injection site and cytosol explicit rather than implicit in a single bioavailability term.

4.5 Quality, CMC and Regulatory Family (3)

4.6 Decision, Intellectual Property and Delivery Family (6)

4.7 Integration Family (1)

GC-BRIDGE receives a payload settled upstream by another design pipeline - an editing-enzyme ribonucleoprotein, an antibody, a de-novo protein - and reformulates it as a delivery problem. The inbound contract is a named, versioned schema rather than an ad-hoc dictionary, which is what allows the upstream pipeline to evolve without breaking this one. Through this bridge the payload's own developability score and anti-drug-antibody risk feed directly into vehicle ranking, and in particular the effect whereby a strongly cationic lipid acts as an adjuvant and amplifies that anti-drug-antibody risk is computed quantitatively rather than noted as a caveat.

5. Biophysical and Biological Rationale - What the Numbers Actually Mean

A predicted number is only useful if the reader knows what physical quantity it estimates, what approximation produced it, and what biology makes it matter. This chapter is therefore longer than a feature list would be. Each section states the mechanism, then the computation, then the honest limit.

5.1 Apparent pKa - The Single Switch the Whole Modern Ionizable Lipid Rests On

The biology. An ionizable lipid must be two different molecules at two moments. In blood at pH 7.4 it should be essentially uncharged, because a cationic particle surface binds complement and immunoglobulin, activates the complement cascade, and is cleared within minutes; it also lyses erythrocyte membranes on contact. Inside the endosome, where a proton pump drives the lumen from pH 6.5 down through 5.5 toward 5.0, the same lipid must become substantially cationic - because protonation is what lets it pair with the anionic phospholipids of the endosomal membrane, form the inverted-hexagonal contacts that destabilise the bilayer, and release the payload into the cytosol. The molecule's job is to sense a pH difference of roughly two units and change its charge state across it.

That is why the apparent pKa is the dominant single design parameter. It is the pH at which half of the ionizable groups at the assembled particle surface are protonated, and it sets where on the endosomal acidification trajectory the switch flips. Too low and the lipid never protonates enough to escape; the particle is taken up, traverses to the lysosome, and is degraded with its payload intact but irrelevant. Too high and the lipid is already cationic in plasma, and the particle never survives long enough to be taken up at all.

Why the assembled-particle value is not the molecular value. Measuring or computing the pKa of an isolated lipid molecule in dilute solution answers the wrong question. On a particle surface the ionizable groups are packed at high density, and each protonation raises the local surface potential, which makes the next protonation harder. The effect is large - typically more than a full pH unit - and it is systematic rather than noise. The platform therefore treats the apparent pKa as a self-consistent electrostatics problem: the surface potential shifts the pKa, the shifted pKa changes the degree of protonation, and the changed protonation changes the surface potential. The governing relation is the surface-potential correction to the intrinsic value, pKa_app = pKa_intrinsic - F*psi/(2.303*R*T), solved to self-consistency rather than applied once.

The intrinsic value itself comes from a structure-based calculation when the supporting tool is available, and otherwise from a descriptor regression of the form pKa ~ 6.5 + 0.10*logP - 0.01*TPSA. Which route was used is recorded in the output alongside the value, because the two carry very different uncertainty, and a reader who cannot tell them apart cannot weight them correctly.

The route dependence. The target window is not a single number but a per-route band (L10656): intravenous 6.2-6.5, intramuscular 6.6-6.9, subcutaneous 6.4-6.8, inhalation 6.6-7.0, and the intrathecal, intraocular and intratumoral routes 6.4-6.9. The intravenous window is the lowest and narrowest because an intravenous particle must survive a full circulatory transit before it is taken up, so the penalty for being cationic in plasma is at its maximum. An intramuscular vaccine particle is injected into tissue and taken up locally within minutes; being slightly cationic costs it much less, and the higher window buys better escape. Scoring a lipid against the wrong route's window is therefore not a rounding error - it can invert the ranking.

The honest limit, stated because it is measurable. The engine also applies a hard sanity gate that fails any candidate whose predicted pKa falls outside 5.0 to 8.0. Measured against the current library, whose values span roughly 6.44 to 6.86, this gate passes every candidate and has zero discriminating power. It is retained as a guard against a grossly wrong prediction, not as a screen, and it should not be cited as evidence that candidates were filtered on pKa. The discriminating work is done by the continuous window-fitness term, which decays linearly to zero one full pH unit outside the route's band.

5.2 Encapsulation Efficiency and Internal Morphology - Two Values Normally Visible Only After Synthesis

The biology. Encapsulation efficiency is the fraction of payload sequestered inside the particle rather than adsorbed on its surface. Surface-associated RNA is not merely wasted: it is exposed to serum nucleases, it is available to pattern-recognition receptors that sense extracellular nucleic acid, and it contributes to the measured payload concentration without contributing to delivered dose. A formulation with 95 percent encapsulation and one with 70 percent may assay identically for total RNA and differ severalfold in potency.

The computation. Rather than infer encapsulation from a binding assay, the platform measures it geometrically on the simulated particle: the fraction of payload beads lying within 5 Å of solvent is the exposed fraction, and encapsulation is its complement. This is a definition rather than a correlation, which is its strength - it does not depend on a dye's accessibility or a quenching efficiency - and also its limit, since the 5 Å threshold is a modelling choice that sets where "surface" begins.

Getting this to compute at all required a specific engineering decision worth naming, because it is the difference between a feasible analysis and an impossible one. A naive all-pairs distance calculation between payload and solvent beads on a large particle allocates on the order of hundreds of gigabytes per frame. The analysis therefore uses a spatially capped neighbour search, which returns only pairs within the cutoff and never materialises the full matrix.

Internal morphology. The platform also classifies internal structure, and the class matters biologically. An electron-dense internal substructure - the bleb morphology - correlates in cryo-electron microscopy with higher expression, and the engine uses that morphology as the expected class for exactly the cargo types that express a protein. That coupling is now derived from the morphology expectation rather than from a hand-maintained list of cargo names, which is what fixed the family-drift defect described in Chapter 2.

Supporting observables are computed from the same trajectories: bilayer thickness and lamellarity; a chain order parameter taken as P2 = (3*cos^2(theta) - 1)/2 over coarse-grained tail bond vectors measured against the local radial normal rather than a global axis, which is the correct reference for a curved surface and is then converted to the deuterium-order convention; lateral diffusion by an Einstein fit to the mean-squared displacement over the middle portion of the curve, avoiding both the ballistic onset and the poorly sampled tail; payload solvent-accessible area by a Shrake-Rupley construction on a Fibonacci sphere; and zeta potential from the charge distribution.

One detail in the zeta calculation is a good example of a trap the code documents rather than hides. The coarse-grained force field runs with a reaction-field permittivity of 15 - a modelling parameter of the simulation, not a property of water. The analytic electrostatics that converts charge to potential must use the real relative permittivity of water, about 78. Silently reusing the simulation's 15 would inflate every computed potential by a factor of roughly five, and the two constants are explicitly distinguished in the code for that reason.

5.3 Cargo Size Is a First-Class Physical Variable

A point that distinguishes a delivery model from a lipid model: the same lipid composition does not make the same particle around different payloads. The engine scales the simulation box with payload length as box_nm = min(120, max(30, 30*(cargo_nt/1500)^0.42)), bounded between 30 and 120 nm.

The sub-linear exponent is the physically meaningful part. A flexible polyanion in a condensed particle does not occupy volume proportional to its length; it occupies volume closer to a power of roughly one-half, because it coils and because the condensing lipid reduces its effective persistence length. An exponent of 0.42 encodes that a tenfold longer transcript needs a box roughly 2.6 times larger, not ten times. Using a linear scale would waste two orders of magnitude of simulation volume on solvent; using a constant box would clip a large transcript against its own periodic image and produce a spurious morphology.

The charge-ratio requirement scales with cargo class for the same underlying reason, and the registry records it per cargo: short oligonucleotide classes use a nitrogen-to-phosphate ratio of 1, conventional mRNA and circular RNA and multiplex payloads use 6, self-amplifying RNA and plasmid DNA use 8, and circular self-amplifying RNA - the longest and most structured cargo - uses 9. A longer, more highly charged payload needs proportionally more ionizable nitrogen to condense, and under-supplying it is what leaves payload on the surface.

5.4 The Free-Energy Battery - The Four Events That Govern Potency

Delivery succeeds or fails at four molecular events, and each is a free-energy difference rather than a rate that can be read off a structure. The platform computes all four by biased sampling along a reaction coordinate.

Table 5-1. The four legs of the free-energy battery. The receptor leg is explicitly skipped when no receptor is present in the box - an uncomputed value is never filled in as zero.

The method and its parameters. Each leg is an umbrella-sampling calculation: the system is restrained at a series of positions along the coordinate, and the biased distributions are combined into a free-energy profile. The main battery uses eight windows from 2.0 to 4.1 nm at uniform 0.3 nm spacing, with restraint stiffnesses chosen per leg - 1000 kJ/mol/nm² for endosomal insertion, 1500 for receptor association, 2000 for RNA-lipid separation. The stiffnesses differ because the underlying profiles differ in steepness: a stiffer restraint is needed where the unbiased force is larger, or the window's sampled distribution will not overlap its neighbours. The membrane-insertion battery uses a finer and longer coordinate - 24 windows from 0 to 6 nm at 0.25 nm - because it must resolve both the interfacial adsorption minimum and the central barrier.

One transparency note about the implementation. The profile integration is a hand-written umbrella integration over the mean force rather than a call to the conventional weighted-histogram utility. The reason is documented in the code: the biasing path used here does not emit the pull-force files that utility reads, so the original call was dead and silently produced nothing. The function names still contain the older term, which is a naming debt rather than a behavioural one, and it is recorded here because a reader inspecting the output would otherwise reasonably assume the standard tool ran.

From free energy to a reported number. The receptor leg converts to a dissociation constant through Kd = exp(-|dG|/RT). That conversion is where a caveat belongs: the value is computed relative, without a standard-state correction, so it is valid for ranking variants against each other and must not be quoted as an absolute affinity.

5.5 Endosomal Escape - Three Multiplied Factors, Not One Score

The escape index is deliberately a product of three terms rather than a single fitted score, because the three describe independent physical requirements and a failure in any one of them cannot be compensated by the others.

The first is the ionisation balance: given the apparent pKa, a Henderson-Hasselbalch calculation gives the fraction of lipid protonated at endosomal pH. This is the term the pKa design work targets.

The second is helper-lipid fusogenicity. Protonation alone does not restructure a membrane; the composition must be able to adopt a non-lamellar geometry. A cone-shaped unsaturated phosphatidylethanolamine has negative spontaneous curvature and promotes the inverted-hexagonal contact that breaches the bilayer, whereas a cylindrical saturated phosphatidylcholine packs into stable lamellae and resists it. The engine carries a per-helper fusogenicity weight for this reason, and the resulting trade-off is real: the fusogenic helper that improves escape is also the oxidisable one that shortens shelf life.

The third is buffering capacity, the proton-sponge contribution. A lipid with several titratable amines absorbs protons as the endosome acidifies, which both drives counter-ion and water influx and osmotically stresses the membrane. The engine counts only titratable nitrogens for this term, not total nitrogen - an amide nitrogen contributes molecular weight and no buffering, and conflating the two overstates the effect.

Multiplying the three encodes the biology correctly: a perfectly tuned pKa on a rigidly lamellar composition escapes poorly, and so does a highly fusogenic composition whose lipid never protonates.

5.6 The Protein Corona - The Surface the Body Actually Recognises

The biology. Within seconds of entering blood, a nanoparticle is coated in plasma protein, and from that moment the biological identity the body reads is the corona, not the lipid. The clearest demonstration is hepatocyte delivery: a conventional ionizable LNP reaches hepatocytes because it adsorbs apolipoprotein E, which is a ligand for the low-density-lipoprotein receptor. The particle is not targeted; it is opsonised into a targeting interaction. Remove that adsorption - by raising PEG density, for instance - and hepatic delivery falls even though nothing about the lipid changed.

This is also why the organ-targeting composition variants work without any ligand. Changing the surface charge changes which plasma proteins adsorb, and it is the adsorbed set that the endothelium of a given organ recognises.

The computation, and one thing it deliberately refuses to do. The engine estimates competitive adsorption of apolipoprotein, albumin and opsonins, and combines them with a multivalent avidity model. The apolipoprotein term saturates with residence time through 1 - exp(-1/tau), which encodes that adsorption is fast and self-limiting rather than linear in exposure.

The refusal is the part worth highlighting. The albumin and opsonin terms are computed only when both PEG surface density and zeta potential are available and finite. When either is missing, the module does not substitute a default, a mean or a zero - it emits an explicit unscored status together with the list of missing inputs, and the ranking layer renormalises that axis out. A numeric guard exists specifically to refuse coercion of a missing value into a number.

This sounds like a small design choice and it is the most consequential one in the chapter. A screening pipeline that defaults a missing corona term to a neutral value converts absent evidence into a passing grade, and because the default is applied consistently it is invisible in every summary statistic. Chapter 8 gives the measurement that shows how badly this distorts a ranking when it is allowed to happen.

5.7 The Length Penalty in RNA Hydrolysis - The Molecular Reason for Ultra-Cold Distribution

RNA degrades chiefly by hydrolysis of the phosphodiester backbone, accelerated by the 2'-hydroxyl that DNA lacks, by hydroxide, and by divalent metal ions. The crucial consequence for formulation is statistical rather than chemical: a transcript is intact only if every one of its linkages is intact. If each linkage has an independent per-unit-time probability of cleavage, the fraction of full-length molecules decays with the product of length and time. A 4,000-nucleotide messenger RNA therefore loses integrity roughly four times as fast as a 1,000-nucleotide one under identical conditions, and a short duplex is effectively immune on the same timescale.

This single fact explains most of the observed difference in cold-chain requirements across modalities, and it is why the platform treats payload length as an input to the stability axes rather than as metadata. It is also why the engine's own defect history around payload length matters: a stage that read the cargo class's *default* length instead of the designed molecule's actual length would compute the stability of a different molecule than the one being ranked. That defect existed, is recorded, and is the reason payload length is now resolved per design.

The mitigations the platform scores follow from the same mechanism: chelation to remove catalytic metals, pH control near the hydrolytic minimum, lyophilisation to remove the water that is both reactant and mobility medium, and the lowest practical storage temperature. Freeze-drying is the one that changes the distribution economics, and it is scored as its own axis precisely because a formulation that survives lyophilisation and reconstitution has a fundamentally different commercial profile from one that requires frozen distribution.

5.8 Colloidal Stability - One Number That Governs Everything Downstream

A lipid nanoparticle suspension is thermodynamically unstable by construction; it is kinetically trapped. Whether it stays trapped is governed by the balance between van der Waals attraction, which is always present, and electrostatic and steric repulsion, which are engineered. The resulting energy barrier, expressed in units of thermal energy, decides whether two particles that collide by diffusion fuse or separate.

The practical value of computing this barrier is that a single number predicts a long list of otherwise unrelated observations: whether the particle-size distribution drifts during storage, whether subvisible particulate counts rise past compendial limits, whether the product survives freeze-thaw cycling, whether it aggregates under the shear of injection through a fine needle, and whether agitation during transport nucleates aggregates. The engine computes the barrier and also the critical coagulation concentration - the ionic strength at which repulsion is screened enough for the barrier to vanish - which is what connects the physics to a buffer specification.

The biology this protects is simple and unforgiving: an aggregate is not a large particle, it is a lost dose and an immunogenicity risk. Aggregated material is cleared by phagocytes rather than delivered, and particulate matter in an injectable is both a compendial failure and a route to infusion reactions.

5.9 Convergence Is Tested, Not Assumed

A molecular-dynamics observable computed from a trajectory that has not equilibrated is not an estimate with a large error bar; it is a measurement of the initial condition. The platform therefore tests for saturation rather than running a fixed duration and reporting whatever it finds.

The test corrects for autocorrelation, which is the essential detail. Successive frames of a trajectory are not independent samples, and treating them as such understates the standard error by the square root of the correlation time - often an order of magnitude. A naive stationarity test on correlated data will therefore declare convergence early and confidently. Only after correcting for the correlation time does a flat mean actually mean a converged mean.

Averages are taken over the trailing fraction of each trajectory, and the central estimate across independent seeds is a median rather than a mean, so that one seed that fell into an unphysical configuration cannot drag the reported value. Multiple seeds are not a refinement here: a single seed cannot distinguish a converged result from a trapped one, so seed spread is part of the measurement rather than an optional check.

Duration is budgeted in explicit tiers - nominally 50 nanoseconds, 1.5 microseconds and 2 microseconds of production sampling - so that a screening question and a mechanistic question can be asked of the same pipeline at very different cost. The integration step is 0.02 picoseconds for the coarse-grained model, reduced to 0.005 for one variant whose chemistry required it; that reduction quadruples the cost of the same simulated time, which is exactly the kind of decision that should be recorded per variant rather than applied globally.

6. The Compendial CMC and Regulatory Layer (42 Modules)

This is the largest single subsystem in the platform, and its existence is the main structural difference between this engine and a lipid-design tool. Its purpose is to answer, at design time, the questions that conventionally surface at registration stability or process validation - when the answer costs a reformulation rather than an afternoon.

All 42 modules are additive risk lenses, not gates. None of them can remove a candidate from the ranked output. Each emits a normalised index between 0 and 1, a categorical tier, and one or two raw physical values in their own units, and the whole set is folded into a single capped composite term. Chapter 8 explains why the composite is capped rather than carried as 42 independent terms.

6.1 Sterile Product and Container (10)

A note on why subvisible particulates deserve a dedicated axis rather than a mention in a stability discussion: for an injectable nanoparticle product, the particulate specification and the product itself are made of the same material. Any aggregation pathway converts product into specification failure, and because the compendial thresholds are counts rather than mass fractions, a very small mass of aggregate can fail a lot. The colloidal-stability barrier computed in Chapter 5.8 is the upstream physical predictor for this axis, which is why the two are wired together rather than reported separately.

6.2 Impurities (4)

The nitrosamine axis is worth singling out because it is the clearest case where a purely structural prediction has real regulatory consequence. Ionizable lipids are built around tertiary amines by design - that is the functional group that titrates. Tertiary and secondary amines are also the substrate for nitrosamine formation in the presence of a nitrosating agent, which can arrive as a nitrite impurity in an excipient. The resulting nitrosamine drug-substance-related impurity falls in a cohort of concern with acceptable intakes in the nanogram-per-day range, which is far below what a routine impurity method would even detect without being designed for it. Identifying this liability from the structure, at design time, is the difference between choosing a different amine and discovering the problem in a registration batch.

6.3 Stability, Storage and Distribution (20)

Two of these encode the genuinely counter-intuitive failure modes, and they are the reason this group is large.

Freezing is a chemical event, not only a thermal one. When a buffered solution freezes, its components do not vitrify together. A buffer whose components have different solubilities will crystallise one component preferentially, and the pH of the remaining unfrozen concentrate shifts - in the worst commonly encountered case by more than a full unit, in the acidic direction. A formulation buffered comfortably at the hydrolytic minimum at room temperature can therefore spend its frozen storage at a pH where hydrolysis is substantially faster. Freezing to protect the payload can accelerate its degradation, and only an axis that models the freeze-concentrate pH rather than the formulated pH will see it.

A ligand can survive drying structurally and not functionally. A surface-conjugated targeting ligand may remain chemically attached through a lyophilisation cycle while losing the conformation or the surface presentation that its receptor requires. A recovery measurement based on size, polydispersity and encapsulation will report a clean reconstitution and say nothing about it. That is why targeting retention is scored separately from reconstitution rather than folded into it.

6.4 Process and Manufacturing (8)

The microfluidic scale-up axis captures a failure that is routine and expensive. Particle size in nanoprecipitation is set by the ratio of mixing time to aggregation time, not by composition alone. A composition that yields a well-behaved particle in a bench microfluidic chip can yield a different size distribution in a production mixer whose mixing time differs, and the change appears as an out-of-specification size on a batch that was formulated correctly. Modelling the mixing regime at design time is what allows the composition to be chosen for scale-invariance rather than for bench performance.

6.5 What the Dossier Module Produces - And What It Refuses to Produce

The dossier module aggregates predicted and known critical quality attributes into a specification-table skeleton, with a test method and acceptance criterion per attribute, plus a completeness percentage and an explicit gap report.

The gap report is the important half. Certain attributes cannot be settled by computation at all - identity confirmation, a real stability dataset, a measured impurity profile, sterility, endotoxin. The module lists these as open gaps rather than filling them with a computed proxy. A tool that quietly populated them would produce a document that looks finished and is dangerous, because a reviewer cannot tell a measured entry from an estimated one. Emitting a table that is explicitly 70 percent complete with the missing 30 percent named is more useful than a table that claims to be complete.

7. Design and Discovery Engines - The Layer That Decides What to Make

Up to v3.0 the platform's centre of gravity was scoring candidates it was given. The layer described here generates them, and it is what changed the platform's character from screen to designer.

7.1 Multi-Objective Pareto Search - Removing the Wrong-Weights Risk

Any weighted score embeds a judgement about relative importance, and that judgement is usually unexamined. If encapsulation is weighted twice as heavily as escape, the ranking asserts that a given improvement in encapsulation is worth twice as much as the same improvement in escape - a claim nobody actually holds with confidence.

The Pareto search sidesteps the question. Rather than collapsing objectives into one number, it identifies the non-dominated set: every design for which no other design is at least as good on all objectives and strictly better on one. The result is a front rather than a winner, and the front is accompanied by a hypervolume coverage metric that quantifies how much of the achievable trade-off space has actually been explored - which distinguishes a genuinely thin front from an under-sampled search.

The practical value is that the front exposes the shape of the trade-off. If the front is sharply curved, a small concession on one axis buys a large gain on another, and that is worth knowing before committing. If it is nearly flat, the axes are not really in tension and the weighting argument was moot.

7.2 Joint Optimisation - Chemistry and Vehicle Do Not Separate

Optimising the lipid, then the composition, then the architecture, in sequence, finds the best composition *for the lipid chosen first*. Since composition changes the assembled particle's apparent pKa, and apparent pKa is what the lipid was selected on, the sequence is circular and the result depends on where it was entered.

The joint optimiser searches chemistry and architecture together. In the current release this is the stage whose cache key was extended to cover its activity model, and consequently the stage whose numbers move - 184 checkpoints re-run. Reporting that plainly is the point: a fix that changes results is a fix that was doing nothing before.

7.3 De-Novo Generation - Beyond the Catalogue

Two generators produce structures rather than selecting them. The ionizable-lipid generator assembles fragments conditioned on a target apparent pKa, which is the right conditioning variable because it is the parameter the downstream biology actually reads. The ligand generator proposes targeting heads for a receptor that the catalogue does not cover.

The honest framing: a generated structure is a hypothesis with a synthesis cost attached, and the platform scores synthesisability alongside predicted performance for that reason. A generator that proposes unmakeable molecules with excellent scores has produced nothing.

7.4 Composition Search and Organ Targeting

The composition search explores the lipid mol-ratio simplex jointly with the choice and proportion of a fifth organ-targeting lipid and the PEG molecular weight. Chapter 2.2 covered the mechanism - charge sign redirects tropism through corona composition - and the search is what turns that mechanism into a quantitative design space.

One limitation is specific and should be read before relying on this layer for a structural prediction. The engine registers 15 candidate fifth lipids, and 9 of the 15 have no dedicated coarse-grained parameter file. Those nine are fully usable for the scoring stages but cannot be built into a simulated particle. A design that selects one of them is therefore scored on chemistry and left without a structural cross-check, and the output records that rather than substituting a similar lipid.

7.5 Active Learning - Feeding Wet-Lab Results Back

Predictions that never meet measurement do not improve. The feedback module ingests wet-lab results, updates a posterior over the design space, and proposes the next batch by an acquisition function that balances expected performance against expected information gain - so that a batch is chosen partly to resolve uncertainty rather than purely to maximise predicted score.

Current status, stated plainly: there is no large prospective wet-lab cohort validating the platform's predictions as of this document [verification required]. The feedback machinery is implemented and tested; the corpus it exists to consume is not yet large. Every performance statement in Chapter 10 therefore rests on internal computational validation, literature directionality and unit-test passage, not on a prospective blinded comparison.

8. The Decision Layer - How Many Axes Become One Number

8.1 The Row, and Why Its Width Is Itself a Risk

The ranking stage reads more than sixty checkpoints per variant and flattens them into one row carrying 382 distinct keys. That row is the unit of comparison, the audit record, and the thing a customer receives.

One property of it deserves disclosure because it is a real fragility rather than a design feature: the output table's header is taken from the keys of the first row. If the first variant happens to lack an axis that later variants carry, that column is absent for everyone. The engine's own documentation flags this. It is stated here because a customer comparing two runs with different variant orderings could otherwise see a column appear or vanish and reasonably conclude that a computation had failed.

8.2 The Success-Probability Fusion

The composite is a gated, weight-renormalised mean:

p = gate * SUM(w_i * v_i) / SUM(w_i)      for every axis i that is PRESENT

Each contributing value is normalised to the interval 0 to 1, and - this is the load-bearing clause - an absent axis is omitted from both sums rather than scored as zero or as a default. The mean is renormalised over what was actually measured.

Weights are calibratable from an external file and fall back to hand-set defaults. The heaviest are encapsulation efficiency at 0.20, predicted transfection efficiency at 0.14, and then a cluster at 0.12 comprising morphology, modification quality, freedom-to-operate, and the entire manufacturability composite. Affinity, therapeutic index, pKa window fitness and potency each carry 0.10. Roughly forty further axes carry between 0.02 and 0.09.

8.3 Why the Quality Layer Is One Capped Term Instead of Forty-Two

This is a subtle arithmetic point with a large effect on ranking resolution, and it is worth stating because it looks at first like an under-weighting of quality.

The 42 quality axes are near-saturated by construction: a well-designed formulation scores highly on most of them, so their values cluster near the top of the range. Carried as 42 independent terms in a renormalised mean, they would dominate the denominator and pull every candidate's composite toward the same high value - compressing the very differences the ranking exists to expose. Two candidates differing materially in escape free energy would appear nearly identical because forty-two near-identical quality terms swamped the difference.

The quality layer is therefore accumulated separately and collapsed into one capped composite term. The effect is that quality still gates attention and still appears in full detail in every row, while the ranking's discriminating power is preserved for the axes that actually vary between candidates. The 42 axes are not down-weighted in importance; they are prevented from acting as forty-two votes for the same proposition.

A related correction appears inside individual axes. Affinity and therapeutic index were both rescaled to logarithmic form because their linear forms saturated - a change from 1 nanomolar to 0.1 nanomolar is enormous biologically and nearly invisible on a linear axis bounded at zero.

8.4 The Weakest-Link Gate - What an Average Hides

A mean rewards compensation. A particle that is outstanding on nine axes and disqualifying on the tenth will out-score a uniformly good one, even though the tenth axis alone determines that it cannot be made or cannot be dosed. Delivery is a serial process, and a serial process is governed by its worst step.

The platform therefore applies a continuous multiplicative penalty derived from the worst category, not the mean: the row is collapsed into five failure categories - delivery, stability, safety, potency, and intellectual property with manufacturing - and the weakest category's score, relative to a threshold of 0.5 and softened by a square-root exponent, multiplies the composite. The softening is deliberate; a hard cliff would make the ranking unstable to small changes near the threshold.

An optional stricter layer exists on top of this, and its three-state design is the notable part. It can floor a candidate's gated probability to zero when its weakest category falls below a limit - but the default is None, meaning the veto layer is not applied at all, which is a distinct state from "applied and passed". Conflating "no veto configured" with "veto passed" is precisely the error the three-state convention exists to prevent.

8.5 The Measurement That Forced an Evidence Tier

This is the most instructive number in the document, and it generalises well beyond this platform.

Because an absent axis is omitted and the mean renormalised, a variant with fewer measured axes is scored only on what it has - and a variant that never ran the expensive physics stages therefore has no opportunity to score badly on them. The consequence was measured directly: median composite probability was 0.7882 across 68 variants with no molecular-dynamics evidence, against 0.6401 across 56 variants that had it, and the top eight rows were all molecule-dynamics-free. The pipeline was systematically rewarding ignorance.

The crucial insight is that no choice of weight fixes this. A weight multiplies a value; it cannot penalise a value that is absent from the sum. Down-weighting the physics axes makes the problem worse, and up-weighting them changes nothing for the rows that lack them.

The fix is therefore not a coefficient but a lexicographic ordering: the wet-lab hand-off list sorts first on whether molecular-dynamics evidence exists, and only then on composite probability. A well-evidenced candidate at 0.64 is ranked above an unevidenced one at 0.79, because the comparison between them is not meaningful on the composite alone. The full ranked table retains both, with the evidence tier as an explicit column.

The general principle: renormalising over present axes is correct for comparing like with like, and becomes a bias the moment the set of present axes differs between candidates. Any pipeline with optional expensive stages has this bias unless it separates evidence depth from score.

8.6 Uncertainty Propagation - Thin Evidence Yields Wide Intervals

Every composite is reported with a credible interval derived from a Beta posterior, so the interval widens when the evidence is thin rather than staying decoratively constant. Ranking is available on the lower bound rather than the point estimate, which is the conservative choice: it prefers a candidate that is confidently good to one that is possibly excellent.

8.7 Failure-Mode-Decorrelated Panel Selection

When a customer can synthesise five candidates, taking the top five by score is usually the wrong choice, because the top five often share a chemical scaffold and therefore share a failure mode. If that mode is real, all five fail together and the experiment returns one bit of information.

The platform selects a panel whose members differ in their weakest category, so that the batch tests several independent hypotheses. The chemical-duplicate groupings from Chapter 2.2 feed this directly: two variants that are structurally identical cannot be panel-diverse no matter how they score.

8.8 The Bench-Eligibility Filter Is Separate from the Score

Five conditions remove a row from the wet-lab hand-off list: a blocking freedom-to-operate verdict, a physicochemical failure, stale upstream evidence, a placeholder target, and chemical divergence between design and scored structure.

In every one of those cases the row remains in the full ranked table with its flag attached. Excluding a candidate from a synthesis list and deleting it from the record are different acts, and conflating them destroys the reviewer's ability to ask why something is missing. Two of the five conditions can be overridden by explicit flags, which is appropriate: stale upstream evidence may be acceptable to a user who knows why it is stale.

9. Long-Running Simulation Reliability

A screening calculation that runs in a minute either works or visibly fails. A multi-day physics campaign across dozens of variants fails in a third way: it appears to work, produces output, and the output does not mean what it says. The modules in this chapter exist because each of these failure modes was observed in operation, not anticipated in design.

9.1 A Proxy Signal Is Not the Fact

Table 9-1. Proxy signals versus real verification. Each row is a failure mode observed in operation and subsequently encoded [internal].

The first row is worth expanding because it is the cheapest lesson in the table. A machine-learning predictor was gated on whether its directory existed. The directory existed; the model weights and two required Python packages did not. The status display reported the tool as available for the entire period during which the corresponding stage wrote null predictions into every row. Nothing errored, and a summary statistic over those rows would have been computed over nulls. The probe now requires the directory, the weight file, and each import - and reports the tool unavailable if any is missing.

9.2 Absence Must Be Representable

The single most repeated lesson across the platform's defect history is that a missing value needs its own representation, distinct from both a passing value and a failure. The engine now implements this in several places with the same shape: a triple of value, record, and status, where the value may be None and the status says why.

Five scoring helpers use this convention, and the corona module uses it with explicit scored and unscored literals. The consequence flows through to the ranking arithmetic, which omits and renormalises rather than defaulting, and to the report layer, which renders an explicit absence row rather than a blank cell.

The brain-targeting axis shows the convention doing real work. Of eight brain-targeted variants, four are unscored on the blood-brain-barrier axis, and the code states the rule explicitly: unscored does not gate. Those four are neither failed nor passed; they are carried with the axis renormalised out and the status visible. A two-state gate would have had to either fail them - discarding four legitimate candidates - or pass them, silently claiming a barrier assessment that was never made.

9.3 Chunked Production and Cross-Restart Resume

Long simulations are run in chunks so that an interruption costs one chunk rather than the campaign, and resume is decided by content rather than by presence. A restart re-reads what was actually produced and continues from there.

Expensive stages receive special treatment in the staleness logic that is worth explaining, because it is an explicit trade-off rather than an oversight. When upstream evidence changes, a cheap stage is simply recomputed. An expensive stage - build, production dynamics, trajectory analysis, the free-energy legs, corona, back-mapping, membrane free energy - is neither recomputed nor deleted. Its result is returned carrying a staleness flag, the basis of the staleness, and the reason. The alternative would be to discard days of computation automatically on an upstream edit, and the judgement is that a human should make that call. Two explicit flags exist to make it: one forces recomputation, the other accepts the stale result and lets the row through with its flag intact.

This is also why the structural invariant of Chapter 3.6 matters. The fingerprint-invalidation path has no expensive-stage guard, so it is safe only while no expensive stage carries a fingerprint - and the program refuses to start if that ever stops being true.

9.4 Shared-Accelerator Reality

The physics layer runs on a shared accelerator alongside other pipelines, which produces failure modes that are invisible in single-tenant testing. The engine bypasses the accelerator's multi-process scheduler by default, sets the bypass before any numerical library is imported so the setting actually takes effect, pins cores per worker, and retries a failed launch a bounded number of times with backoff before falling back to a slower processor-only path.

The bypass default is the outcome of an investigation rather than a preference, and the honest summary is mixed: routing through the scheduler eliminated one class of hardware fault but did not eliminate the crashes under concurrent load, so the bypass was retained and the underlying crash remains open. What keeps campaigns progressing is the retry-to-quorum loop, in both configurations. Stating this is more useful than claiming a solved problem, because a customer running the mechanistic tier on shared hardware will see occasional retries and should know they are expected and absorbed.

9.5 Determinism in Practice

Determinism is asserted through a fixed hash seed set at import, seeded stochastic steps, and sorted, stably serialised checkpoints. The value of this is narrow and real: it means a re-run reproduces numbers exactly, so a discrepancy between two runs is evidence of a code or input change rather than of sampling noise. That property is what makes the fingerprint system meaningful - without determinism, a changed number could never be attributed.

10. Performance Benchmarking Against Industry Standards

The platform is positioned, honestly and deliberately, as an in-silico triage and design layer that sits upstream of - and is confirmed by - the established experimental gold standards rather than replacing them. Its job is to decide where to spend scarce wet-lab capacity, so that the assays which follow are run on the molecules most likely to survive them.

Table 10-1. Positioning by capability. No row claims that computation replaces measurement.

10.1 Honest Limitations

Every module is explicitly labelled a heuristic rather than a measured result, and the label is enforced in the output rather than buried in a footnote. The list below is longer in this edition than in v3.0. That is intentional: the growth reflects better auditing, not new weakness, and a limitation that is named is one a customer can plan around.

Validation status

·       No prospective wet-lab cohort. As of this document there is no large prospective blinded dataset validating the platform's predictions [verification required]. Every performance statement rests on internal computational validation, agreement with literature directionality, and unit-test passage.

·       Tropism is rodent-trained. The composition-graph tropism model is built predominantly on rodent in-vivo observation; primate and human values are translations by route-dependent literature factors, not a retrained human model [literature].

·       Free energies are relative. Receptor-binding free energy carries no standard-state correction. It is valid for comparing variants and must not be quoted as an absolute dissociation constant.

Coverage gaps that a reader could otherwise mistake for coverage

·       Eleven of thirteen modalities are latent. All 170 entries in the discovered-variant registry carry siRNA cargo. The other modalities are implemented, unit-tested and reachable, but unexercised by any live registry entry.

·       Nine of fifteen organ-targeting lipids have no coarse-grained parameters. They are scoreable in stages 0 through 2 and cannot be built into a simulated particle, so a design selecting one of them gets no structural cross-check.

·       The sterol layer is externally catalogued and defaults silently if unqualified. A formulation record that gives a sterol percentage without naming the sterol is defaulted to cholesterol, with the defaulting recorded. In one measured campaign all 60 variants were in exactly that state, so sterol scoring was skipped for 60 of 60. The defaulting is now recorded per variant; it is not a reason to omit the sterol name.

·       PEG-lipid is one species, not a catalogue. A single PEG-lipid identity is hardcoded. PEG molecular weight and density are searchable; PEG-lipid chemistry is not.

·       A third of the discovered registry carries no vehicle class. 56 of 170 entries have no vehicle class assigned, which routes them to default handling rather than to class-specific treatment.

Reporting and interface fragilities

·       The engine's own version banner is wrong. There is no version constant in the program. The command-line description and the runtime log banner both name a release number hundreds of revisions old. Version identity is carried by the filename and the changelog only. A user who reads the banner to identify what ran will be misled; the filename and the run manifest are authoritative.

·       Table columns come from the first row. The output table's header is the key set of whichever variant is emitted first, so an axis absent from that variant is absent as a column for all of them.

·       A naming debt in the free-energy layer. Function names retain the term for a standard histogram-reweighting utility that is not invoked; the integration is a hand-written umbrella integration. Behaviour is as documented in Chapter 5.4; the names are misleading.

·       The engine produces no plots. All output is tabular or textual. Figures in customer-facing material are generated separately.

Operational

·       One scoring axis carries a commercial-licence obligation the code does not enforce. The force-field cross-check of a targeting-head binding pose is driven through the CNS engine, which is free only for non-profit use; commercial deployment requires a licence from its vendor. The source records the obligation at the point of use and deliberately does not gate on it, so a commercial run will produce the number regardless. Appendix A.4 gives the full position and the two ways to resolve it.

·       A crash under concurrent accelerator load remains open. The physics layer occasionally crashes when several simulations share one accelerator. The retry-to-quorum loop absorbs it and campaigns complete, but the root cause is unresolved and the mechanistic tier should be budgeted with that in mind.

What this means in aggregate. The platform's genuine value is narrower and more durable than "replacing the lab": reproducible, auditable, pre-experiment risk triage and candidate ranking that concentrates finite wet-lab effort on the most promising and most manufacturable particles, and that flags manufacturing landmines early enough to design around them. In a regulated setting a tool that overclaims is worse than useless, because it invites a reviewer to trust a number no measurement supports.

11. From a Drug- and Vaccine-Development Perspective

Lipid nanoparticles are the enabling delivery technology behind the first approved siRNA drugs, the messenger-RNA vaccines administered at global scale, and a growing pipeline of messenger-RNA therapeutics, in-vivo editing payloads and tissue-targeted RNA. Across all of them a consistent and counter-intuitive lesson has emerged about where programmes actually fail: not in the biology of the encoded product, which is often the easiest part to get right, but in the vehicle.

11.1 Where It Accelerates Each Development Phase

11.2 The Vaccine-Versus-Therapeutic Distinction, Stated Mechanistically

This distinction governs more design decisions than any other and is frequently collapsed into a single notion of "tolerability", so it is worth separating.

A prophylactic vaccine wants a controlled amount of innate stimulation. The innate response to the particle and its payload is part of the mechanism: it recruits and activates the antigen-presenting cells that generate the adaptive response. A vaccine particle is dosed once or twice, so slow liabilities - lipid accumulation, anti-PEG titre, chronic hepatic exposure - have limited opportunity to matter. The binding constraint is acute reactogenicity.

A chronically dosed therapeutic wants the opposite on every count. Innate stimulation is pure toxicity with no benefit. The slow liabilities become the dominant ones. Two deserve specific mention because they are invisible on a single-dose study:

*Accelerated blood clearance.* PEG is immunogenic. A first dose can raise anti-PEG antibodies that bind the second dose's PEG layer, opsonise it, and strip it from circulation within minutes - so the second dose fails not because the particle is wrong but because the immune system learned the first one. A single-dose experiment cannot see this, and the platform scores it as its own axis for that reason.

*Ionizable-lipid accumulation.* The lipid must persist long enough to deliver and clear fast enough not to accumulate across a chronic schedule. A biodegradable ester linkage is the standard resolution, and it trades directly against the hydrolytic stability of the drug product itself - the same bond that clears in vivo hydrolyses in the vial.

For an editing payload the interaction is different again. A ribonucleoprotein or a nuclease-encoding payload is a foreign protein, and a strongly cationic lipid acts as an adjuvant toward it - amplifying anti-drug-antibody formation against the very payload the therapy depends on. Because editing is typically a single administration, an anti-drug-antibody response may be acceptable; if redosing is ever contemplated, it is not. The platform computes this coupling quantitatively through the bridge module rather than noting it as a caveat.

11.3 Strategy Summary by Modality

11.4 Where the Platform Changes a Go / No-Go Rather Than a Ranking

Three cases are worth naming because in each the platform's output is decision-grade rather than merely informative.

A nitrosamine liability found from structure. If the ionizable lipid's amine architecture carries formation potential, the correct response is to change the amine, and that is a design-stage decision costing an afternoon. The same finding in a registration stability batch costs a reformulation and a year.

A deliverable-volume or injectability failure. These are geometric and rheological facts about the final product. A formulation that cannot be withdrawn at the labelled volume, or cannot be expressed through the intended needle, fails regardless of biological excellence - and neither depends on any biological uncertainty, so a design-stage prediction is close to definitive.

A vehicle-class reassignment. For a short duplex the platform may conclude that a conjugate is the better vehicle than an encapsulating particle. That is not a ranking adjustment; it removes the ionizable-lipid toxicity budget, the encapsulation specification and much of the cold-chain problem from the programme altogether.

12. Customer Guide - What You Provide and What You Receive

You do not need a supercomputer, a cluster allocation or a resident molecular-dynamics specialist to get value from the platform. The two layers most users rely on day to day - the quality risk layer and the success-probability ranking - run as fast pure logic on an ordinary processor, with neither an accelerator nor network access required. The heavy physics stages are entirely optional and engage only when mechanistic, physics-based structure prediction is specifically wanted.

12.1 Execution Tiers and Turnaround

Table 12-1. Execution tiers. Tier 1 alone produces every principal deliverable, including the ranked list and the specification table; the dynamics-derived axes are marked not-adjudicated and their weight is renormalised out.

Two points about choosing a tier, both of which follow from Chapter 8.5 rather than from cost.

Tier 1 is a complete answer, not a degraded one - provided every candidate is run at tier 1. The renormalisation is sound when the set of present axes is the same across candidates.

Do not mix tiers within one comparison. Running some candidates at tier 3 and others at tier 1 and then ranking them together reintroduces exactly the bias that the evidence tier exists to contain: the tier-1 candidates will score higher on average because they have no physics axes to score badly on. If budget permits the mechanistic tier for only part of the set, use it to confirm a tier-1 ranking, not to produce a mixed one.

12.2 The Standard Engagement Flow

12.3 Deliverable Specification

·       Full ranked table - every variant, with success probability and lower credible bound, all scored axes, and explicit not-adjudicated markers

·       Wet-lab priority table - the top candidates after the eligibility filters, with failure-mode-decorrelated panel selection applied and the evidence tier as the primary sort key

·       Per-cargo split tables - partitioned so comparison is within a modality rather than across modalities

·       Per-variant design package - design summary, sequence, lipid composition, freedom-to-operate report, bench protocol, microfluidic parameters, active-targeting specification, barcode and primers

·       Specification table - critical quality attributes with test methods and acceptance criteria, plus the gap report and completeness percentage

·       Plate map and synthesis order file - directly usable at wet-lab hand-off

·       Run manifest - run metadata, cargo distribution, and the module versions used

·       Report document - an eleven-section report in Markdown and Word form, covering summary, payload, composition, manufacturing method, chemistry, self-assembly, scoring axes, biodistribution, specification conformance, open items, and reproducibility information

12.4 How to Read the Output Without Being Misled

This section exists because the most likely way to misuse the platform is to read a number without its status.

1.     Check the status column before the value. An axis reported as unscored or not-adjudicated has had its weight renormalised out. It is not a zero and not a pass.

2.     Prefer the lower credible bound for decisions. The point estimate is the platform's best guess; the lower bound is what survives thin evidence.

3.     Read the evidence tier alongside the probability. A high probability on a candidate with no physics evidence means fewer opportunities to score badly, not superior performance.

4.     Treat the gap report as the specification table's most important page. Completeness percentage tells you how much of the document is prediction.

5.     Identify the run by the manifest, not the log banner. The banner's version string is stale (Chapter 10.1).

6.     Check whether flagged rows were meant to be excluded. A row can remain in the full table while being filtered out of the bench list; the flag says which.

12.5 What You Do Not Get - And Why We Say So Plainly

The platform does not replace release testing, a real stability study, an aseptic media fill or a measured impurity determination. It does not certify a lot, it does not generate data a regulator can accept as evidence of quality, and it relieves no one of running the validated assays.

Every number it produces is a prediction whose purpose is to concentrate and de-risk the experimental programme - to tell the laboratory where to look first and what to watch for - not to stand in for the experiment. This honesty is built into the software rather than left to the reader: each module carries an explicit caveat string in its output, and the generated dossier flags measured-only attributes as open gaps rather than quietly filling them with a computed proxy.

13. Requirements-Compliance Comparison (Publications, Open Tools, Commercial Software)

The table below compares the platform, line by line, against requirements that published methods, open tools and commercial packages typically demand or claim to satisfy. Comparators are described as categories rather than by product name: a category-level comparison is the more verifiable claim, and naming a vendor product would invite a feature-by-feature dispute that no customer benefits from. Where a measured method or an established suite does a job better, the verdict says so.

Table 13-1. Requirements comparison. The last four rows are included deliberately. A comparison table in which the authoring platform wins every row is not a comparison, and a reader who finds no losing row correctly discounts the whole table.

13.1 Requirements Drawn from Publication Practice

Three requirements come from peer review rather than from regulation, and they are worth addressing separately because they are the ones a reviewer of a manuscript would raise.

Method reproducibility. A computational result should be reproducible from the description. The platform satisfies this by determinism plus archived snapshots: 372 engine snapshots are retained, so any past result can be re-run against the exact code that produced it. This is stronger than a version number in a methods section.

Stated uncertainty. A point prediction without an interval is not a scientific claim. Every composite carries a credible interval whose width is proportional to evidence.

Negative and absent results reported. This is where most tool documentation fails and where this document deliberately does not: the latent modalities, the pKa gate with no discriminating power, the nine unparameterised lipids, the sterol scoring skipped for a whole campaign, the stale version banner, and the open crash under concurrent load are all stated in Chapter 10.1 rather than omitted.

Where publication practice is not met. A held-out prospective benchmark against measured outcomes is the standard a published predictor is held to, and the platform does not yet have one. Unit-test passage and literature directionality are weaker evidence, and they are what exists today.

14. Reproducibility, Validation and Auditability

A recurring requirement in regulated software - and separately in academic peer review - is that a result be reproducible and traceable to the process that produced it. The platform was built around that requirement rather than retro-fitted to it. The engine is deterministic and seeded, so identical input yields identical output on every run; it is checkpointed at every stage, so the provenance of each value is recoverable; and the pure-logic layer is backed by an automated suite that runs on an ordinary processor in minutes without an accelerator or a network.

14.1 Test Status - The Actual Numbers, Measured for This Document

Table 14-1. Test status measured against the engine described here for this document. It is a convention here neither to round the pass count nor to omit the failures.

There is a reason for writing the table this way. "All tests pass" reads better, but it conveys nothing to a reader who cannot check it and costs credibility with one who can. Naming which modules currently fail makes every other claim in this document more believable.

14.2 Where the Failures Are, and What They Mean

The 33 failures and 6 errors are not scattered. They fall into four groups, and 24 of the 33 concentrate in one subsystem.

The activity-model cluster is the substantive one, and it is an open defect rather than a stale expectation. Three distinct facts, each read directly from the failure output:

7.     A trained model checkpoint was not regenerated when its schema changed. The module declares schema version 2; the shipped checkpoint is version 1. The guard refuses the mismatch and states the semantic difference explicitly - a version-1 checkpoint treats one feature as a binary flag where the current module treats it as a graded quantity - and names the retraining script required. This is a schema guard working exactly as designed, reporting that a release step was missed.

8.     A scoring function exists as a copy rather than an import. A parity guard asserts that the environment-coupled module holds *the canonical function object*, and finds a copy instead. The guard's own message states its purpose: "that is the drift this guard exists to catch."

9.     The copy has already drifted numerically. The same parity test compares the two paths on a siRNA conjugate case and finds a difference of 4.4 x 10^-3 against a tolerance of 10^-9. The drift is small in magnitude and unambiguous in kind.

Taken together these three are the same lesson Chapter 3.5 draws from the cache history, appearing here in a different guise: duplicated logic drifts, and a trained artefact is a cached consumer of the code that trained it. The current release extended its cache key to cover this activity model precisely because it reaches a scoring stage through four layers of plain imports. The test suite is now reporting that the corresponding *artefact* - the trained checkpoint - was not carried along.

What this document does not claim. Whether the 29 single-strand-RNAi failures represent stale test expectations (the engine deliberately changed and the tests pin superseded behaviour) or a genuine regression (the tests are right and an opt-in gate stopped working) is not resolved here. Both readings are consistent with the evidence available from the failure output alone. The honest statement is that this subsystem has 32 failing assertions, the cause is identified as a schema and duplication issue, and the resolution is pending. A reader relying on the single-strand RNAi modality should treat it as under repair; the twelve other modalities are unaffected by this cluster.

14.3 Two Infrastructure Failures Worth Naming

Snapshot-pinned tests dangle. Two test files fail at collection because they import engine snapshots by absolute numbered path - and those snapshots have since been archived. The files are present in the archive directory; the tests look for them in the project root. A third test in the opt-in group fails the same way.

This is the same failure class as a stale index pin, and it is worth generalising: a test that pins a specific numbered snapshot becomes a dangling reference the moment that snapshot is archived. It is a useful pattern for asserting cross-version parity and it carries a maintenance obligation that is easy to forget, because the failure appears as a missing file rather than as a wrong answer.

One suite must not be run against a live campaign. Four errors arise in a watchdog test exercising monitoring behaviour. Separately - and this is an operational warning rather than a defect - the suite must be scoped to the platform's own test directory. Invoking the test runner without that scope lets it discover the snapshot archive, where a module-level exit call in an unrelated archived test terminates the collector outright. In this document's measurement that mis-scoped invocation produced 1,033 spurious errors before the correct scope produced the numbers in Table 14-1. Anyone reproducing these figures should scope the run explicitly.

14.4 Version Lineage and Snapshots

Beyond per-run reproducibility, the development process preserves history: every release is snapshotted and the prior snapshot archived, giving 372 retained engine versions in which any past result can be reproduced against the exact code that generated it. Determinism, checkpointing, an extensive suite and an archived lineage together satisfy the "show your work" bar - and do so more transparently than a closed product whose internal logic cannot be inspected, re-run independently or audited.

One caveat on identifying a version, repeated from Chapter 10.1 because it matters here most: the program carries no version constant, and both its command-line description and its runtime log banner name a release hundreds of revisions old. Version identity comes from the filename, the changelog and the run manifest. A reviewer reconstructing which code produced a result must use those and must not trust the banner.

14.5 Data-Integrity Conventions

These are stated as principles because each was derived from a specific defect and each generalises.

·       When a number crosses a file boundary, its uncertainty and provenance travel with it. If they do not, that is treated as a bug, not a simplification.

·       Arithmetic decides gates, rather than proliferating conditional branches. Inverse-variance weighting acting as the gate is more robust than a cascade of special cases, because a branch that was never written is a silent pass.

·       Epsilons are set in the meaningful units of the axis. Inserting an arbitrary small constant when a denominator degenerates amplifies rounding error into extreme standardised scores and can exclude every ordinary candidate.

·       Validation evidence is bound at the moment of production. A stamp attached afterwards buys retroactive applicability at the cost of correctness.

·       Logs append, so current state is read from the last entry. Reading the first match returns the oldest state.

·       Absence is representable. Three-state returns rather than a sentinel value; a missing number is never coerced into a number.

·       A trained artefact is a cached consumer of its training code. Section 14.2 is the current instance of this rule being enforced by a guard rather than by convention.

15. Conclusion

This platform is unusually broad in scope: nucleic-acid and ionizable-lipid design, physics-based structure prediction by coarse-grained dynamics, and a 42-module compendial quality and regulatory layer, inside one reproducible in-house engine. These three domains are almost never found inside one tool, let alone fused into a single decision. And they are fused: every signal collapses into one calibrated success probability and one specification table, so the molecule that ranks first is also the molecule whose manufacturing and regulatory risks are already enumerated.

The change of character in v4.0 reduces to one sentence: the evidence layer grew faster than the physics layer. The engine nearly doubled in size, and the larger part of the growth went not into new physics but into machinery that decides whether a computed number deserves to be believed - schema-versioned caches so a fix reaches the cached consumers that depend on it, three-state gates so an unmeasured axis is not scored as a passing one, a lexicographic evidence tier so a candidate cannot win by having been measured less, and a structural invariant the program refuses to start without.

That emphasis is not incidental to the scientific goal; it is what makes the scientific output usable. The measurement in Chapter 8.5 is the clearest illustration: a pipeline that renormalises over present axes will systematically rank the least-examined candidate highest, and no choice of weight repairs it. A platform that did not look for that bias would have shipped a ranking that quietly inverted its own purpose, and every downstream number would have been wrong in a way no summary statistic would reveal.

It does not replace the laboratory, and it is not designed to. What it does is tell the laboratory where to look, what will be hard to manufacture and which regulatory boxes are still empty - before a single milligram of lipid is synthesised. For an RNA-medicine developer that timing is the entire point: it is the difference between discovering a nitrosamine liability, an aggregation cliff or a deliverable-volume shortfall in a Phase 1 stability batch, where it costs a year and a reformulation, and designing that same liability out at the bench, where it costs an afternoon.

15.1 Adoption Checklist

·       [ ] The target gene or cargo sequence and its modality are settled

·       [ ] The target organ and route of administration are defined, including whether the hepatic default is acceptable

·       [ ] The route is declared explicitly, since it selects the apparent-pKa target window (Chapter 2.4)

·       [ ] A candidate ionizable-lipid set is chosen, or de-novo generation is delegated to the platform

·       [ ] The dosing regimen (single against repeat) is decided - it determines whether the anti-PEG and lipid-accumulation axes are engaged

·       [ ] The intended storage condition (refrigerated / frozen / lyophilised) is fixed, as the reference point for the shelf-life axes

·       [ ] If a sterol other than cholesterol is intended, it is named rather than given only as a percentage (Chapter 10.1)

·       [ ] The execution tier is chosen, and the same tier is applied to every candidate in the comparison (Chapter 12.1)

·       [ ] The number of candidates that can be committed to wet-lab validation is agreed

·       [ ] Any existing wet-lab data is prepared in the active-learning feedback format

·       [ ] An internal route exists for passing freedom-to-operate output to legal review

·       [ ] A plan is in place to obtain assays for the measured-only attributes named in the gap report

·       [ ] The reader of the output knows to check status columns before values, and to identify the run by its manifest rather than the log banner

15.2 Glossary

Appendix A. Third-Party Component Licences and Commercial-Use Status

A.1 The Scope This Assessment Covers

This appendix answers one narrow question: may the platform operator run these components as part of a commercial service? The scope is use only. It explicitly does not cover selling, sublicensing, redistributing, or embedding any of these components in a product delivered to a third party.

That distinction does most of the work here, and it is worth stating plainly because it is routinely conflated. The copyleft obligations of the GPL and LGPL - source availability, licence propagation, notice requirements - are triggered by distribution, not by internal use. An organisation running GPL software on its own machines to produce its own results incurs no distribution obligation. Consequently a licence that looks restrictive on a compliance checklist may be entirely unrestrictive for the use actually being made of it, while a licence that looks permissive may carry a use restriction that matters far more. Both cases appear in the table below.

Every entry's licence was read from the licence text or package metadata present on this system; the "Basis" column names where. Where no licence file was present locally, the entry is marked for verification rather than asserted.

A.2 Components Whose Licence Permits Commercial Use

Table A-1. Components whose licence permits commercial use under a use-only scope.

A.3 One Component with a Condition Attached - ViennaRNA

ViennaRNA 2.7.0 is not under an OSI-approved licence and therefore deserves its own entry rather than a table row. Its licence text, read from the installed package metadata, grants use explicitly:

> "Permission is granted for research, educational, and commercial use and modification so long as 1) the package and any derived works are not redistributed for any fee, other than media costs, 2) proper credit is given to the authors and the Institute for Theoretical Chemistry of the University of Vienna."

followed by:

> "If you want to include this software in a commercial product, please contact the authors."

Under the use-only scope of this appendix the position is therefore permitted, with two live obligations: the platform must not redistribute the package for a fee, and it must credit the authors and the Institute for Theoretical Chemistry of the University of Vienna. The "contact the authors" sentence attaches to including the software in a commercial product - that is, distribution - and would become operative if the platform were ever packaged and shipped rather than operated as a service. Anyone contemplating that change of business model should treat this as an open item, not a settled one.

A.4 The One Genuine Commercial-Use Restriction - CNS, via HADDOCK

This is the only entry in this appendix that is not cleared for commercial use, and unlike the copyleft cases the restriction bites on use, so the narrowing of scope does not resolve it.

The platform's force-field cross-check of a predicted ligand-receptor pose is driven by haddock3, which is itself Apache-2.0 and unproblematic. But the two modules used - topology generation and energy-minimisation scoring - are both modules of the CNS (Crystallography & NMR System) engine, and CNS is free only for non-profit use. Commercial deployment requires a CNS licence from BIOVIA.

The platform's own source records this obligation at the point of use rather than in a compliance document, and it is worth quoting because it also records the decision made about it:

> "LICENCE NOTE: haddock3 itself is Apache-2.0, but the CNS binary it drives is free only for non-profit use; topoaa and emscoring - the two modules used here - are both CNS modules. Commercial deployment needs a CNS licence from BIOVIA. Wiring it in was an explicit instruction; this note records the obligation, it does not gate the code."

Three consequences follow, stated plainly because a customer or an acquirer will need them:

10.  The code does not enforce this. The note explicitly declines to gate. A commercial run will invoke CNS and produce a number, and nothing in the software will object. The obligation is organisational, not technical.

11.  It is confined to one axis. CNS is reached only through the force-field second opinion on a targeting-head binding pose. Every other axis in this document - the entire quality layer, the ranking, the free-energy battery, the structural prediction, the corona model - is CNS-free. Disabling that one cross-check removes the obligation without touching anything else in the pipeline, at the cost of that axis reverting to a single learned confidence with no independent check.

12.  The remedy is a licence, not a rewrite. A CNS licence from BIOVIA clears it. If that is not obtained, the honest options are to leave the axis unscored, or to replace the force-field opinion with a component under a permissive licence.

A.5 Entries Requiring Confirmation

Two binaries are installed without an accompanying licence file, so their terms could not be verified from this system and are not asserted here.

Table A-2. Components whose licence could not be read from this system.

The second row carries a general caution worth naming: a single statically linked scientific binary can carry the licence obligations of everything compiled into it, and those obligations are not visible from the executable. Establishing the position for such a binary means examining its build manifest, not its filename.

A.6 Summary

The actionable conclusion. For a service operated in-house, the stack is licence-clean with one exception and one condition. The exception - CNS - affects exactly one scoring axis, is documented at its point of use in the source, and is resolved either by a BIOVIA licence or by disabling that single cross-check. The condition - ViennaRNA attribution - costs a credit line. Neither is a reason to alter the platform's architecture, and both are stated here so that no one discovers them during diligence instead of before it.

Disclaimer. All in-silico outputs are predictive heuristics intended for triage and prioritisation, not measured release data. Compendial confirmation by validated methods remains required. Freedom-to-operate output is automated preliminary screening and is not a legal opinion. The limitations in Chapter 10.1 and the failing test groups in Chapter 14.2 are part of this document's factual content and should be read before relying on any specific axis.