The SIM Framework

What this surface is

The Science Report for gate_20260917T182405Z, served as written. It is an authored analysis of one completed run: the gate at a glance, the three-pillar argument, findings for each of the 10 claims, package-level findings, and what the report cannot establish. A human committee nominates the Development Candidate; SIM emits no go/no-go and neither does this report.

It is deliberately not Provenance, which carries per-claim forensic depth -- numeric provenance, dropped replicates, digests, citation resolution -- and recomputes every figure it shows at build time.

How to read A and B

Every claim is run through an adversarial loop, 5 times per claim. Each repetition opens two positions and an arbiter adjudicates between them.

A verdict of A means the arbiter found the constructive case carried; a verdict of B means it found the skeptical case carried. Neither is a recommendation, and neither is a decision.

Majority verdicts observed on this gate: A, B.

1. What this report is, and what it is not

This report presents the adjudicated reasoning and the evidence SIM assembled for ten claims at the Development Candidate gate, for a synthetic oral small-molecule NLRP3 inflammasome inhibitor in Alzheimer's disease.

Six statements govern how every page below must be read.

SIM emits no go/no-go, and this report emits none. The package's own posture field says it: SIM produces the adjudicated reasoning and the evidence for the human committee that nominates a Development Candidate and commits it to IND-enabling studies. Nothing here is a nomination decision or a recommendation toward one.

The candidate is fictional and its readouts are synthetic. Every readout in this gate carries synthetic: true. The candidate was authored with deliberate liabilities, so a skeptical majority is the designed outcome of the exercise and is not a finding about a molecule.

The precedents are real and every statement about them is a pointer to a record, not an authority. The adjudication characterizes real programs' pharmacology, exposure and safety in order to test shared basis. A reader must not treat any statement here about a named real drug as authoritative; read it at the source.

Verdict counts are over all five repetitions of each claim. Quoted prose is not. The verdict sequence, the low-confidence sequence and the degradation record are carried for every repetition. The arms' own assessments were read for two repetitions of the centerpiece claim -- the representative one and the one that dissented -- so any statement about what an arm argued is over the repetitions read, never over all fifty.

A verdict is not a result. Verdict spread across repetitions is information for the scientist and is published as a readout. No per-claim figure in this report is a result about the molecule.

SIM does not certify that the claim set is complete. There is no canonical dataset for this gate. The ten claims are a stated subset of a Candidate Drug Target Profile the demonstration does not have, and their adequacy is not a SIM output.

2. The candidate, and the decision the claims ask about

The decision is to nominate the molecule as a Development Candidate and commit it to IND-enabling GLP toxicology. It is not the decision to run an in-vivo efficacy study; by the time a nomination is written, in-vivo efficacy is an input to the package rather than the thing being decided. This gate is defined by companies and by contracts, not by any regulator, and what a nomination must contain is drawn from the program's own Candidate Drug Target Profile.

The candidate's authored load-bearing flaw is a thin chronic-dosing safety margin. The readout states it in numbers: a projected efficacious unbound plasma concentration of 312 nM against a lowest-effect exposure of 1060 nM, a 3.4-fold margin, established in a 14-day repeat-dose rat study with day-14 toxicokinetics. No non-rodent repeat-dose study has been run and no non-rodent species has been selected. The finding at the lowest effect exposure is dose-dependent transaminase elevation in the high-dose group, reversible on washout. The intended clinical regimen is chronic daily oral dosing in an elderly Alzheimer's population for twelve months or longer.

The readout's own note is the whole problem in one line: the margin is a 14-day, single-species, plasma-compartment figure read against dosing intended to run for years.

3. The gate at a glance

Every value in this table is copied from the package. SEQ is the verdict sequence in repetition order, where A is the constructive arm's position and B the skeptical arm's. QUALITY is the build log's own read of the evidence, not a SIM output about the molecule.

Ten claims requested, ten resolved, none failed, none excluded. Four carry resolved_low_conf. The gate ran against fred 0.3.0 on wire 1.2 with the configured model tag kimi-k2.6:cloud, a cloud-served route, on engine commit 8ddbf9aa with a clean working tree.

Two claims have no majority verdict at all and are still reported as resolved. cns_penetrance split two to two with a null verdict in repetition four; developability split two to two with an explicit unresolved in repetition two. Both carry caveat_reason: no_strict_majority. A reader who takes status: resolved at face value on those two rows will take a tie for a conclusion.

4. The three-pillar argument, which is the thing to look at first

Pfizer's three pillars -- free drug at the site of action, target binding, and expression of pharmacological activity from that site -- are what a Development Candidate package has to be able to argue in the compartment that matters. For a CNS indication, that compartment is the brain.

Pillar one, free drug in brain: inferred, not measured. The candidate has a measured rodent Kp,uu of 0.22 with a total brain-to-plasma ratio of 0.6 and a P-gp efflux ratio of 4.8 from MDCK-MDR1. Those are measurements. The projected efficacious brain unbound concentration of 69 nM is not: it is the plasma figure multiplied by the partition coefficient. The package records no cerebrospinal fluid concentration of the candidate at any dose or timepoint, in any species.

Pillar two, target binding: measured in blood. Potency is a whole-human-blood IL-1beta release IC50 of 17.7 nM, which evidences engagement in blood. The declared secondary assay -- primed macrophage or microglia IL-1beta -- was never reported. The IC50 itself rests on two of four replicates after two exclusions, recorded as cv_over_threshold and ref_cpd_out_of_range. The central_target_engagement readout states the position without hedging: central engagement status is not demonstrated, with no CSF drug exposure, no CSF pharmacodynamic biomarker, and no occupancy measurement.

Pillar three, pharmacology expressed from the site of action: an effect at a dose, not at an exposure. The in-vivo pharmacology readout records reduced brain IL-1beta and IL-18 with partial cognitive rescue in one amyloid precursor protein overexpressing transgenic model, unreplicated, with no power calculation, pre-specified effect size or group sizes recorded, and no result in a knock-in model. The unbound brain concentration achieved at the effective dose was never measured. So pillar three cannot be joined to pillar one: the package cannot state what free brain concentration produced the pharmacologic effect it relies on. The translational fluid biomarker -- the endpoint that would have carried into a clinical program -- is trending only and does not reach significance.

And the argument is not written down. The human_dose_projection readout records three_pillar_argument.written_down: false, alongside no predicted human clearance, no allometric or PBPK basis, no projected clinical dose or regimen, and no statement of fractional target coverage over a dosing interval in either compartment. The species basis of the carried efficacious exposure is itself unstated -- the package does not record whether 312 nM is a mouse exposure, a rat-scaled exposure or a projected human exposure.

That last point deserves its own sentence, because it propagates. The safety margin is a ratio, and the numerator's species basis is unrecorded. A 3.4-fold margin computed between a rat tox exposure and an efficacious exposure of unstated species is a number whose meaning is not established by the package that carries it.

5. Per-claim findings

5.1 chronic_safety_margin -- B 4-1, evidence weak

The centerpiece, and the claim the candidate was authored to fail on. The skeptical arm carried four of five repetitions.

The constructive arm's case, read at repetitions one and five, is that a 3.4-fold margin with a reversible, dose-limited, monitorable transaminase signal is sufficient when read against the NLRP3 class safety record; that the signal is distinguishable from the furan-ring hepatotoxicity that halted MCC950, a motif the candidate lacks; and that dapansutrile, a direct NACHT-domain inhibitor sharing the mechanism, has been dosed to 2000 mg for 14 days in heart-failure patients without significant hepatotoxicity.

The skeptical arm's case is that the comparison is broken rather than weak. Dapansutrile's tolerability was established in non-chronic, primarily peripheral settings; the longest published human exposure for a CNS-penetrant NLRP3 inhibitor in neurodegeneration is twelve weeks, most are twenty-eight days or less, and no NLRP3 inhibitor has ever been dosed in an Alzheimer's population. Fourteen days in one species addresses neither duration-dependent toxicity nor species translation, and no exposure multiple addresses the mechanism-based risk of chronic innate-immune suppression that the criterion flags.

The arbiter sided with the skeptical arm on shared basis rather than on plausibility, and said so in those terms across four repetitions: the class record offers category alignment but no shared basis for a chronic safety margin, and the evidence reaches the subchronic-tolerability mark but not the chronic-safety-margin mark.

The one dissent, repetition five, is the most interesting repetition in the gate. Of the two repetitions whose arm records were read, the constructive arm made no tool calls in repetition one and nine across four iterations in repetition five -- and repetition five is the one it won. Repetitions two through four were not read at that level, so this is a pairing observed twice, not a pattern established over five. The arbiter in repetition five accepted that MCC950's halt is attributed to a furan motif the candidate lacks, that dapansutrile and NT-0796 extend tolerability into neurodegenerative populations, and judged that the skeptical arm overreached. That is a verdict flip that tracks retrieval effort, not new evidence, and it is a caution about reading any single repetition as the answer.

The build log's own next step names the remedy: 90-day rodent and non-rodent repeat-dose hepatic safety margins, measured unbound brain exposure in the toxicology arm, and chronic human NLRP3 safety data.

5.2 target_rationale -- A 3-2, low confidence

The only claim the constructive arm carried. The liability is that the human genetic argument does not reach the target: NLRP3 itself is not an established genome-wide significant Alzheimer's risk locus, and the package's human-genetic support runs through microglial modifiers -- reduced INPP5D activity inducing inflammasome formation in human iPSC-derived microglia, and the protective PLCG2-P522R variant reported to suppress it. That establishes the pathway is disease-relevant. It does not establish that inhibiting this node is beneficial.

The build log flags three citation problems on this claim, and they are not cosmetic: a PMID cited for MCC950 protection in an APP/PS1 model reports Parkinson's pharmacology rather than Alzheimer's; one PMC identifier resolves to an unrelated knee-ligament paper; a second PMC identifier could not be retrieved. Repetition four lost an arm and is recorded low confidence.

5.3 target_engagement -- B 5-0, evidence weak

Unanimous for the skeptical arm, with the build log rating citation quality strong. The objection is compartmental rather than numerical: 17.7 nM in whole blood is engagement in blood, the declared microglial assay was never submitted, and against a Kp,uu of 0.22 the inference to central engagement is a judgment rather than a measurement. The build log's next step is a human microglial engagement assay or a CSF pharmacodynamic readout.

5.4 central_target_engagement -- B 4-1, low confidence

The claim with the starkest record: no central measurement of any kind. The build log's note is worth reading as written -- peer precedents NT-0796 and VENT-02 provide robust human CSF data, and that data highlights the gap rather than bridges it. Two of five repetitions ran one-sided after an arm failed.

The readout also records why occupancy is not an easy substitute here: CNS tracers for this mechanism are reported preclinically and in non-human primates, none is clinically validated, and the target is hard to image because expression is low and largely myeloid. CSF drug exposure paired with a CSF pharmacodynamic biomarker is the practicable route for this class.

5.5 in_vivo_pharmacology -- B 3-2, evidence weak

Split three to two. The build log records that the constructive arm omitted the clinical trial identifiers and a key PMID entirely, that the curated negative record on the absence of positive controlled neurodegenerative efficacy went uncontested, and that the constructive arm's use of dapansutrile as direct support despite its superior exposure verification is a comparability misalignment. That is the named failure mode of this field -- precedent used as unearned support -- detected and named per claim.

5.6 selectivity -- B 4-1, comparability call counts

One of two claims where the comparability call is a clean counts with no caveat. Selectivity itself is not in doubt: 17.9 nM at NLRP3 against greater than 10,000 nM at AIM2, NLRC4 and NLRP1. The liability is a modest hepatic-transporter activity that must be weighed alongside the hepatic safety readout. The build log notes that both arms agree on the numbers and diverge on interpretation and precedent comparability, and that the candidate's own IC50 values were not independently retrievable because the molecule is synthetic.

5.7 cns_penetrance -- no majority, 2-2 with one null verdict

The claim the comparability basis was actually authored for, and the only one with a null verdict. Repetitions one and two went to the skeptical arm on the absence of CSF exposure as the translatable surrogate the criterion dimension names; repetitions three and five went to the constructive arm on a stage-appropriate standard -- that Kp,uu and efflux ratio are the standard nonclinical characterization at this gate and that human CSF exposure is a Phase 1 readout rather than a DC-gate prerequisite. Repetition four produced no verdict at all.

That disagreement is not noise. It is two defensible readings of what the gate requires, and it is exactly the kind of question the package exists to put in front of a committee rather than resolve.

The build log's citation notes on this claim are unusually pointed: it identifies redundancy in the constructive arm's citations, records that the usnoflast ALS Phase 2a result is reported but not confirmed as published, that the claim it missed its primary endpoint is unverified by search, and that one of the skeptical arm's registry identifiers returned no matches. It also observes that the two negative precedents the skeptical arm relies on appear to be curated records without independent verification -- which is true, and is the subject of section 6.2.

5.8 admet_pk_hepatic -- B 5-0, comparability call counts

Unanimous, and the second clean counts. Oral bioavailability 41.9 percent, half-life 6.5 hours, moderate metabolic stability, and the mild reversible transaminase elevation that mirrors the real class caution. The build log records that the cited abstracts verify NLRP3 biology and general dapansutrile tolerability but do not verify the specific hepatic-safety claims asserted: the MCC950 citation does not mention furan-ring hepatotoxicity or clinical transaminase elevation in its abstract, and the dapansutrile citation mentions general safety without discussing hepatotoxicity. So the structural-differentiation argument that the constructive arm leans on across two claims is unverified at the record-fragment level.

5.9 developability -- no majority, 2-2 with one explicit unresolved

The other tie. The liability is that the exposure any in-vivo package rests on depends on an enabling amorphous dispersion that has not been made; an amorphous form is metastable and can recrystallize, so exposure achieved in an efficacy study may not be reproducible in the form carried into toxicology, and dose escalation may be form-limited rather than tolerability-limited.

Repetition two is the only repetition in the gate where the arbiter returned an explicit unresolved, and its reasoning is the cleanest statement of the failure mode in the whole package: the class precedents establish that oral CNS delivery of NLRP3 inhibitors is feasible, but none addresses the candidate-specific dependency on an unmanufactured metastable dispersion, so the precedents establish a viable path without bridging the specific gap. Repetition five sharpens it -- the retrieved record contains no data on the precedents' own P-gp status, CNS-MPO scores, formulation requirements or solid-state stability, so they cannot serve as direct comparators for a formulation question.

Repetition one lost an arm: the skeptical arm failed its retry ladder, and the arbiter judged the constructive case alone, saying so explicitly before doing it. This claim has the weakest citation quality in the gate, with one PMC identifier returning empty and another resolving to an unrelated 2026 systematic review on coronary angiography.

5.10 human_dose_projection -- B 5-0, low confidence, evidence weak

Unanimous, and the claim where both arms agree on the facts. The build log records that both positions anchor to the same acknowledged gaps and that no citation was asserted and found absent by search. There is no human dose projection, no written three-pillar argument, and no species basis for the carried efficacious exposure. Two of five repetitions ran one-sided.

This claim connects the others. It is the artifact that ties exposure, engagement, pharmacology and margin into one argument, and the package does not contain it.

6. Package-level findings

6.1 The evidence route: 1,398 citations, and the corpus carries them

Across the ten claims the package accounts for 1,398 citations. Curated records dominate on every claim without exception -- 126 of 167 on the centerpiece, 158 of 178 on target rationale, 101 of 140 on dose projection. Retrieved-and-evidenced citations range from zero on target rationale to 39 on human dose projection.

This is consistent and it matters for how the demonstration is described: live retrieval corroborates, and the curated corpus is the dominant evidence source. It is also a lower bound, for a reason the package states itself -- fred caps its tool-result report at 503 bytes after reasoning completes, so an identifier the model genuinely read can be absent from every payload SIM holds. The retrieved bucket is a lower bound and the unevidenced bucket an upper bound, always, never the reverse.

Retrieval effort varied widely and unevenly: 26 tool calls on cns_penetrance against 56 on target_engagement. Free-text searches returned zero far more often than accession lookups -- on the centerpiece, 15 of 28 free-text calls returned nothing while all 12 accession-only calls returned something. An empty search is not an absent literature, and the zero rate for the query form is the number to read a negative claim against.

6.2 The negative-precedent records are argued and cannot be attributed

Three authored negative-precedent records travel in this package: that no NLRP3 inhibitor has been dosed in Alzheimer's disease, that no positive controlled neurodegenerative efficacy exists for the class, and that the longest recorded human exposure is twelve weeks. They are the skeptical arm's strongest available argument, because a corpus assembled by curating papers cannot contain the proposition that no paper exists.

All three are recorded untestable on all ten claims, for the same reason: their public_ref names no parseable accession. The supplied-citation accounting is explicit that untestable is not not-cited -- it means the item could not be tested at all, and filing it as not-cited would assert we looked and found nothing when the truth is we could not look.

So the package reasons on three propositions it cannot attribute. The arbiter relies on them by name repeatedly -- on the centerpiece it calls them accurate within their stated scope and says they undermine the comparability a chronic safety claim requires -- while the citation accounting reports all three as untestable everywhere. This is a curation defect, it is the same defect the previous science report named, and it is still present.

6.3 The two arms read every precedent differently

On every one of the ten claims, deterministic_diff.shared_equal is empty. Not one precedent was rated the same way by both arms on any claim. The divergence counts run from two of two precedents on selectivity to ten of ten on target rationale, and the recurring pattern is that the constructive arm rates the comparability basis precedent yes or with_caveat where the skeptical arm rates it no.

The claim-level detail is more useful than the count. On the centerpiece, the constructive arm alone relied on four precedents and the skeptical arm alone on three, two of which are the authored negatives. On target_engagement the constructive arm introduced none of its own and the skeptical arm five. On human_dose_projection all three authored negatives are rated with_caveat by the constructive arm and no by the skeptical arm -- the two arms disagreeing about whether an absence counts.

6.4 Degradation, and one repetition that is not counted

Six of the fifty repetitions ran one-sided after an arm failed its stage gate: one on target_rationale, two on central_target_engagement, one on developability, two on human_dose_projection. Each is recorded with path: sentinel_one_sided, cause: one_side_failed, and low confidence set. The degradation record is honest about what it is -- the cause is the route the repetition took, not a diagnosis, and it never says whether the surviving verdict is sound.

One repetition is missing from that accounting. Repetition four of cns_penetrance produced a null verdict and null arbiter notes, yet that claim's n_degraded is zero and every one of its repetitions is recorded as a clean two_sided path with cause: none. The claim's counts_no_verdict is 1, so the non-verdict is visible in one field and invisible in the other two. A reader auditing degradation on the degradation record will not find it.

6.5 The arbiter's own read diverges from the collapse, on every claim

Every claim carries at least one fred_read_vs_collapse_divergence finding in c1_findings: fred's in-trace judgment that a precedent counts, or does not count, diverges from the collapsed comparability call. Five claims carry two such findings, one in each direction, and five carry one. This is recorded rather than reconciled, which is the right behavior, but it means the comparability call a reader sees is not always the call the reasoner made.

6.6 The comparability basis is authored for one claim and used by ten

The gate carries a single gate-level finding: the comparability basis is build-team authored and carries no independent SME validation. It is worse than that in a specific way that the package states plainly. The basis was authored for cns_penetrance, and every other claim runs against it marked basis_alignment: shared_basis_not_claim_specific. Their surfaced precedent is not independently comparability-justified for the claim it was used on.

Every claim in this gate surfaces the same precedent, precedent_pc7b071. Nine of ten claims are therefore adjudicated for comparability against a basis written for a different question.

7. What the package says about itself

The package carries its own limits blocks, and they are unusually strict. They are worth summarizing because they constrain every number above.

The citation partition has four buckets, not three, and the fourth is not a rounding error: unparseable means no accession could be derived and membership was never tested. not_evidenced is not a claim that the model did not retrieve something, because of the 503-byte reporting cap. retrieved_evidenced is a substring test and is an upper bound on its own bucket.

The retrieval profile says what SIM asked for and whether anything came back, and says nothing whatever about what the literature contains. at_reporting_cap marks a claim read off bytes SIM never saw as unverifiable -- it does not verify one, and it cannot.

The supplied-citation record says whether an item was cited and explicitly refuses to emit a per-item match count, on the grounds that such a figure is a relevance score wearing a different name. Nothing in the package ranks, scores, weights or orders evidence.

8. Defects and observations owed

Recorded here so they are visible; none is closed by this report.

1. The negative-precedent records cannot be attributed (6.2). Three authored negatives, untestable on all ten claims, relied on by name in the adjudication. A curation defect in the public_ref field, not a reasoning defect.
2. A non-verdict is invisible to the degradation record (6.4). cns_penetrance repetition four: null verdict, n_degraded zero, path recorded clean.
3. The claim-level low_confidence key does not exist. It is absent on all ten claims and the fact lives in the status string as resolved_low_conf. A consumer querying the key counts zero where the correct count is four.
4. Two claims with no majority verdict are reported resolved. cns_penetrance and developability are two-two ties. The caveat is carried in caveat_reason, not in the status.
5. Broken and mis-resolved citations reached the adjudication. One PMC identifier resolves to an unrelated knee-ligament paper on one claim and to a coronary angiography review on another; a second PMC identifier could not be retrieved on two claims; one PMID cited for Alzheimer's pharmacology reports Parkinson's.
6. The structural-differentiation argument is unverified at the record-fragment level (5.8), while being load-bearing for the constructive arm on both the safety-margin and hepatic claims.
7. A possible association between retrieval effort and the centerpiece verdict flip (5.1). In the two repetitions read, the constructive arm made zero tool calls and lost, then nine and won. Three repetitions were not read at that level, so this is a pairing observed twice and not a pattern; establishing one would need the remaining repetitions read and then a designed comparison.

9. What this report cannot establish

It cannot say whether the ten claims are the right ten. There is no canonical dataset for this gate and the claim set is a stated subset of a profile the demonstration does not have.

It cannot say whether the curated records are scientifically right. No SME has reviewed them. The comparability basis is itself sme_validated: false.

It cannot say what the arms argued across all fifty repetitions. Two repetitions of one claim were read in full for the arms' prose; the rest of this report's per-claim narrative rests on the arbiter's notes, which are carried for all fifty, and on the build log's quality read.

It cannot tell a live gate from a porous one. Where IND-enabling work has already been started at risk, a nomination review ratifies a commitment already made, and nothing in this package establishes that this gate is a real kill point.

And it is n=1: one corpus, one model, one candidate. The results are a case study with internal validity only. Generalization requires a different study, not a better run.

10. Provenance

Compiled from ~/sim_recon/archive/gate_20260917T182405Z, whole-gate evidence package md5 34234cecc4f96c6b6785f3d6be928935, manifest md5 853a42366fd5bbb0eadbf9c81a5ac712, and the ten per-claim trace_sinks.json files under outputs/. Read via extract_gate_slices_v2.py md5 67f7bd002df8a85c47d3b96d2da4b338, which copies fields and computes nothing except two counts it labels.

The gate predates ruling R428, so its manifest carries no schema and no config block. Runs from the next gate onward record the frozen configuration they executed under.

The run label for this gate is the operator's to set in the register; this report is titled by gate id because the archive holds an earlier science report under a run name and two documents with one name is a worse outcome than a long one.

Carried limitations

Every heading below is a limitation this demonstration carries. The reasoning behind each one is a click away; the limitation itself is not.

1. SIM emits no go/no-go
   
   What this means

   Nothing in this interface is a recommendation. A human committee nominates the Development Candidate and commits the program to IND-enabling studies.
2. Every readout was authored with a liability, so a skeptical majority is the designed outcome
   
   What this means

   A readout produces a claim only if it carries an authored liability; a readout with no authored liability produces NO CLAIM. All ten claims here therefore exist because a flaw was written into the readout behind them. THE PACKAGE CONTAINS NO CLAIM WHOSE EVIDENCE IS CLEAN, so it has no positive control, and the verdict distribution cannot by itself distinguish a correct adjudication from a biased one. Read the adjudications, not the tally.
3. The verdict spread, and what it does not measure
   
   What this means

   Ten claims at five repetitions. Across the fifty: 36 found for the skeptical reading, 12 for the constructive, 1 was unresolved on the record and 1 returned no verdict at all. A verdict of B is not a finding that the candidate fails; it records that on that repetition the skeptical reading was better anchored. SPREAD PARTLY MEASURES THE ARMS' OWN STANDARD-SETTING RATHER THAN THE EVIDENCE: every claim is named for sufficiency, sufficiency is relative to a standard, and the arms are deliberately given the criterion's dimension and no threshold. The claims are presented in run order and are never sorted by verdict, because a sorted table reads as a ranking.
4. One repetition returned no verdict, and no field records why
   
   What this means

   On cns_penetrance, one repetition of five carries a null verdict. Both opening arms completed and both are recorded with their citations; the arbitration returned nothing at all -- no verdict, no cited set, no reasoning. THE REPETITION IS NEVERTHELESS RECORDED AS CLEAN: its path reads two-sided, its confidence flag reads normal, and the claim's degradation record reports zero degraded repetitions with the cause of every one of the five reading none. That record's declared scope is which ARM failed its own stage gate, so it is truthful and it is blind here -- THE ARBITRATION STAGE HAS NO DEGRADATION COVERAGE AT ALL. The claim's own no-verdict counter does record the loss, so it is visible in the tally and absent from the record that exists to explain it. Why the arbitration returned nothing is not in the package and is not asserted here.
5. Two claims are reported resolved with no majority verdict
   
   What this means

   cns_penetrance and developability each split their five repetitions evenly. Both carry a no-strict-majority caveat, both record a null majority verdict, both carry no evidence band at all, AND BOTH ARE REPORTED WITH STATUS RESOLVED. Resolved here means the loop completed, not that the repetitions agreed. A reader taking it as a finding about the evidence would be reading the wrong thing, and these two claims are where that misreading costs most.
6. Four claims are resolved carrying a low-confidence flag
   
   What this means

   target_rationale, central_target_engagement, developability and human_dose_projection each contain at least one repetition where one arm failed its stage and the surviving arm was read against a sentinel. Six repetitions of the fifty took that path and all six carry the flag; the other forty-four ran two-sided. The flag marks a claim CONTAINING a weak repetition, not a weak claim, and the sentinel trace names which side failed. The centerpiece claim is not among them.
7. Read-vs-collapse divergence, on every claim
   
   What this means

   The engine's in-trace self-assessment and C1's distribution-level collapse disagree somewhere on all ten claims: fifteen findings, every one of the same kind, with five claims carrying two. This is a named calibration question rather than a defect. The count is stated because an earlier version of this entry put the divergence on eight of ten and named two claims as carrying no finding at all, and neither is true of this gate.
8. C1 keys on more than one ground, and the band is only one of them
   
   What this means

   Eight claims take counts-with-caveat and two take counts. Four band weak, four band moderate and two have no band at all. THOSE TWO DO NOT LINE UP: the four moderate claims split two and two across the two calls, so a reader inferring the call from the band is wrong on all four of them, and the two claims with no band still take counts-with-caveat. Grouped on the full five-key verdict counts, ten claims return six distributions between them and TWO OF THOSE GROUPS HOLD MORE THAN ONE CLAIM, covering 6 of 10: A=0, B=5 is shared by target_engagement, admet_pk_hepatic and human_dose_projection; A=1, B=4 by chronic_safety_margin, central_target_engagement and selectivity. MEMBERS OF A GROUP DO NOT THEREBY SHARE A CALL OR A BAND, so a reader inferring either from the verdict counts is comparing draws rather than independent evidence. THE GROUNDS ARE NAMED IN THE PACKAGE AND ARE NOT ONE: majority_quality_weak_or_absent, minority_low_confidence and no_strict_majority are all live on this gate, and one claim carries two of them at once. Each claim's own caveat reason is the authority on its own ground rather than this paragraph.
9. Unanimity here is reproducibility, not corroboration
   
   What this means

   three of ten claims return the same verdict in every one of their five repetitions. The five draws behind any one claim share one model, one prompt and one evidence base, so agreement measures how reproducibly that evidence is read, not how independently it is corroborated. A count is not a tally of independent judgements, and on this gate it is not even a tally of judgements: one repetition did not resolve at all and one returned no verdict, so NOT EVERY DRAW RESOLVED.
10. Breadth is bounded
    
    What this means

    ONE precedent record is surfaced against ONE comparability basis for all ten claims. That basis is build-team authored and carries no independent SME validation, which the package records as a finding of its own. Nine of the ten claims carry a basis alignment stating that the basis is not claim-specific.
11. Negative precedent was reached and cannot be attributed
    
    What this means

    Three negative-precedent records were authored for this gate, carrying the propositions that no molecule of this mechanism has been dosed in Alzheimer's disease, that none has a positive controlled efficacy readout in a neurodegenerative indication, and that the longest published central exposure is measured in weeks. A corpus assembled by curating papers cannot contain the proposition that no paper exists, so these had to be authored. The reasoning arms relied on them by name and the adjudication argued about their scope. THE ACCOUNTING NEVERTHELESS REPORTS ALL THREE AS UNTESTABLE ON ALL TEN CLAIMS, because each record's public reference names no parseable accession. Against 56 supplied items recorded as cited and 54 as not cited, thirty entries are untestable. The records were used; the provenance layer cannot attribute them. That is a defect in the reference format, not in the records.
12. Citation coverage is bounded
    
    What this means

    The package carries 1398 citation rows: 1169 resolve to the curated corpus, 132 were evidenced by a recorded tool result, 93 ARE EVIDENCED BY NEITHER, and 4 do not parse. THE THIRD BUCKET IS THE SAFETY-RELEVANT ONE -- an identifier neither in the corpus nor returned by any recorded search. It is spread unevenly rather than concentrated: four claims hold 85 of the 93 and three claims hold none at all. All four unparseable rows sit on a single claim. THESE ARE IDENTIFIERS, NOT RECORDS: two identifiers in different namespaces can name the same paper and nothing in this pipeline resolves that. Across the ten run directories the two opening arms cite 17 distinct PubMed identifiers, which is the one figure in this entry an instrument derives: build_citation_metadata.py owns it, the bake stamps it, and a probe checks this page against it. The arbitration cites further identifiers that neither arm cited and the package holds more again than are cited anywhere; NEITHER OF THOSE COUNTS IS STATED HERE. The retrieved bucket is a LOWER bound and the unevidenced bucket an UPPER bound, always, because an identifier appearing only in a query the system issued is not thereby retrieved.
13. The arbitration introduces identifiers neither arm cited
    
    What this means

    On 12 of the 50 repetitions the arbitration cites at least one identifier that appears in neither opening argument, 8 distinct identifiers in all. SIM's own composition has no retrieval or ranking stage -- both arms are handed the same fixed menu on every repetition -- but the inference service issues its own literature searches while it reasons. Whether such an identifier came back from one of those searches or came from model weights is not resolved here.
14. Resolution is not verification
    
    What this means

    A resolving identifier proves a record exists. It does not prove the record supports the proposition it was cited for. Existence is machine-decidable and correspondence is a human judgment, and that judgment has not been made anywhere in this package.
15. The citation sidecar reaches PubMed only
    
    What this means

    17 distinct PubMed identifiers are cited across the ten run directories and all 17 resolved, with none not found and none failing lookup. THAT IS A STATEMENT ABOUT PUBMED IDENTIFIERS AND NOT ABOUT EVERY IDENTIFIER THE PACKAGE CITES: accessions in other namespaces are outside the resolver's reach and are counted nowhere in that figure.
16. Arbiter reasoning is retained in full, not audited
    
    What this means

    The package keeps per-repetition arbiter prose for 49 of the 50 repetitions run. The missing one is the repetition whose arbitration did not return, so the gap is where the loss was rather than where retention chose. Completeness of retention says nothing about the quality of the reasoning retained, and nothing here checks it.
17. The profile carries more criteria than there are claims
    
    What this means

    The pre-registered candidate profile holds fourteen criteria against ten claims. Four carry no claim: a CNS safety-pharmacology position including seizure liability, exposure in the intended population, a structurally independent backup compound, and freedom to operate. They are carried so the record holds criteria with NOTHING ADJUDICATED AGAINST THEM. Regulatory, intellectual property and commercial own no claim. NO COVERAGE FRACTION IS PUBLISHED: counting coverage against a stated floor is certifying completeness as a fraction, so the omissions are named instead.
18. The seed pin covers manifested files only
    
    What this means

    It fingerprints the files listed in the committed manifest, not a whole-tree scan, and the pin and the manifest agree on the count. The manifest lists 45 of the 45 files the seed holds. Only 8 of the cited identifiers appear anywhere in that seed; the rest appear nowhere in it, so the pin bounds the corpus and bounds NOTHING about the evidence the repetitions actually reached. Staleness is detectable within that scope and not beyond it.
19. Candidate readouts are synthetic; precedents are real
    
    What this means

    The candidate molecule and its measurements are fabricated for this demonstration and NO SME HAS REVIEWED THEM, nor the profile's criteria and thresholds. The surfaced precedents are real public records, and every statement about a named real molecule is a POINTER TO A RECORD rather than an authority: it must be read at the source. Reading the candidate data as real measurements misreads the entire artifact.
20. Engine value claimed is capability presence
    
    What this means

    The expert key is grade-lifted rather than firewalled by an independent SME, the outstanding SME ruling is externally blocked, and the key is authored against an older claim set than this gate ran. Any comparative headline remains circular.
21. The convergence-amplifier critique is load-bearing
    
    What this means

    The Scannell and Rogozinska argument that repeated model agreement amplifies rather than corrects error applies to this architecture and is carried on its own merits.
22. Elapsed time measures something outside this system
    
    What this means

    Generation happens on an inference service whose behavior is not observable from the machine running the composition, and no artifact this pipeline writes counts tokens. Any timing figure therefore measures the service and the network as much as the reasoning. A run that took longer did not thereby think harder, and nothing in this package can tell the difference.
23. SIM cannot tell a live gate from a porous one
    
    What this means

    Where IND-enabling work has already been started at risk, a nomination review ratifies a commitment already made and becomes a documentation exercise. SIM's value is contingent on the gate being a real kill point and it has no way to establish that it is.
24. This demonstration is n=1
    
    What this means

    One corpus, one model, one candidate, one run. What is shown here is a case study with internal validity only: it records what this configuration did on this evidence, and carries no claim about how SIM would behave on another program, another corpus or another model. Generalization requires a different study, not a better run.

Contact

raminderpal@anduril.cc