Provenance
Where every figure came from. Per-claim verdict distribution, surfaced precedent, numeric provenance and citations, with every baked input downloadable and digest-checkable.
What is on this site
Three pages and the files they were generated from. Everything below is reachable from here.
On this page
- What this is -- the gate, the science, and what the system does not do.
- The decision being demonstrated -- what the Development Candidate gate is and why it is expensive to get wrong.
- The science, in plain terms -- NLRP3, Alzheimer's, the candidate molecule, and the flaw built into it. Includes a glossary.
- How the results were produced -- the reasoning architecture: what goes in, the adversarial loop, and what comes out.
- Findings from the reasoning engine -- deliberately not written yet, and why.
- Carried limitations -- the full set, repeated on every surface rather than linked from them.
The two apps
- Science Report -- The gate committee's readable evidence view.
- Provenance -- Where every figure came from. Per-claim verdict distribution, surfaced precedent, numeric provenance and citations, with every baked input downloadable and digest-checkable.
The files these pages were generated from
Served alongside the pages. Fetch any of them and check it against the digests published on Provenance, which is where the source data now sits.
- evidence_package.json -- 2124172 bytes
- manifest.json -- 148362 bytes
- seed_pin.json -- 580 bytes
- seed_manifest.json -- 10334 bytes
- numeric_index.json -- 38479 bytes
- citation_metadata.json -- 70627 bytes
- qa_context.md -- 999789 bytes
- bake_manifest.json -- the digests every file above is checked against. It is listed separately because it cannot appear in its own file list: the digest would have to be taken over a document that already contained it. It is served, and it is not self-attested.
The decision being demonstrated
A candidate drug spends years in discovery before any of the regulated work begins. Laboratory assays, animal pharmacology and exploratory toxicology all sit inside that stage. At some point a team has to nominate one molecule and commit it to the formal safety program a regulator will read: IND-enabling toxicology run to GLP.
That is the Development Candidate gate. In industry practice and in deal contracts, Development Candidate, Candidate Drug and GLP Toxicology Candidate all name the molecule cleared to enter that work, and some contracts define the selection literally as the decision to start GLP toxicity studies. It matters because it is expensive in every currency a research organization has. The regulated program costs real money, takes months to read out, absorbs the attention of the people who would otherwise be improving the molecule, and carries an ethical weight that no budget line captures. Committing to a molecule that was never going to work burns a year. Declining one that would have worked can end the program.
What kind of question this is, and what SIM assembles
It is worth being precise about the kind of question this is. It is not a verdict on whether the underlying biology is right, and it is not a prediction that the drug will succeed. It is a judgment about one molecule at one moment, made on incomplete evidence, by a committee of people who will disagree with each other and who have to justify the call afterwards.
SIM's job is to assemble what that committee reads: for each question bearing on the decision, the evidence found, the case for and the case against, the precedents that are genuinely comparable rather than merely similar, and the citations behind all of it. The committee decides. SIM does not, and nothing on these pages is a recommendation.
Every figure in this demonstration is computed from the served files and can be checked against a published digest. This section cannot be: it is authored rather than derived, written to make the rest readable by someone who does not work in this field. No figure anywhere rests on it.
The science, in plain terms
Alzheimer's disease is usually described through two hallmarks: amyloid plaques and tau tangles. A third strand of research concerns inflammation. The brain has its own immune cells, called microglia, and in Alzheimer's they are persistently switched on. The open question is whether that inflammation is merely a consequence of the damage or part of what drives it.
NLRP3 sits at the center of that question. It is a protein inside immune cells that behaves like an alarm: when it detects a danger signal it assembles, and the assembled form releases potent inflammatory messengers, chiefly IL-1beta. Amyloid can trip that alarm. If chronic NLRP3 firing is part of what pushes the disease forward, then a molecule that blocks NLRP3 ought to slow it down.
That is the therapeutic bet. It is a live hypothesis rather than a settled result, and it is a popular one -- which is a hazard in its own right, and one this demonstration carries as a standing limitation rather than a footnote.
Why getting into the brain is the hard part, the candidate, and the glossary
Why getting into the brain is the hard part
A molecule can shut NLRP3 down perfectly in a dish and still be useless. The brain is protected by the blood-brain barrier, a tight lining of blood vessels that keeps most substances out, reinforced by pumps -- P-glycoprotein chief among them -- that actively eject foreign molecules back into the bloodstream.
Two measurements decide whether a candidate clears it. Kp,uu compares the free drug concentration in brain tissue against the free concentration in blood; only unbound drug can act, so this is the figure that matters rather than the total amount present. The P-gp efflux ratio says how hard the pump is working against the molecule. A weak Kp,uu on its own is a concern. A weak Kp,uu together with a strong efflux ratio is the combination that experienced teams have learned not to wave through.
The candidate on these pages
The candidate is a fictional molecule: an oral, small-molecule, direct NLRP3 inhibitor designed to reach the brain. It was built for this demonstration with a specific weakness deliberately designed into it, because a decision-support system is only interesting on a hard case. Its weakness is exactly the one described above, and the gate states it in its own words:
On the CNS penetrance / brain exposure (Kp,uu, brain:plasma, P-gp efflux) readout, the noted liability -- borderline unbound brain exposure (Kp,uu ~0.22) with a P-gp efflux liability (ER ~4.8); whether in-vivo central target engagement will be sufficient is the make-or-break translation question at this gate -- is sufficient to nominate the candidate as a Development Candidate and commit it to IND-enabling GLP toxicology studies.
Quoted from the evidence package this page was built from, not written into this section.
Glossary
Every term used above and on the two evidence surfaces.
| term | in plain English |
|---|---|
| Alzheimer's disease | A progressive brain disease. Usually described through two hallmarks, amyloid plaques and tau tangles; a third strand of research, the one this demonstration sits in, concerns chronic inflammation. |
| Microglia | The brain's own immune cells. They carry NLRP3, and in Alzheimer's disease they are persistently switched on. |
| NLRP3 | A protein inside immune cells that acts as an alarm. When it detects a danger signal it assembles, and the assembled form triggers inflammation. It is the target this candidate is designed to block. |
| Inflammasome | The assembled machine NLRP3 forms part of. Putting it together is what converts a danger signal into active inflammation. |
| IL-1beta | One of the inflammatory messengers released when NLRP3 fires. Measuring it is the standard way to tell whether a molecule has shut NLRP3 off. |
| Blood-brain barrier | A tight lining of the brain's blood vessels that keeps most substances out. Any drug for a brain disease has to get past it. |
| Kp,uu | The ratio of free drug in brain tissue to free drug in blood. Only unbound drug can act, so this is the number that matters rather than the total. Around 1 means the barrier is no obstacle; well below 1 means most of the drug never reaches the target. |
| P-gp efflux | P-glycoprotein is a pump in the blood-brain barrier that actively throws foreign molecules back into the bloodstream. An efflux ratio well above 1 means the pump is working hard against the drug. |
| In vivo | In a living animal, as opposed to in a dish. In-vivo pharmacology and efficacy are INPUTS to the nomination package rather than the thing the gate decides; what follows the gate is regulated toxicology. |
| APP/PS1 | A strain of mouse engineered to develop Alzheimer's-like brain pathology. It is one of several such models, and they do not always agree with each other -- which is itself one of the claims below. |
| ADMET / PK | Absorption, distribution, metabolism, excretion and toxicity, and the time course of drug levels in the body. What the body does to the drug, rather than what the drug does to the body. |
| Developability | Whether a molecule can realistically be turned into a manufacturable, stable, dosable medicine. |
| Precedent | A published molecule whose record can be set beside this candidate. On these pages the precedents are real and carry real identifiers. |
| Comparability basis | A written statement of what makes a precedent comparable: the shared mechanism, the shared readout, and what differs. Without one, a precedent is a resemblance rather than evidence. |
Every figure in this demonstration is computed from the served files and can be checked against a published digest. This section cannot be: it is authored rather than derived, written to make the rest readable by someone who does not work in this field. No figure anywhere rests on it.
How the results were produced
Every claim is argued twice, for and against, by two openings that receive the same records. An arbiter reads both, checks the citations and rules, and may decline to rule. That happens repeatedly per claim, and the repetitions are not averaged: their spread is the result.
The reasoning itself is done by a frozen inference service that may search the literature while it works, which is why every citation is recorded as curated, retrieved or neither, and why search queries leave the machine.
The outlined box in the diagram is SIM. The frozen inference service, fred, runs inside it: SIM calls it and never modifies it. The evidence on the left goes into SIM, the package on the right comes out of it, and the human committee sits outside it.
Click the diagram to open it full screen. Press Esc or Close to return.
Generated when this page was built, from the same constants the application runs on, and drawn inline so it follows the page theme and inverts when printed. The repetition count shown is read from the gate rather than written into the picture.
Findings from the reasoning engine
An account of what the engine actually did, written against the gate rendered on this page. Deferred deliberately. An earlier account described a different run, and rather than publish it beside data it contradicts, this section waits for findings grounded in the gate being shown.
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.
- 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. - 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. - 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. - 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. - 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. - 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. - 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. - 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. - 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. - 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. - 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. - 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. - 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. - 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. - 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. - 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. - 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. - 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. - 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. - 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. - 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. - 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. - 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. - 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.
Why the limitations are repeated
The same set is carried on all three surfaces, stated beside the claims and the provenance it qualifies rather than linked from them, because a limitation a reader has to go looking for is a limitation that will not be read. It is repeated rather than summarized: what is bounded, what is contingent, what is not yet independently validated, and what is a named open question rather than a result.
This page carried a pointer to the other two surfaces until the convergence-amplifier entry made the case against it. That entry qualifies the architecture described here rather than any single claim, so a reader who stopped at this page would never have met it.