The arsenal — four regulated product types, one mechanistic oracle

The spine

Connect, don’t merge.

An arsenal where anything may import anything becomes one tangled program with four names. Two rules keep every piece independently useful — and they are not aspirations in a README, they are grep-verifiable facts, re-counted today. This is what lets a bench be adopted on its own, and what lets the operator be swapped out without touching the science.

csb imports no consumerThe mechanistic oracle cannot reach back into the products that use it.

0violations

The map feeds the oracle, never the reversemrm (associative — what should I try?) is consumed by csb (mechanistic — does it work?). Nothing in mrm reaches back, so the evidence graph stays usable on its own.

0mrm → csb

The two oracles

One proposes. One disposes.

They answer different questions and stay separable. mrm is the idea bank — breadth of evidence across 59 million typed couplings. csb is the mechanism — narrow, cited, and willing to refuse. A design that consulted only one of them would be either ungrounded or unimaginative; running both is what turns a shortlist into a decision.

Associative oracle · mrmWhat connects to what

Twenty-eight sources, curated and derived, reconciled into one weighted, typed, directed graph — 172 distinct relation kinds, so an edge says how two things are coupled, not merely that they are. Counted against the store on this disk today.

typed couplings: 59,161,677
distinct sources in store: 28
relation kinds: 172
store size: 1.99 GB

34,181,718open

23,470,039gated

The gate is a column, not a filter bolted on at the end: licence_track is NOT NULL, so an edge that cannot say how freely it may travel cannot be stored at all. export(OPEN) is then exactly 34,181,718 rows of 59,161,677 — a redistributable subgraph you can hand to a partner at 59M scale without a lawyer reading it first.

Identity is resolved, not assumed. Obesity (MONDO:0011122) carries same_as edges to DOID, EFO, ICD9, MEDGEN, NCIT, SCTID and UMLS — eight vocabularies, one node. 244,766 identity edges do that across the graph, so a query written in one namespace does not quietly miss evidence filed under another.

Mechanistic oracle · 3017_csbWhether it actually works

A five-tier epistemic contract decides what may be said. A chain is only as quantitative as its weakest hop, and it never upgrades — pulling a bare number out of a qualitative chain raises rather than returns.

tiers: REFUSED → QUALITATIVE → COMPENSATED → REDUCED → QUANTITATIVE
binding constants, uncited: 0 of 60
affinities refused for want of a citation: 4

Drug–drug interaction · midazolam + ketoconazole

Predicted ×1.57–6.26 sits below the observed clinic ×10–15. csb reports the miss and names the cause — a conservative geometric-mean K_i and no dynamic PBPK — rather than tuning a constant until the number lands. At the conventional 0.015 µM K_i the upper corner reaches ×18.8, so the clinic is bracketed: K_i‑sensitive, not structurally wrong. Both values cited.

The four benches

Four product types. One oracle underneath.

Drug, cosmetic, device, nutrition — each drives the same mechanistic core through its own regulatory and physical reality, so a formulation, a dose ceiling and a statutory route are decided in one pass instead of three teams. Every readout below is a real run, and every card ends with the decision that run actually returns.

Drug`cldd`

cldd explore adc --pareto

design space: antibody–drug conjugate
linker / payload: peptide‑cleavable / auristatin
drug‑antibody ratio: 3.96
therapeutic index: 59.0
antigen selectivity: 0.85
evaluations: 335

What the run returnsA Pareto front over 335 evaluations, not one lucky molecule — linker chemistry, payload, DAR and antigen selectivity moved together until the therapeutic index stopped improving. Delivery is scored as an intracellular-release surrogate, and the card states that in its own output rather than in a footnote.

Cosmetic`clcd`

clcd design --claim whitening

active: niacinamide
concentration: 4.69 %
binding constraint: 고시 notified range
MFDS route: 보고
data waiver: eligible
without the claim: 10.0 % · use‑level bound

What the run returnsThe marketing claim is priced in statute. Declaring 미백 tightens the ceiling from the 10 % documented use level to the 5 % 별표4 band and sets the MFDS route to 보고 with a data waiver — so the cost of a claim is known before the formula is locked, not after a filing comes back.

Device`clmd`

clmd credibility --product-id demo-infusion-01

framework: ASME V&V 40
code verification: passed
model validation: missing
uncertainty quantification: warning
evidence complete: 2 of 5
allowed output: internal review only

What the run returnsA credibility verdict that binds. The bench audits its own evidence against ASME V&V 40, finds 2 of 5 complete at this consequence level, and downgrades the allowed output to internal review — it will not let a result be quoted further than its evidence supports. Enforcement, not a disclaimer.

Nutrition`clnd`

clnd interaction spinach_high_oxalate --drugs warfarin

mechanism: VKOR antagonism
kind: pharmacodynamic
effect factor: 0.65
exposure change: none — null
severity: moderate

What the run returnsThe mechanism, not just a flag: VKOR antagonism, pharmacodynamic, effect factor 0.65, exposure unchanged. So it gives the correct counterintuitive advice — keep vitamin K intake consistent, do not avoid greens — and it refuses four upper-intake verdicts by name where no UL exists to quote.

The loop

What it closes, end to end.

Each engine is useful alone. Chained, they answer the question a design programme actually asks — does this molecule do what I want in a body, and what will the regulator want to see about it — without a human carrying numbers between two tools.

one call

Design → body effect → dossier gap.arsenal_assess scores a molecule on the mechanistic oracle and hands that payload, verbatim and in process, to the regulatory ledger. The chain exists because the hand-assembled version made the model relay a ~3.6k-token payload between two tool calls — a transport limit no prompt fixes. It is registered only when both engines are actually operable.

72 rows

Which studies you no longer have to commission.Korea’s CTD ledger, per product class and per section, with the guidance behind each row: 21 answerable by computation, 7 of them moved by biology the oracle just computed. A study budget planned in advance instead of discovered during review.

tiers

A number arrives with its own epistemic status.Five tiers from REFUSED to QUANTITATIVE; a chain is only as strong as its weakest hop and it never upgrades. Pulling a bare figure out of a qualitative chain raises rather than returns — which is why the figures that do come back can be put in front of a reviewer.

offline

Auditable with the network off.3,234 curated MFDS/NIFDS documents, cut into 198,266 chunks, pinned by sha256 and addressable through 896,222 article-level anchors in a local FTS5 index. A citation you can open is worth more than one that merely sounds right — and nothing here needs a vendor to stay up.

context

The tool surface follows the bill.The full arsenal is 42 tools and ~8,966 tokens of schema before the model reads the question; the composed surface is 7 tools and ~978. A self-hosted model is handed the small one automatically, because on your own box every schema token is paid in wall clock — roughly six minutes of prompt processing per turn at 24 tok/s.

one limit

Magnitudes are for contrast, not calibration.A predicted interval can sit below the clinic — the midazolam–ketoconazole case above does — and csb reports that miss rather than tuning a constant until it lands. Rank and compare with these engines; 51 of the 72 dossier rows still route to a bench, each named with the guidance behind it.

Verification

How this page knows.

Nothing here is quoted from a document. Each figure below carries the command that re-derives it, so a reader can disagree with the page by running it.

mrm

59,161,677 edges · 28 sources · 172 relation kinds.Counted straight out of the store: select count(*) from edges against data/mrm/derived/mrm.duckdb. The licence split — 34,181,718 open / 23,470,039 gated / 1,509,920 share-alike — is a group by licence_track on the same table.

the spine

Two laws, zero violations.Two greps, run today: no csb module imports a consumer bench, and nothing in mrm imports csb. Each returns nothing, against paths that exist — a rule checked on an empty directory proves nothing.

the ledger

72 rows, read from the manifest that enforces them.The sibling page inlines its figures at build time from rapl.readiness.manifest, the product-IR module namespace and the corpus pipelines’ own summaries — uv run python scripts/export_showcase.py in RAPL. No figure on it was typed by hand.