CSB — The Mechanistic Oracle | Simplicity Board

csb computes what a perturbation does to a body, organ by organ, from cited mechanism — and refuses, by name, whatever it cannot compute. The refusal is not the caveat; it is half the product.

Why it exists

Ask what a compound does to a body and most systems answer by lookup — this drug, these recorded effects. That answer does not exist for a compound nobody has dosed, it cannot say why, and it cannot tell you what it is missing. csb computes the answer from mechanism instead, and where the mechanism is not in hand it says so by name rather than filling the gap, because a confident wrong number costs more here than a refusal with an address.

How it computes

A drug is (physchem, target binding, PK), never a name. A perturbation routes through one shared PBPK field → target occupancy → an expression gate → a per-organ effect. Nobody types "myopathy"; it emerges because HMGCR is expressed in muscle. The test is blunt — could a lookup table have produced this answer? If yes, the mechanism was decoration.
One shared field, so one changed fact moves every organ at once. Vancomycin oral vs IV is a single bioavailability fact, and it collapses the field everywhere together: the heart drops below the MRSA MIC (no endocarditis cover) while the kidney goes quiet. The safety/efficacy split falls out of the route, not out of a table.
Contrast, not magnitude. A magnitude is NOT_CALIBRATED unless a calibration test says otherwise. The deliverable is the ratio that survives the uncited constants.
A refusal is a result with an address. An organ that cannot be computed says why, by name — expression_unknown, reserve_not_characterised, no_pbpk_compartment — never a silent 0. "We know the steps; we need k_cat for reaction X" is the most useful sentence the platform can hand a scientist.
Two solver worlds, one contract. Mechanistic kinetics (ODE/PDE, PBPK, genome-scale flux balance) and physics-informed ML (PINNs, neural ODEs) implement the same model contract, so a disease can be solved classically and learned, then the two cross-checked.
Every claim re-checkable offline. A caveat ships inside the JSON, not in a docstring nobody reads. Models are sha256-pinned under their real names, and the literature is a ledger of what was actually fetched.

The corpus it computes on

Kinetics cannot say what is connected to what, so underneath the oracle sits its data layer: 59,161,677 typed, weighted, directed couplings over 28 sources, molecule to clinic. csb reads it through one flag, csb effect --with-corpus, and the corpus never reaches back.

Strength and sign stay orthogonal. Negative feedback (TSH↔T4) is coupled and signed. A probability model has nowhere to put the minus sign, so inhibition ends up encoded as a number and one of biology's most reliable couplings reads as weak evidence. STRING supplies coupling unsigned, SIGNOR supplies sign without strength — joined on the same accessions, they fill both fields of one edge.

Dossier	Target · Indication	Regulatory shape
GLPI-103 →	GLP-1 / Apelin dual agonist · Type 2 Diabetes	Full CTD, Modules 1–5 · NDA
TILA-278 →	anti-TL1A × IL-22R bispecific · Ulcerative Colitis	Biologic · Phase 2b package · BLA
OBX-319 →	anti-CD19 × anti-CD20 bispecific · Systemic Lupus	Biologic · Phase 3 submission · BLA

(페이지 작성중; Page under construction) 방명록을 남겨주세요! Please leave a comment in my guestbook!

Why it exists

How it computes

The corpus it computes on

The stack built on it

What makes an answer usable — IR · RA · Operator

Stated plainly

목적

계산 방식

계산의 바탕이 되는 코퍼스

그 위에 올라가는 스택

답을 쓸 수 있게 만드는 것 — IR · RA · 오퍼레이터

밝혀 두는 한계