v2.2 · discovery orchestration

Autonomous Discovery Loop v2.2

How TWOG drives its own scientific crux: it proposes the next falsifying test, pre-registers a hashed kill-criterion before compute, resolves its own inputs, dispatches real GPU work, and audits the result through confound and provenance gates before deciding what to test next.

activeapplies to · candidate cruxesclaims · falsification attempt

What this governs

The loop is falsification-first. For a leading hypothesis it proposes the test most likely to kill it, commits the kill-criterion before any compute runs, and never auto-promotes what survives. A supporting result is evidence the idea is harder to kill, not proof that it is true.

Falsification first. The human write-gate is strictly terminal.

01 · Propose the kill test

Given a candidate, the loop selects the next experiment most likely to falsify the leading hypothesis rather than the one most likely to confirm it.

02 · Pre-register the criterion

A kill-criterion is hashed and committed before any compute runs, so the threshold for failure cannot be moved after the result is known. This is the anti p-hacking gate.

03 · Resolve inputs

From a candidate’s named target and therapy, the loop fetches its own inputs: PubChem SMILES for the compound and an RCSB structure for the target, with no hand-curation.

04 · Dispatch compute

It runs real GPU compute on Modal and ledgers the artifact. In CI it falls back to an in-process mock provider so the loop is reproducible without a GPU.

05 · Gate and terminate

A result must clear mandatory confound and provenance pre-gates before it can be accepted; a supports result cannot be accepted until its known confounds survive controls. The loop then decides what to test next, and the human write-gate remains terminal.

Why falsification-first

A confirmation-seeking engine drifts toward agreeable results. The loop instead proposes the test most likely to kill the leading hypothesis, so the work that survives is the work that resisted a genuine attempt to break it.

Why the criterion is pre-registered

The kill-criterion is hashed and committed before any compute runs. Pre-registration removes the freedom to redefine success after seeing the numbers, which is the main route to p-hacking in an automated pipeline.

How inputs are resolved

The loop starts from a candidate’s named target and therapy and fetches its own inputs — the RCSB structure for the target and the PubChem SMILES for the compound — instead of relying on a curated file someone prepared by hand.

How results are gated

Confound and provenance pre-gates are mandatory. A supports result cannot be accepted until its known confounds survive controls, and verdicts are never recorded as true. Nothing the loop concludes is auto-promoted; the human write-gate is the only path to a record change.

What it does not claim

Everything here is in-silico and hypothesis-generating. A surviving hypothesis is not a validated treatment. The bar the loop optimizes for is integrity and reproducibility, not clinical proof.

What a reader can verify

What is being tested: The leading hypothesis for a candidate and the falsifying experiment chosen to attack it.
Pre-registered criterion: The hashed kill-criterion committed before compute, with the hash and commit time.
Resolved inputs: The named target and therapy, the fetched RCSB structure, the PubChem SMILES, and their source provenance.
Compute run: Provider (Modal GPU or CI mock), job ID, docking and pose outputs, cost, and ledger reference.
Gate outcome: Confound-control results, provenance checks, the resulting verdict, and the explicit non-promotion record.

How to interpret this method

A surviving hypothesis is harder to kill, not proven true.
A pre-registered kill-criterion cannot be revised after the result is known.
A supports verdict is invalid until its known confounds survive controls.
No loop result is auto-promoted; the human write-gate is terminal.
In-silico survival is hypothesis-generating, not clinical validation.

What this method does not certify

Autonomous Discovery Loop v2.2 does not certify efficacy, safety, or clinical readiness. It runs reproducible, pre-registered falsification attempts in silico and leaves every promotion decision to a human operator.