NeuroTopo — architecture-candidate qualification toolkit and evidence base

This repository contains the code, protocols, results and audit records of the NeuroTopo project's architecture-candidate line (ARCH-0001) together with the qualification tooling built while running it.

Status: negative and closed, honestly. The candidate (SC-REL-v3/v4: a recurrent core that performs two hypothetical next-command rollouts and applies a bounded state correction) showed no replicable advantage over a compute-matched strong control on the tasks tested, and the route was closed in the tested scope. The project's stage gates contain no positive architecture claim. What this repository does contain is a reproducible record of that outcome plus a set of checks that repeatedly caught the project's own false positives.

Headline results (all reproducible from the included artifacts)

Result	Verdict
Candidate learning qualification, 15 seeds, fixed budget	PARTIALLY_SUPPORTED — 14/15 passed (one HOLD_COLLAPSE failure retained)
Old one-bit command-response task family	Not discriminative — strong control at 100% on 4/8 pre-registered conditions, and an explicit memoryless rule solves it exactly
DELAY-2 memory-task carrier qualification	INCONCLUSIVE — the pre-registered compute-matched control misses the learnability gate (5-seed mean 0.6455 vs 0.65)
Exploratory candidate-vs-control probe	SUPPORTED / EXPLORATORY (+11.26pp), not A2-creditable: the control used 1.82x the candidate's compute
Independent, compute-matched replication (fresh seeds)	FALSIFIED / NO_REPLICATED_ADVANTAGE — candidate 0.6637 vs control 0.7261; the no-correction arm ties the candidate

Two positive signals were produced and both were retracted by pre-registered checks; the full chain (EXP-0030 -> EXP-0031, EXP-0032 v1 -> v2, EXP-0033 -> EXP-0034) is preserved under research/evidence/ and research/decisions/.

Repository layout

Path	Contents
src/neurotopo/	Package: task generators, models, qualification tooling, loop/runner utilities
tests/, scripts/	368-test acceptance suite, verification/replay scripts, experiment drivers
configs/, experiments/	Frozen run configs and pre-registered protocols (EXP-0001 .. EXP-0034)
artifacts/EXP-0028..0034, artifacts/validation	Full records for the six most recent experiments plus the acceptance suite
research/	Programme: hypotheses, prior-art audits, design docs, decisions (0001-0109), evidence reports, stage gates, packet-only reflection records
paper/	Contribution dossier and venue assessment for the write-up
docs/	Workflow, reproduction and protocol documentation

Large raw artifacts from earlier experiments (hundreds of MB of .npz state/trajectory dumps for EXP-0015..0027 and DYN-0001) are not included; the scripts and configs that regenerate them are. See PROVENANCE.md.

Running the checks

uv sync --frozen --extra dev
uv run --frozen --extra dev python -m pytest -q
uv run --frozen --extra dev python scripts/verify.py

The verification suite writes artifacts/validation/suite.json; the qualification runners refuse to start unless that suite passes and its recorded source hashes match the working tree.

What this repository does not claim

• no positive architecture effect, no Transformer/Mamba comparison, no scaling, language or biological claim;
• no claim that the qualification checklist items are novel phenomena (see research/prior_art/ for the audits and the explicit overlap list);
• the scoped structural result (the strong control's function class contains the candidate's on the tested family) is a statement about that model family and those tasks only.

License and citation

Code: MIT (LICENSE). Documents, protocols, results and audit records: CC BY 4.0 (LICENSE-DATA). Please cite the repository and the specific decision/evidence files you rely on; no paper exists yet.