Repo Alignment Plan#
Current State#
The repository is the neurotwin Python research package and CLI. It already contains the leakage-controlled Neural Translation infrastructure, prepared-manifest training and evaluation, claim gates, model cards, A100 handoff packaging, the legacy NeuralStateSpaceTranslator, the historical NeuroTwinPairOperator, and the experimental NFC implementation under src/neurotwin/models/nfc/.
The current branch is ptlnextdoor/kahlus-v2-prs at cdc3ee9, with a clean worktree at the start of this pass. The graph report identifies the core abstractions as the old translator, Pair-Operator, split/leakage machinery, prepared tasks, baseline runners, architecture registry, and the new NFC synthetic suite.
Trusted#
Split manifests, event manifests, leakage audits, prepared-task generation, and claim gates are trusted infrastructure lanes.
Synthetic and MOABB smoke paths are trusted as plumbing checks, not scientific evidence.
A100 packaging and handoff scripts are trusted only when built from a clean committed HEAD and verified by checksums.
The current claim boundary is trusted: no model-superiority, clinical, first-foundation-model, exact TRIBE, exact BrainVista, or fNIRS support claim is allowed.
Not Trusted Yet#
NFC synthetic results are not scientific evidence.
The NFC synthetic falsification gate is the immediate trust boundary before Algonauts/CNeuroMod or 6x A100 runs.
The benchmark layer still needs scrutiny for over-broad labels, ablation thresholds, uncertainty calibration semantics, and whether the gate proves more than plumbing.
Old short-step synthetic signals are invalidated as evidence after the corrected gate.
fNIRS and TurboQuant/TurboVec are research notes or optional future infrastructure, not implemented model claims.
Planned Changes#
Add/update the top-level NFC pivot project-state document.
Add/update the NFC mathematical constitution and no-loss equation ledger.
Add/update a code architecture map tying research concepts to files and claim boundaries.
Add/update a master research-state dossier.
Add fNIRS observation-operator notes as docs only.
Add TurboQuant/TurboVec retrieval/audit notes as docs only.
Tighten claim hygiene docs.
Verify and, only if necessary, harden the NFC synthetic gate and A100 handoff strictness.
Out of Scope#
No A100 jobs.
No raw public neural data.
No clinical, MDD, diagnostic, or treatment claims.
No full fNIRS implementation.
No TurboVec dependency or vendored retrieval stack.
No deletion of old translator, Pair-Operator, MOABB, leakage, reporting, model-card, or packaging systems.
No claim that NFC beats baselines unless the strict evidence gate actually passes.
Validation Plan#
git diff --checkPYTHONPATH=src python3 -m unittest discover -s tests -vPYTHONPATH=src python3 -m neurotwin.cli doctorPYTHONPATH=src python3 -m neurotwin.cli eval --suite nfc_synthetic --out-dir /tmp/neurotwin_nfc_alignment_smoke --train-steps 1 --seed 0PYTHONPATH=src python3 -m neurotwin.cli eval --suite nfc_synthetic --out-dir /tmp/neurotwin_nfc_alignment_strict --train-steps 5 --seed 0 --require-pass || truebash scripts/run_smoke.sh /tmp/neurotwin_alignment_smokegraphify update .Package only after validation and only from a clean committed HEAD if a commit is made.