Skip to content

PIPELINE.md

Latest commit

 

History

History
89 lines (76 loc) · 5.22 KB

PIPELINE.md

File metadata and controls

89 lines (76 loc) · 5.22 KB
Raw

BSFF — the whole pipeline, in one picture

BSFF takes a claim about a signal and returns a machine-readable, provenance- bound verdict. Nothing is ever promoted to "true": the strongest an empirical claim can earn is survived falsification under stated conditions. Everything below is one chain, not a pile of parts.

   raw signal                         a publication's claim
   (.edf/.bdf/.csv/.npy)              (verbatim quote + proposer)
        │                                      │
        ▼                                      ▼
   normalize ──► raw-signal guard         anchor (verbatim, or QUARANTINED_UNANCHORED)
   (EDF/BDF reader,                            │
    physical units)                            ▼
        │                              classify falsifiability tier
        ▼                                      │
   ┌─────────────────────── route ───────────────────────┐
   │ EMPIRICAL_STATISTICAL → surrogate battery (MIAAFT + Bayes conjunction)
   │ EMPIRICAL / causal     → conditional transfer entropy (directed)
   │ LOGICAL                → argument-structure check (not soundness)
   │ DEFINITIONAL/NORMATIVE/NON_FALSIFIABLE → quarantine
   └──────────────────────────────────────────────────────┘
        │
        ▼
   verdict ──► hash-chained truth ledger ──► human-readable report (HTML/MD)
                    │
            corpus / batch: many sources at once, with extraction accountability

Stages

stage module guarantee
ingest normalize, datasets.load_series reads EDF/EDF+/BDF/CSV/NPY in pure Python; fail-closed
raw-signal guard datasets.check_rawness refuses feature tables / accuracy matrices / labels — tests the signal, not someone's preprocessing
anchor adjudication.source a claim absent from its source is quarantined, never judged (anti-fabrication floor)
classify adjudication.falsifiability deterministic, auditable tier; default fail-closed NON_FALSIFIABLE
route → empirical verdict_engine, surrogate_engine MIAAFT surrogate null + effect-size conjunction gate
route → causal transfer_entropy directed, conditional; calibrated operating characteristic
route → logical adjudication.argument detects argument structure, never claims soundness
ledger adjudication.ledger append-only hash chain; tampering breaks it
report adjudication.report_render faithful HTML/Markdown with provenance hash
corpus adjudication.batch many sources, consolidated, with proposer accountability

Dispositions (none of them is "true")

SURVIVED_FALSIFICATION · DIRECTED_COUPLING_SURVIVED · DIRECTED_COUPLING_UNCONDITIONED · REFUTED · UNSUPPORTED · PENDING_EVIDENCE · ARGUMENT_STRUCTURE_DETECTED / ARGUMENT_STRUCTURE_INCOMPLETE / NOT_AN_ARGUMENT · QUARANTINED_{UNANCHORED,DEFINITIONAL,NORMATIVE,NON_FALSIFIABLE}

Command index

command does
bsff selftest run the operational-kernel self-validation
bsff falsify --claim spec.json --signal s.npy falsify one signal claim with a ClaimSpec
bsff normalize --input rec.edf [--list|--channel|--out] read raw EDF/BDF → canonical array
bsff adjudicate-data --data s.edf --test nonlinear_structure|directed_coupling data-driven verdict (bring your own raw signal)
bsff ingest --arxiv <id> fetch a paper's abstract as a provenance-stamped source
bsff adjudicate --source-text t.txt --claims c.json anchor, classify, route, ledger a source's claims
bsff adjudicate-batch --manifest corpus.json adjudicate a corpus + extraction accountability
bsff render --report r.json --format html|md render a verdict for human reading
bsff ledger-verify --ledger l.jsonl check the hash-chain integrity of a ledger

Calibration (measured, not asserted)

The empirical engines are validated against labelled ground truth (bsff.operating_characteristic, bsff.te_operating_characteristic, and the v0.2.0 validation corpus): genuine nonlinear structure survives, matched linear/IID nulls are refuted, a causal pair is called directionally, a null pair is not, and a common-drive confound that fools pairwise transfer entropy collapses under conditioning. The residual at small sample sizes is documented, not hidden — see OPERATING_CHARACTERISTIC.md and TRANSFER_ENTROPY.md.

Boundaries

  • BSFF does not decide truth; it decides what scrutiny a claim admits and applies it. It does not prove BCI claims true.
  • The signal engines are linear/spectral. Nonlinear directed coupling (k-NN transfer entropy) and non-time-series designs (two-group, cohort) are out of scope and would need their own validated tests.
  • It is not externally validated against TISEAN and carries not regulatory validation. It is one auditable, reproducible pass — not a replacement for peer review.