habitat

Your process is validated. Did a hand move the measure?

habitat doesn’t replace a validation, a binder, or a guardrail. It adds one thing they can’t: a certificate — recomputable on your bench, from the artifact alone — that a data-handling step left the protected geometry byte-identical, or a precise reading of how far it moved. Your validated process stands; here is a floor your counterparty can recompute for themselves.

What every receipt here is held to — the measured baseline, what each digest can and cannot see, and the open items — is stated in two sentences on the method page, dated, with the benches behind every clause.

non-influence — did your own hand move what you were measuring?
pharmacovigilance · drug safety
Did the pipeline move the signal?
A signal-detection pipeline over adverse-event data — dedup, coding, filtering, disproportionality. Every step can quietly move a real signal.
Read the case →
model risk · fair lending
Did data-prep move the protected coupling?
Model risk management under SR 26-2. Reweighting, sampling, imputation, remediation — each can perturb the protected-attribute geometry an examiner will look for.
Read the case →
AI provenance · agent traces
Did the agent’s own actions leak into its record of what happened?
A trace interleaves collected observations and generated turns. The fear is the system treating its own output as observed fact — silent, compounding.
Read the case →
exclusion — can you prove what never entered — and what fully left?
information governance · walls, barriers, holdouts
Did a readable record shape what the system learned?
Ethical walls, information barriers, revoked consent, eval holdouts — the record stays readable, and today “it never entered” is procedure, not proof.
Read the case →
privacy · right to erasure
Did the erased record actually leave the state?
Deleting the row is easy; the value lives in the consolidated state. The field’s best is a probabilistic bound — a place for residual influence to hide.
Read the case →
the frontierThree receipts past the shipped gallery — real pinned data where named, constructed corpora where marked, every number recomputable on your bench. Each exists only because the record lives on the substrate: adopt the substrate to get the receipt.
frontier · quants — did the state ever see the future — and did the revision change the history?
quantitative research · backtests
Did the backtest’s state ever see the future?
Point-in-time databases attest availability, not use; lineage attests events, not effect. Here every state in the history is a pure function of its own past.
Read the frontier case →
data operations · vendor restatements
Did the revision change the history — or only its packaging?
Byte-diff the files and everything changed; compare within an ε and you handed the vendor a window. Here the revision is graded against the corpus’s own floor.
Read the frontier case →
frontier · texts — which acts actually shaped the accumulated state?
authorship · AI disclosure
Which acts actually moved the record?
Diff and blame attribute characters, not effect. Per act: did it move the accumulated semantic state past what re-treading the field’s own history costs?
Read the frontier case →

Instruments

Receipts are records we published and you recompute. An instrument runs on your data. There is one so far: Trace your noise — drop a CSV and your browser reads whether its covariance spectrum’s tail could be noise at all, or only after a constant is added. Nothing you drop leaves your tab, and the engine must tie to a published reference readout before it reads anything. The same verdict is pre-built for every public card on this page.

The ledger

Every card is one attested line: the question it answers, the register it answers in, and a fingerprint you recompute yourself. Not a wall of green — a set of receipts. See the ledger →