Evidence infrastructure for financial AI research and AI-generated outputs.
We build a governance suite that makes research easier to audit and harder to overstate.
ashare-data-immunityfor data quality and snapshotspit-adjusterfor point-in-time price meaninglookahead-freefor timing checksfactor-qcfor backtest qualityfalsification-ledgerfor claims and evidence trailslesson-bookfor surfaced past mistakesholdout-governancefor AI-assisted research receipts
The backbone is one flow:
data -> adjust -> timing -> backtest -> falsify -> review -> publish
holdout-governance sits across that flow as the release gate. It checks what was used, what passed, what is missing, and whether a human approved the result.
| Flow stage | Repository | What it does |
|---|---|---|
| entry / release gate | holdout-governance |
the wrapper: one artifact, one verdict (also gov mcp for agents) |
| data | ashare-data-immunity |
A-share daily-bar quality, snapshots, SHA-256 manifests |
| adjust | pit-adjuster |
point-in-time back-adjustment with drift detection |
| timing | lookahead-free |
verifiable look-ahead-freedom for pipelines |
| backtest | factor-qc |
fail-closed backtest quality gate (DSR/PBO/MinTRL) |
| falsify | falsification-ledger |
pre-registration, hash-chained ledger, adjudication |
| learn (loop-back) | lesson-book |
surfaced past mistakes become the next round's checks |
The pinned repositories on this profile follow this order: release gate first, then the data pipeline in flow order — read the org alphabetically and you miss the chain; read it as pinned and the pipeline reads top-down.
A holdout set is the data you don't touch until the very end — it keeps your story honest. A holdout juror is the one who refuses to go along until the evidence is in. Every research claim deserves both.
- We do not place orders.
- We do not change trading rules.
- We do not give investment advice.
- We do not treat one passing check as proof of profit.
If you are doing AI-assisted financial research, start with
holdout-governance. It records the evidence cutoff, checks that passed,
the AI identity, and the human review state in one manifest.
- Three data incidents in one night —
a production watchdog caught a ×100 scale corruption, then the same
investigation surfaced a missing corporate-action event and a mixed-source
unit defect. No single check caught all three — the layers did. Both
failure signatures are reproducible offline in
pit-adjuster's examples.
GitHub lists an org's repositories alphabetically and offers no custom ordering, so the profile pins carry the design order. Keep the pinned repositories in flow order (max 6):
holdout-governance— entry / release gateashare-data-immunity— datapit-adjuster— adjustlookahead-free— timingfactor-qc— backtestfalsification-ledger— falsify
To pin: https://github.com/orgs/holdout-labs/repositories → pin each repo in
that order (the pinned section shows them top-down in pin order). lesson-book
(learn / loop-back) and .github are intentionally not pinned; their role is
documented in the flow table above. When a new family member arrives, decide
its flow stage first, then update the table and the pins together — the
alphabetical list below is not the message.