asx-cert / README.md
mrfullcream's picture
seed: cert-input dataset card
b971602 verified
|
Raw
History Blame Contribute Delete
1.96 kB
metadata
configs:
  - config_name: cert-bundle
    data_files:
      - split: dev
        path: splits/dev/*
      - split: holdout
        path: splits/holdout/*

mktmind/asx-cert

Durable cert inputs for asx-lake certify in the ASX document lake.

Why this exists

asx-lake certify runs a frozen route against a sealed gold split and emits a content-addressed ADOPT/REJECT decision. The gold ships in git (probe/labels/*.json). The rest of what certify reads was not durable —

  • the frozen route document,
  • the per-gold-page route predictions (+ natives),
  • the job receipts (render digests from the run), and
  • the render manifest (the independent side of page attribution).

When the v2 lake and its predictions were lost, recertifying required regenerating predictions and surfaced real defects. This dataset is the fix: it mirrors the served-lake durability pattern (mktmind/asx-lake + lake.ref) so a fresh checkout + asx-lake cert pull reproduces certification from a pin.

Layout

splits/<split>/
  frozen-route.json       # the pinned route certified against
  predictions/             # per-gold-page route predictions
  natives/                 # native arm outputs (*.md per stem) [if used]
  job-receipts/            # page_id -> render_sha256 (one or more)
  render-manifest/         # render manifest entries (one or more)
  decision.json            # the sealed CertificationDecision/v1 (publish never rewrites it)

Pin

A 1-line <repo>@<sha> pin lives in cert.ref at the repo root (git-tracked). asx-lake cert publish --split <split> gathers + verifies the decision is sealed, uploads this tree, then writes cert.ref last (pin-last). asx-lake cert pull fetches the pinned revision so certify reproduces.

The publish only reads an already-written decision; it never recomputes one to make a pin succeed (refuse to pin a half-cert).