Skip to content

Releases: hseshadr/assay

avow 0.4.1

Choose a tag to compare

@hseshadr hseshadr released this 13 Aug 16:06
3121df1

Avow 0.4.1 is the corrective release for Python CLI ledger writers and supersedes 0.4.0.

The ledger append and convenience-head update now share one bounded, durable process-lock boundary. Recovery fails closed on stale or missing pins; malformed and oversized ledgers are bounded; receipt, keypair, and head persistence use staged atomic writes; and CLI failures redact caller values. Mixed 0.4.0/0.4.1 writers continue to coordinate on the legacy ledger-file lock.

The release also freezes deterministic latency and RSS budgets, brings every Python function under the repository's Grade A complexity contract, and keeps Python/TypeScript receipt bytes and metric answers pinned by shared vectors.

Published as avow==0.4.1 on PyPI with PEP 740 attestations and @edgeproc/avow@0.4.1 on npm with SLSA provenance. Users of assay score --ledger should upgrade from 0.4.0.

Full detail: CHANGELOG.md.

avow 0.4.0

Choose a tag to compare

@hseshadr hseshadr released this 13 Aug 12:37
4484b86

Adds an agreement face for ordered raters: quadratic-weighted Cohen's kappa, Kendall tau-b, exact and weighted agreement, with explicit undefined results instead of flattering degenerate scores.

The Python scoring face now exposes confusion counts and false-negative rate. The pair-versioned @edgeproc/avow browser package gains recall@k, precision@k, F1@k, MRR, and binary confusion metrics, pinned to Python by 23 shared hand-computed vectors.

The release gate breaks 46 named claims—17 in TypeScript—and requires every guard to turn red before restoring the whole suite green. A __proto__ document-id parity defect and mutation-harness stale-bytecode defect are fixed.

Published as avow==0.4.0 on PyPI with PEP 740 attestations and @edgeproc/avow@0.4.0 on npm with SLSA provenance. Full detail: CHANGELOG.md.

avow 0.3.0

Choose a tag to compare

@hseshadr hseshadr released this 03 Aug 19:31
3c3c89f

Adds assay.ranking — precision@k, recall@k, nDCG@k, MRR and MAP over ir_measures/trec_eval, the IR field's reference implementation.

scikit-learn was rejected: its ranking metrics are multilabel-classification metrics and needed two hand-written semantic corrections at the boundary — without one, retrieving 1 of 4 relevant documents scored 1.0. Both corrections are deleted.

Also renames ReplayMismatchPayloadHashMismatch (breaking). Published to PyPI as avow and npm as @edgeproc/avow, both with provenance.

Full detail in CHANGELOG.md.