eval-toolkit 0.50.0__tar.gz → 1.0.1__tar.gz

{eval_toolkit-0.50.0 → eval_toolkit-1.0.1}/.gitignore

@@ -54,6 +54,14 @@ mutants/
 gate3-audit-prompt.md
 gate3-audit-report.md
 gate3-audit-round-*-report.md
+# R8-C10 audit fix: extend to cover the comprehensive-audit-* and
+# audit-verification-* naming conventions introduced at v0.50/v0.51.
+codex-comprehensive-audit-*-report.md
+codex-microaudit-*.md
+gemini-microaudit-*.md
+audit-gemini.md
+comprehensive-audit-codex.md
+audit-verification-*.md
 
 # Claude Code project settings (machine-local)
 .claude/

{eval_toolkit-0.50.0 → eval_toolkit-1.0.1}/CHANGELOG.md

@@ -5,6 +5,304 @@ All notable changes to this project will be documented in this file.
 The format is based on [Keep a Changelog](https://keepachangelog.com/en/1.1.0/),
 and this project adheres to [Semantic Versioning](https://semver.org/spec/v2.0.0.html).
 
+## [1.0.1] — 2026-05-25 — audit_citation_alignment + RC4 docs polish
+
+First v1.x patch release. Ships the `audit_citation_alignment` validator
+that's been pre-staged by consumer `prompt-injection-detection-prototype`
+(see #77), plus the smallest #76 cleanup item (RC4).
+
+### Added
+
+- **`audit_citation_alignment` module** — flat-module per [ADR 0001](docs/source/adr/0001-flat-module-layout.md)
+  (stay-flat-through-v1.x; subpackage restructure deferred to v2.0).
+  Exports `validate_citations(...)`, `ADRSubject`, `CitationMisalignment`,
+  and `extract_adr_subject_category` as Tier 1 STRICT (per
+  [ADR 0003](docs/source/adr/0003-stability-contract-and-gate3-methodology.md)).
+  Validator catches the bug class where a markdown surface cites "per
+  ADR-NNN" but the cited ADR's actual subject doesn't match the
+  surrounding claim category — motivated by the V1.3.2 P1-2 finding in
+  the consumer `prompt-injection-detection-prototype` audit where
+  `docs/REPRODUCIBILITY.md:76` cited ADR-029 (test markers) for a
+  tier-lock claim that should have cited ADR-034 (reproducibility tier
+  ladder). The mis-citation went undetected by lychee (URL-resolves
+  check), consumer's `audit_numbers.py` (numeric values), and consumer's
+  `audit_adr_count_claims.py` (count claims). Closes #73.
+- **Pre-tag dogfood**: `validate_citations()` exercised against
+  eval-toolkit's own docs (95 files including README + audit_findings +
+  methodology + migration guides) — 0 misalignments found. Validator
+  proven in production use before consumers adopt.
+
+### Fixed
+
+- **RC4** (#76 cleanup) — v0.51 documentation count-tally reconciliation
+  across `docs/source/audit_findings.md`, `docs/source/migration/v0.51.md`,
+  and `CHANGELOG.md` `[0.51.0]` section. Canonical tally now consistent:
+  **13 confirmed → fixed in v0.51 / 3 refuted (R8-G2 + R8-G5 + R8-V1+V2
+  paired) / 2 deferred (R8-G3, R8-G4) = 18 total**. Prior drift was
+  "left 2 undecided" (migration/v0.51) vs "deferred" (CHANGELOG) +
+  ambiguity in the audit_findings ship-status section. Closes RC4 of #76.
+
+## [1.0.0] — 2026-05-25 — Stability contract activates per ADR 0003
+
+v1.0 is a **stability-contract activation**, not a code delta from v0.51.
+Every fix that landed at v0.51 is what v1.0 ships; the new thing at v1.0
+is that the [ADR 0003](docs/source/adr/0003-stability-contract-and-gate3-methodology.md)
+Tier 1 / Tier 2 / Tier 3 stability contract becomes load-bearing.
+Breaking changes to Tier-1 surfaces after v1.0 require a major bump (v2.0).
+
+### Contract activation
+
+- **Tier 1 STRICT** — public-API signatures captured in
+  `tests/golden/public_api/snapshot.json`. Any signature drift bumps to v2.0.
+- **Tier 2 ADDITIVE** — the 9 strict Protocols (`Scorer`, `LeakageCheck`,
+  `Splitter`, `ThresholdSelector`, `DatasetLoader`, `MetricSpec`,
+  `MetaLearner`, `Probe`, `TextTransform`) + 1 opt-in (`Versioned`).
+  Method shapes are frozen; additive subprotocols + new Protocols allowed.
+- **Tier 3 FREE** — internal modules (prefixed `_`). Refactors don't
+  bump major.
+
+### Gate 3 audit closure (Rounds 5 → 10)
+
+The multi-LLM cross-review sequence is closed. Per ADR 0003, Gate 3
+substitutes Codex + Gemini + Claude independent reads for external
+academic peer review. Outcomes:
+
+- **Round 8** (against v0.50): 13 confirmed → fixed in v0.51; 3 refuted;
+  2 v1.x-deferred (R8-G3 custom exceptions; R8-G4 joblib OOM capping).
+- **Round 9** (against v0.51 RC): 6 confirmed of 10 source items + 3
+  third-audit findings in modules neither auditor cited. 2 candidate-
+  blocker items fixed in-PR before tag.
+- **Round 10** (micro-audit on R9 follow-on commit): 3 Codex confirmed →
+  fixed in v0.51; 1 accept-as-design (Gemini); 1 refuted (Gemini
+  Pattern-1 violation).
+
+Full ledger at `docs/source/audit_findings.md` Rounds 5 → 10. v1.0.1
+cleanup batch tracked at GH issue #76.
+
+### Carried-over deprecations
+
+The R8-C1 `DeprecationWarning` on multi-seed `evaluate_folded(seeds=...)`
+calls without an explicit `reseed_splitter` callback **persists past v1.0
+by design** (pre-v1.0 deprecation window is one minor; `DEPRECATION.md`
+requires ≥2 minors to close a cycle). Single-seed callers see no change;
+multi-seed callers should pass `reseed_splitter` for true seed variance.
+
+### Migration
+
+If your consumer is on v0.51, nothing changes. If on v0.50 or earlier,
+follow [`docs/source/migration/v0.51.md`](docs/source/migration/v0.51.md)
+for the actual migration steps. Downstream projects should pin
+`eval-toolkit>=1.0,<2.0` to opt into the stability contract.
+
+## [0.51.0] — 2026-05-24 — Round 8 rectification batch
+
+The 18-item rectification batch following the Round 8 multi-LLM audit
+(Codex + Gemini reports verified at
+`audit-verification-codex-gemini-v0.50.0.md`, 2026-05-24). Per Decision
+Y.2 + the staggered-pre-v1.0 plan, v0.51.0 is a BREAKING-allowed
+minor bundling all fixes before v1.0 tags. Round 9 audit runs against
+the v0.51 RC.
+
+**Audit outcome**: 13 confirmed → fixed in this release; 2 deferred
+(R8-G3 custom exceptions, R8-G4 joblib OOM capping) to v1.x as Tier-2
+additive; 3 refuted (R8-G2 cyclic-import framing; R8-G5 cherry-picked
+weak test; R8-V1 + R8-V2 over-confident Gemini validations). See
+`docs/source/audit_findings.md` Round 8 section for the full ledger.
+
+**Round 9 follow-on**: a Round 9 multi-LLM cross-review (Codex + Gemini)
+ran against the v0.51 RC pre-tag. Verified by Claude at
+`audit-verification-round-9-v0.51.0.md` (6 confirmed / 3 refuted / 1
+partial; plus 3 third-audit fixes in modules neither auditor cited).
+**Two third-audit findings + one source-report regression fix shipped
+in this RC pre-tag** (commit-graph below); the remaining 4 deferred
+items go to v1.0.1. See `audit_findings.md` Round 9 section for the
+full ledger.
+
+### Added (Round 9 follow-on)
+
+- **R9-F-sweep-1** (CANDIDATE v1.0 BLOCKER closed) — `_sweep.py:
+  _validate_scorer_output()` now validates scorer output is finite
+  (no NaN / +inf / -inf), not just shape. Pre-R9 follow-on, NaN/inf
+  scores passed R7-C's shape check and silently propagated into the
+  sweep DataFrame, then silently zeroed the ASR flag (NaN >= threshold
+  is False). Closes the "no silent failures" invariant gap R7-C
+  established for shape but didn't extend to finiteness. Brings sweep
+  validation to parity with `stacking.py`'s `_validate_fit_inputs` /
+  `_validate_predict_inputs`. Tier-2 additive — callers whose scorers
+  were silently producing NaN now get a clear `ValueError` with
+  diagnostic context.
+
+- **R9-F-bootstrap-1** — `bootstrap.bootstrap_ci(...)` emits a
+  `UserWarning` when scipy's BCa method degenerates (returns
+  `ci_low == ci_high == point` or non-finite bounds). Pre-R9, the
+  R8-C4(b) RNG bug spuriously varied bootstrap streams and could mask
+  BCa degeneracy on small-n + ceiling/floor-metric inputs; post-R8 with
+  correct RNG, the brittleness is exposed. Warning text recommends
+  `method='percentile'` as the safer fallback at small n. The default
+  remains `method='BCa'` (preserves bit-stability for non-degenerate
+  cases); auto-fallback is deferred to v1.0.1 if user demand.
+
+- **R9-F-bootstrap-2** — `bootstrap.mde_from_ci(...)` now explicitly
+  rejects NaN CI width with `RuntimeError`. Pre-R9, NaN width
+  (possible when scipy BCa returns NaN bounds) bypassed the
+  `if width <= 0` check (NaN <= 0 is False in IEEE float) and
+  silently returned `MDEEstimate.mde = NaN`. Bundled with F-bootstrap-1.
+
+### Fixed (Round 10 follow-on)
+
+Pre-tag scoped Codex + Gemini micro-audit on `edadddc` surfaced 3
+Codex-confirmed findings (all fix-recommended / minor; no v1.0
+blockers). Verified by Claude; 1 Gemini accept-as-design + 1 Gemini
+refuted (Pattern-1 violation; calibration record in
+`audit_findings.md` Round 10 section). All 3 confirmed findings
+shipped in this RC pre-tag:
+
+- **R10-F1** — `protocols.Scorer.predict_proba` docstring + `_sweep.py`
+  error message clarification. Pre-R10, `_validate_scorer_output`'s
+  runtime error said "finite floats in [0, 1]" but the boundary check
+  only enforced finiteness (no range validation); the Scorer Protocol
+  docstring also lacked an explicit `[0, 1]` contract statement. R10-F1
+  extends the Protocol docstring to document calibrated-probability
+  semantics + reword the sweep runtime message to drop the unenforced
+  `[0, 1]` claim. Range enforcement is intentionally deferred to a
+  future minor once consumer usage patterns clarify whether the
+  Protocol should be strict (`[0, 1]`) or permissive (ranking scores).
+
+- **R10-F2** — `tests/test_bootstrap_unit.py::test_bootstrap_ci_bca_degeneracy_emits_warning`
+  test predicate hardening. Pre-R10, the test's assertion block used
+  `if ci.ci_low == ci.ci_high == ci.point_estimate:` — but NaN==NaN is
+  False in IEEE float, so the assertions were silently skipped on the
+  current scipy fixture (which returns NaN bounds). The test passed
+  WITHOUT proving the warning fires for the common degeneracy mode.
+  R10-F2 mirrors the production predicate exactly:
+  `(not np.isfinite(low)) or (not np.isfinite(high)) or (low == high == point)`.
+  The assertion block now runs whenever ANY degeneracy mode fires.
+
+- **R10-F3** — `bootstrap.mde_from_ci` docstring update for the
+  R9-F-bootstrap-2 non-finite-width branch. Pre-R10, the Raises section
+  said "non-positive width" only; the implementation has also been
+  rejecting non-finite width since `edadddc` but the docstring lagged.
+  R10-F3 updates the Raises text to "non-positive or non-finite width"
+  and adds a 4-line note explaining the scipy-BCa NaN-bound motivation
+  so callers understand the new behavior is intentional, not incidental.
+
+### Added
+
+- **R8-C6** — `calibration.reliability_curve(...)` and
+  `calibration.maximum_calibration_error(...)` now call
+  `_validate_calibrated_score(y_score)` BEFORE the sklearn dispatch.
+  Pre-v0.51 these functions silently accepted raw logits (any range);
+  sibling `metrics.expected_calibration_error*` variants already
+  validated input range via the same helper. Now symmetric — out-of-range
+  scores raise `ValueError` with the same actionable diagnostic. Also,
+  `calibration.fit_temperature(...)` now validates the `bounds` tuple
+  (finiteness, positivity, `lo < hi`) BEFORE forwarding to
+  `scipy.optimize.minimize_scalar` — cryptic optimizer errors replaced
+  with actionable input-validation errors.
+
+- **R8-F1** — `losses.RecallAtLowFPR.__init__(...)` now validates
+  `pos_weight > 0` at construction time, matching the sibling-kwarg
+  validators for `fpr_target` and `fpr_smoothing_beta`. Pre-v0.51
+  non-positive `pos_weight` produced degenerate-but-bounded loss
+  values silently.
+
+- **R8-F2** — `metric_specs.ece(n_bins=, strategy=)` factory now validates
+  `n_bins` eagerly at spec-construction time (matches the eager
+  `strategy` validation already present). Pre-v0.51 `n_bins`
+  validation was deferred to compute time.
+
+- **R8-F3** — `analysis.CsvPredictionReader.read_predictions(...)` now
+  detects missing CSV columns at read time and raises a
+  `ValueError(f"CSV file at {uri!r} is missing required column(s) ...")`
+  with the file path + available columns. Pre-v0.51 missing columns
+  were silently filled with empty strings, causing cryptic
+  `ValueError: invalid literal for int() with base 10: ''` downstream
+  in `load_prediction_arrays`'s dtype conversion. Root cause now
+  surfaces at the boundary.
+
+- **R8-C1** — `harness.evaluate_folded(...)` now accepts an optional
+  `reseed_splitter: Callable[[Splitter, int], Splitter] | None`
+  callback. When provided, each seed iteration calls
+  `reseed_splitter(splitter, seed)` to produce a fresh splitter for
+  that seed's fold iteration. Default `None` preserves the historical
+  behavior (the same splitter instance is reused across the seed loop,
+  so multi-seed × CV only varies the bootstrap RNG, not fold
+  partitions) AND emits a `DeprecationWarning` when `len(seeds) > 1`.
+  The warning persists past v1.0 because the pre-v1.0 deprecation
+  window (v0.51 → v1.0) is one minor and ADR 0003 / DEPRECATION.md
+  require ≥2 minors to close a cycle. Migration example::
+
+      from dataclasses import replace
+      evaluate_folded(
+          scorers, splitter, slice_,
+          seeds=(1, 2, 3),
+          reseed_splitter=lambda sp, s: replace(sp, seed=s),
+          ...
+      )
+
+  R8-C1 audit fix.
+
+### BREAKING
+
+- **R8-C2** — `SourceDisjointKFoldSplitter.iter_folds(...)` now caps
+  the fold count at `min(self.k, n_sources)` (matching
+  `get_n_splits(...)`). Pre-v0.51 the loop ran `range(self.k)` and
+  yielded EMPTY test partitions for the surplus folds when
+  `k > n_sources` while `get_n_splits` returned `min(k, n_sources)`
+  — the two methods silently disagreed on fold count. v0.51 caps both
+  at the same value AND emits a `UserWarning` when `k > n_sources` so
+  the caller knows the cap was applied. Callers that consumed the
+  surplus empty-test folds will see fewer iterations now; that was
+  the bug. (Probe-verified at
+  `audit-verification-codex-gemini-v0.50.0.md`.)
+
+- **R8-C3** — `thresholds.recall_at_fpr(...)` fallback semantics changed
+  when no threshold satisfies `target_fpr`. Pre-v0.51 the fallback set
+  `threshold = 1.0` and then computed `y_pred = (y_score >= 1.0)` —
+  inclusive comparator — which classified any negative-class sample
+  with score exactly 1.0 as predicted-positive. The probe
+  `recall_at_fpr(y=[0,1], scores=[1.0,1.0], target_fpr=0.0)` returned
+  `actual_fpr=1.0, fp=1` in silent violation of the function's
+  FPR-ceiling invariant. v0.51 returns a SENTINEL
+  `RecallAtFprResult(threshold=np.inf, recall=0.0, actual_fpr=0.0,
+  fp=0, tn=n_val_neg)` whenever the constraint is unsatisfiable.
+  Callers detect via `np.isinf(result.threshold)`. The
+  `actual_fpr ≤ target_fpr` invariant is now preserved by construction.
+  Migration: any caller filtering on `result.threshold` should add an
+  `np.isinf(...)` branch — pre-v0.51 the sentinel value was `1.0`.
+  (Verified at `audit-verification-codex-gemini-v0.50.0.md`.)
+
+- **R8-C4(a)** — `harness.evaluate(...)` with a `Generator`-typed `rng`
+  is now bit-stable across `n_jobs` values. Prior to v0.51, the same
+  `rng` object was attached to every `(slice, scorer)` work_unit;
+  joblib forked copies at the SAME generator state into N parallel
+  workers, so every worker used identical bootstrap sample streams —
+  silently producing non-independent CIs across `(slice, scorer)`
+  pairs in parallel mode and divergent results vs sequential mode.
+  The v0.51 implementation spawns one independent `SeedSequence` per
+  work unit at the dispatch boundary in `_score_all_slices` (depends
+  on the R8-C4(b) `spawn_seed_sequences` fix). Each pair now sees an
+  independent bootstrap stream; sequential (`n_jobs=1`) and parallel
+  (`n_jobs>1`) modes produce bit-identical CIs per the SPEC 7
+  contract at `docs/source/methodology/parallelism.md`. Integer `rng`
+  callers (the common case) are unaffected. (Verified by multi-slice
+  probe at `audit-verification-codex-gemini-v0.50.0.md`.)
+
+- **R8-C4(b)** — `eval_toolkit._rng.spawn_seed_sequences(rng, n)` now
+  respects `Generator` state. Prior to v0.51, the function extracted
+  the bit-generator's seed_seq and called `.spawn(n)` on it — so a
+  `Generator` advanced by prior draws produced the same children as a
+  fresh `Generator` with the same construction seed. The new
+  implementation draws `n` fresh entropy values FROM the generator
+  via `rng.integers(0, 2**63-1, size=n)` and wraps each in a
+  `SeedSequence`. Each call advances generator state, so repeated
+  calls on the same instance yield different children. This was the
+  root cause of bootstrap non-independence across `(slice, scorer)`
+  pairs in `harness.evaluate` — when the same `Generator` was shared
+  across bootstrap callsites, all callsites silently used the same
+  resample stream. (Verified probe at
+  `audit-verification-codex-gemini-v0.50.0.md`.)
+
 ## [0.50.0] — 2026-05-23 — SPEC 7 `rng` parameter adoption
 
 The SPEC 7 follow-up to v0.49.0. The `_rng.py` scaffold shipped at

{eval_toolkit-0.50.0 → eval_toolkit-1.0.1}/PKG-INFO

@@ -1,6 +1,6 @@
 Metadata-Version: 2.4
 Name: eval-toolkit
-Version: 0.50.0
+Version: 1.0.1
 Summary: Reusable evaluation contracts for binary classification: metrics, bootstrap CIs, calibration, artifacts, and evidence gates.
 Project-URL: Homepage, https://github.com/brandon-behring/eval-toolkit
 Project-URL: Documentation, https://brandon-behring.github.io/eval-toolkit/
@@ -114,7 +114,8 @@ format changes.
 │  gpu_info + leakage_report (NeurIPS-aligned)           │
 ├─ Tier 2 ─ Protocol-based orchestration ────────────────┤
 │  Scorer / SliceAwareScorer / LeakageCheck / Splitter   │
-│  ThresholdSelector / DatasetLoader / SimilarityStrategy│
+│  ThresholdSelector / DatasetLoader / MetricSpec        │
+│  MetaLearner / Probe / TextTransform (9 strict)        │
 │  Versioned (opt-in: per-object versions in manifest)   │
 ├─ Tier 1 ─ Functional core ─────────────────────────────┤
 │  pr_auc / roc_auc / ECE variants / Brier / bootstrap_ci│
@@ -129,69 +130,65 @@ run: capture the manifest.
 
 ## Documentation
 
-- **[Getting started](docs/getting-started.md)** — end-to-end
+- **[Getting started](docs/source/getting-started.md)** — end-to-end
   walkthrough for new users: install, define a Scorer, build slices,
   run `evaluate()`, persist results, add a claim, render a plot.
-- **[Methodology curriculum](docs/methodology/README.md)** — 16
+- **[Methodology curriculum](docs/source/methodology/README.md)** — 16
   chapters on splits, metrics, calibration, evidence gates,
   prediction artifacts, and more.
-- **[Schema reference](docs/schemas.md)** — field-by-field semantics
+- **[Schema reference](docs/source/schemas.md)** — field-by-field semantics
   for `results.v1.json`, `results_full.v1.json`, `manifest.v1.json`.
-- **[Migration guides](docs/MIGRATION.md)** — v0.6→v0.7, v0.7→v0.8,
-  v0.8→v0.9.
-- **[Extending](docs/extending.md)** — Protocol-by-Protocol guide for
+- **[Migration guides](docs/source/MIGRATION.md)** — per-version migration
+  hub (v0.7 onward).
+- **[Extending](docs/source/extending.md)** — Protocol-by-Protocol guide for
   custom Scorers, Splitters, LeakageChecks, ThresholdSelectors,
   DatasetLoaders, EvidenceGates.
-- **[Repo strategy](docs/repo-strategy.md)** — how the package is
-  organized, the 6-bucket target shape, and the checklist that
-  governs when to extract a sub-package into its own repo.
+- **[Repo strategy](docs/source/repo-strategy.md)** — how the package is
+  organized, the flat-module layout per ADR 0001, and the v2.0 trigger
+  criteria for any future subpackage split.
 
 ## Methodology
 
 What good binary-classification evaluation looks like, with each
 concern mapped to the toolkit primitive that operationalizes it.
 
-- [`docs/methodology/`](docs/methodology/README.md) — the curriculum
-  (16 chapters). Recommended reading order:
-  [`leakage`](docs/methodology/leakage.md) →
-  [`splits`](docs/methodology/splits.md) →
-  [`thresholds`](docs/methodology/thresholds.md) →
-  [`calibration`](docs/methodology/calibration.md) →
-  [`comparison`](docs/methodology/comparison.md) →
-  [`bootstrap`](docs/methodology/bootstrap.md) →
-  [`length_stratification`](docs/methodology/length_stratification.md) →
-  [`text_dedup`](docs/methodology/text_dedup.md) →
-  [`versioning`](docs/methodology/versioning.md) →
-  [`fairness`](docs/methodology/fairness.md) →
-  [`reproducibility`](docs/methodology/reproducibility.md) →
-  [`testing`](docs/methodology/testing.md) →
-  [`reading_list`](docs/methodology/reading_list.md).
-- [`docs/MIGRATION.md`](docs/MIGRATION.md) — per-version migration
-  guides (v0.6→v0.7, v0.7→v0.8).
-- [`docs/roadmap.md`](docs/roadmap.md) — forward-looking tracker;
-  v1.0.0 path; consumer gap-doc cross-links.
+- [`docs/source/methodology/`](docs/source/methodology/README.md) — the
+  curriculum (16 chapters). Recommended reading order:
+  [`leakage`](docs/source/methodology/leakage.md) →
+  [`splits`](docs/source/methodology/splits.md) →
+  [`thresholds`](docs/source/methodology/thresholds.md) →
+  [`calibration`](docs/source/methodology/calibration.md) →
+  [`comparison`](docs/source/methodology/comparison.md) →
+  [`bootstrap`](docs/source/methodology/bootstrap.md) →
+  [`length_stratification`](docs/source/methodology/length_stratification.md) →
+  [`text_dedup`](docs/source/methodology/text_dedup.md) →
+  [`versioning`](docs/source/methodology/versioning.md) →
+  [`fairness`](docs/source/methodology/fairness.md) →
+  [`reproducibility`](docs/source/methodology/reproducibility.md) →
+  [`testing`](docs/source/methodology/testing.md) →
+  [`reading_list`](docs/source/methodology/reading_list.md).
+- [`docs/source/MIGRATION.md`](docs/source/MIGRATION.md) — per-version
+  migration guides (v0.7 onward; v0.49 / v0.50 / v0.51 included as of
+  v0.51.0).
+- [`docs/source/roadmap.md`](docs/source/roadmap.md) — forward-looking
+  tracker; v1.0.0 path; consumer gap-doc cross-links.
 
 ## Extending eval-toolkit
 
 How to plug your own scorers / leakage checks / splitters / loaders /
 threshold selectors into the harness.
 
-- [`docs/extending.md`](docs/extending.md) — Protocol-by-Protocol
+- [`docs/source/extending.md`](docs/source/extending.md) — Protocol-by-Protocol
   guide, ~50-line full-harness recipe, project-layout pointer.
 
 ## Worked examples
 
-- [`docs/examples/prompt_injection_walkthrough.md`](docs/examples/prompt_injection_walkthrough.md)
-  — End-to-end prompt-injection eval on a synthetic OWASP LLM01:2025
-  fixture; cross-links to the
-  [showcase repo](https://github.com/brandon-behring/prompt_injection_classifier_showcase)
-  for the real Lakera PINT walkthrough.
-- [`docs/examples/pytorch_scorer_example.md`](docs/examples/pytorch_scorer_example.md)
-  — HuggingFace transformer + LoRA `Scorer` adapter (batched inference,
-  GPU/CPU placement, deterministic-mode setup).
-- [`docs/examples/claims_and_gates.md`](docs/examples/claims_and_gates.md)
-  — Composing reference + custom `EvidenceGate`s into a `ClaimSpec` and
-  running `evaluate_claims()` for release-time go/no-go checks.
+- [`docs/source/examples/`](docs/source/examples/index.md) — Sphinx /
+  MyST-NB executable notebooks covering: the evaluation harness,
+  metrics + bootstrap, calibration, claims-and-gates, leakage
+  detection, cross-corpus contamination scanning, character-injection
+  adversarial sweeps, callable-embedder dedup, and the activation-delta
+  probe.
 
 ## Install
 
@@ -291,7 +288,7 @@ with tempfile.TemporaryDirectory() as run_dir:
 | `eval_toolkit.splits` | `Splitter` Protocol + 5 reference impls (holdout / stratified / group / source-disjoint / time-series) |
 | `eval_toolkit.loaders` | `DatasetLoader` Protocol + 4 reference impls (DataFrame / SingleSlice / ParquetGlob / HF datasets) with Croissant-compatible `describe()` |
 | `eval_toolkit.manifest` | `RunManifest` (NeurIPS-aligned) + source-role / guardrail metadata + `make_manifest` / `write_manifest` |
-| `eval_toolkit.claims` | `EvidenceGate` class (frozen dataclass: name + callable check + severity), reference gate factories (`required_metric_gate`, `minimum_slice_size_gate`, `metric_threshold_gate`, etc.), `evaluate_claims()`, and `ClaimReport` for claim-mode vs exploratory-mode checks. See [`docs/extending.md`](docs/extending.md) for writing custom gates and [`docs/examples/claims_and_gates.md`](docs/examples/claims_and_gates.md) for a worked end-to-end example. |
+| `eval_toolkit.claims` | `EvidenceGate` class (frozen dataclass: name + callable check + severity), reference gate factories (`required_metric_gate`, `minimum_slice_size_gate`, `metric_threshold_gate`, etc.), `evaluate_claims()`, and `ClaimReport` for claim-mode vs exploratory-mode checks. See [`docs/source/extending.md`](docs/source/extending.md) for writing custom gates and [`docs/source/examples/claims_and_gates.md`](docs/source/examples/claims_and_gates.md) for a worked end-to-end example. |
 | `eval_toolkit.text_dedup` | `SimilarityStrategy` Protocol + 5 strategies (TF-IDF / hash / embedding / Jaccard / MinHash-LSH); `near_dedup` / `cross_dedup` orchestrators |
 | `eval_toolkit.plotting` | PR curves, reliability diagrams, confusion matrices, score histograms, lift CIs |
 | `eval_toolkit.provenance` | File hashing, run-directory layout, figure metadata sidecar |

{eval_toolkit-0.50.0 → eval_toolkit-1.0.1}/README.md

@@ -31,7 +31,8 @@ format changes.
 │  gpu_info + leakage_report (NeurIPS-aligned)           │
 ├─ Tier 2 ─ Protocol-based orchestration ────────────────┤
 │  Scorer / SliceAwareScorer / LeakageCheck / Splitter   │
-│  ThresholdSelector / DatasetLoader / SimilarityStrategy│
+│  ThresholdSelector / DatasetLoader / MetricSpec        │
+│  MetaLearner / Probe / TextTransform (9 strict)        │
 │  Versioned (opt-in: per-object versions in manifest)   │
 ├─ Tier 1 ─ Functional core ─────────────────────────────┤
 │  pr_auc / roc_auc / ECE variants / Brier / bootstrap_ci│
@@ -46,69 +47,65 @@ run: capture the manifest.
 
 ## Documentation
 
-- **[Getting started](docs/getting-started.md)** — end-to-end
+- **[Getting started](docs/source/getting-started.md)** — end-to-end
   walkthrough for new users: install, define a Scorer, build slices,
   run `evaluate()`, persist results, add a claim, render a plot.
-- **[Methodology curriculum](docs/methodology/README.md)** — 16
+- **[Methodology curriculum](docs/source/methodology/README.md)** — 16
   chapters on splits, metrics, calibration, evidence gates,
   prediction artifacts, and more.
-- **[Schema reference](docs/schemas.md)** — field-by-field semantics
+- **[Schema reference](docs/source/schemas.md)** — field-by-field semantics
   for `results.v1.json`, `results_full.v1.json`, `manifest.v1.json`.
-- **[Migration guides](docs/MIGRATION.md)** — v0.6→v0.7, v0.7→v0.8,
-  v0.8→v0.9.
-- **[Extending](docs/extending.md)** — Protocol-by-Protocol guide for
+- **[Migration guides](docs/source/MIGRATION.md)** — per-version migration
+  hub (v0.7 onward).
+- **[Extending](docs/source/extending.md)** — Protocol-by-Protocol guide for
   custom Scorers, Splitters, LeakageChecks, ThresholdSelectors,
   DatasetLoaders, EvidenceGates.
-- **[Repo strategy](docs/repo-strategy.md)** — how the package is
-  organized, the 6-bucket target shape, and the checklist that
-  governs when to extract a sub-package into its own repo.
+- **[Repo strategy](docs/source/repo-strategy.md)** — how the package is
+  organized, the flat-module layout per ADR 0001, and the v2.0 trigger
+  criteria for any future subpackage split.
 
 ## Methodology
 
 What good binary-classification evaluation looks like, with each
 concern mapped to the toolkit primitive that operationalizes it.
 
-- [`docs/methodology/`](docs/methodology/README.md) — the curriculum
-  (16 chapters). Recommended reading order:
-  [`leakage`](docs/methodology/leakage.md) →
-  [`splits`](docs/methodology/splits.md) →
-  [`thresholds`](docs/methodology/thresholds.md) →
-  [`calibration`](docs/methodology/calibration.md) →
-  [`comparison`](docs/methodology/comparison.md) →
-  [`bootstrap`](docs/methodology/bootstrap.md) →
-  [`length_stratification`](docs/methodology/length_stratification.md) →
-  [`text_dedup`](docs/methodology/text_dedup.md) →
-  [`versioning`](docs/methodology/versioning.md) →
-  [`fairness`](docs/methodology/fairness.md) →
-  [`reproducibility`](docs/methodology/reproducibility.md) →
-  [`testing`](docs/methodology/testing.md) →
-  [`reading_list`](docs/methodology/reading_list.md).
-- [`docs/MIGRATION.md`](docs/MIGRATION.md) — per-version migration
-  guides (v0.6→v0.7, v0.7→v0.8).
-- [`docs/roadmap.md`](docs/roadmap.md) — forward-looking tracker;
-  v1.0.0 path; consumer gap-doc cross-links.
+- [`docs/source/methodology/`](docs/source/methodology/README.md) — the
+  curriculum (16 chapters). Recommended reading order:
+  [`leakage`](docs/source/methodology/leakage.md) →
+  [`splits`](docs/source/methodology/splits.md) →
+  [`thresholds`](docs/source/methodology/thresholds.md) →
+  [`calibration`](docs/source/methodology/calibration.md) →
+  [`comparison`](docs/source/methodology/comparison.md) →
+  [`bootstrap`](docs/source/methodology/bootstrap.md) →
+  [`length_stratification`](docs/source/methodology/length_stratification.md) →
+  [`text_dedup`](docs/source/methodology/text_dedup.md) →
+  [`versioning`](docs/source/methodology/versioning.md) →
+  [`fairness`](docs/source/methodology/fairness.md) →
+  [`reproducibility`](docs/source/methodology/reproducibility.md) →
+  [`testing`](docs/source/methodology/testing.md) →
+  [`reading_list`](docs/source/methodology/reading_list.md).
+- [`docs/source/MIGRATION.md`](docs/source/MIGRATION.md) — per-version
+  migration guides (v0.7 onward; v0.49 / v0.50 / v0.51 included as of
+  v0.51.0).
+- [`docs/source/roadmap.md`](docs/source/roadmap.md) — forward-looking
+  tracker; v1.0.0 path; consumer gap-doc cross-links.
 
 ## Extending eval-toolkit
 
 How to plug your own scorers / leakage checks / splitters / loaders /
 threshold selectors into the harness.
 
-- [`docs/extending.md`](docs/extending.md) — Protocol-by-Protocol
+- [`docs/source/extending.md`](docs/source/extending.md) — Protocol-by-Protocol
   guide, ~50-line full-harness recipe, project-layout pointer.
 
 ## Worked examples
 
-- [`docs/examples/prompt_injection_walkthrough.md`](docs/examples/prompt_injection_walkthrough.md)
-  — End-to-end prompt-injection eval on a synthetic OWASP LLM01:2025
-  fixture; cross-links to the
-  [showcase repo](https://github.com/brandon-behring/prompt_injection_classifier_showcase)
-  for the real Lakera PINT walkthrough.
-- [`docs/examples/pytorch_scorer_example.md`](docs/examples/pytorch_scorer_example.md)
-  — HuggingFace transformer + LoRA `Scorer` adapter (batched inference,
-  GPU/CPU placement, deterministic-mode setup).
-- [`docs/examples/claims_and_gates.md`](docs/examples/claims_and_gates.md)
-  — Composing reference + custom `EvidenceGate`s into a `ClaimSpec` and
-  running `evaluate_claims()` for release-time go/no-go checks.
+- [`docs/source/examples/`](docs/source/examples/index.md) — Sphinx /
+  MyST-NB executable notebooks covering: the evaluation harness,
+  metrics + bootstrap, calibration, claims-and-gates, leakage
+  detection, cross-corpus contamination scanning, character-injection
+  adversarial sweeps, callable-embedder dedup, and the activation-delta
+  probe.
 
 ## Install
 
@@ -208,7 +205,7 @@ with tempfile.TemporaryDirectory() as run_dir:
 | `eval_toolkit.splits` | `Splitter` Protocol + 5 reference impls (holdout / stratified / group / source-disjoint / time-series) |
 | `eval_toolkit.loaders` | `DatasetLoader` Protocol + 4 reference impls (DataFrame / SingleSlice / ParquetGlob / HF datasets) with Croissant-compatible `describe()` |
 | `eval_toolkit.manifest` | `RunManifest` (NeurIPS-aligned) + source-role / guardrail metadata + `make_manifest` / `write_manifest` |
-| `eval_toolkit.claims` | `EvidenceGate` class (frozen dataclass: name + callable check + severity), reference gate factories (`required_metric_gate`, `minimum_slice_size_gate`, `metric_threshold_gate`, etc.), `evaluate_claims()`, and `ClaimReport` for claim-mode vs exploratory-mode checks. See [`docs/extending.md`](docs/extending.md) for writing custom gates and [`docs/examples/claims_and_gates.md`](docs/examples/claims_and_gates.md) for a worked end-to-end example. |
+| `eval_toolkit.claims` | `EvidenceGate` class (frozen dataclass: name + callable check + severity), reference gate factories (`required_metric_gate`, `minimum_slice_size_gate`, `metric_threshold_gate`, etc.), `evaluate_claims()`, and `ClaimReport` for claim-mode vs exploratory-mode checks. See [`docs/source/extending.md`](docs/source/extending.md) for writing custom gates and [`docs/source/examples/claims_and_gates.md`](docs/source/examples/claims_and_gates.md) for a worked end-to-end example. |
 | `eval_toolkit.text_dedup` | `SimilarityStrategy` Protocol + 5 strategies (TF-IDF / hash / embedding / Jaccard / MinHash-LSH); `near_dedup` / `cross_dedup` orchestrators |
 | `eval_toolkit.plotting` | PR curves, reliability diagrams, confusion matrices, score histograms, lift CIs |
 | `eval_toolkit.provenance` | File hashing, run-directory layout, figure metadata sidecar |

{eval_toolkit-0.50.0 → eval_toolkit-1.0.1}/src/eval_toolkit/__init__.py

@@ -52,6 +52,14 @@ _EXPORTS: dict[str, str] = {
     # CharacterInjectionStrategy + character_injection SimpleNamespace
     # removed at v0.47 (Decision N + plan §4E). TextTransform Protocol +
     # the 12 concrete dataclasses are now the only public path.
+    # --- audit_citation_alignment ---
+    # Flat-module per ADR 0001 (Stay flat through v1.x; subpackage
+    # restructure deferred to v2.0). Closes #73. Motivated by consumer
+    # V1.3.2 P1-2 ADR-029 mis-citation finding.
+    "ADRSubject": "eval_toolkit.audit_citation_alignment",
+    "CitationMisalignment": "eval_toolkit.audit_citation_alignment",
+    "extract_adr_subject_category": "eval_toolkit.audit_citation_alignment",
+    "validate_citations": "eval_toolkit.audit_citation_alignment",
     # --- losses ---
     "RecallAtLowFPR": "eval_toolkit.losses",
     # --- preprocessing ---