0.4.0 full-history default identity

2026-10-04-default-full-history-identity.csv contains all 24 default-mode studies (eight problems, three seeds, 3,000 proposals): 72,000 compared proposal/value records and zero differences versus pinned 0.3.0. Its adjacent metadata and SHA-256 sidecar record verified paired trace hashes, binaries, sampler sources, and pins. See the identity report.

This directory contains the compact machine-readable evidence for `results-2026-07-18.md`. These are current reruns of the documented profiles, not reconstructed metadata for an older unpublished CSV.

Profile	CSV	Rows	CSV SHA-256	Metadata SHA-256
Full, startup 10	`2026-07-18-full.csv`	60	`7dcc63346df87367d7fe0c151faef369f103508eb3117b2f50009bcb3f83e629`	`2ca193efffc0c25b23948f2c2a8d7c67f819310b4cab633ef7f25cdbc01364ae`
Million discrete, startup 200	`2026-07-18-startup-200.csv`	10	`479c28e9f3825fb152e1b0193350af98922be6bf528c618c04091eb75c207e9a`	`896cd01dec6039c75057935250ae668c51772aa549b008c25c00406f1ba38214`

Each CSV has an adjacent metadata sidecar:

Both runs use:

source snapshot SHA-256 12abddc13f35933acaebf80772a74e3b9fa0af3d637c05f6afc0713a616fb404;
native executable SHA-256 2b624c6bcad5b5bb9bcbb2198258054d53658fd32a5ab67198d15b354bd4d732.

The metadata records base revision 297df69294ac9bbfce0d59deae47724165d7ea63 and repository_dirty: true, because the double-domain gamma_count portability fix and evidence updates were not committed at generation time. Exact source identity is supplied by the sidecar's ordered 20-file manifest: every benchmark/build input has a relative path, size, and SHA-256, plus the aggregate above. Configuration and native binary hashes provide the remaining run identity.

The sidecars retain a replayable invocation under execution.command. Published artifact and native-binary paths are repository-relative, and the command uses the logical executable python; the exact Python implementation and version remain recorded without exposing a contributor's home directory.

Ad-hoc benchmark outputs belong under ignored build/. Files enter this directory only when they are small, tied to a reviewed result, and published with a verified metadata sidecar.

Native scaling 0.4.0

The reviewed measurements in `../../scaling/README.md` publish three compact CSVs with adjacent .csv.metadata.json sidecars:

Profile	CSV	Verified CSV SHA-256
Acquisition versus history	`2026-10-03-scale-ask.csv`	`47f9d9c85add8c06ae9fd0e26edb481e8d02bae4a4ccb717fca52e1d331e05be`
Native throughput, retention, billing	`2026-10-03-scale-native.csv`	`7aa11607484bce378946ce7d66faac68797a8f290be63393e0c4bfc464a763d5`
Paired normalized-regret quality	`2026-10-03-scale-quality.csv`	`c37edecfcae2881c19c63ee6e3f753978ec739117f442ed929d37df96b427574`

The sidecars include exact source and measured binary hashes, compiler/platform, engine/codegen/COCO pins, row counts, and verified CSV hashes. Acquisition rows distinguish baseline snapshot asks from after batch-eight feedback. Native rows distinguish the light million-trial gate, output-only matrix, hourly real-strategy shapes, and the slow magnifier shape. Quality includes the pinned hard-suite replica and 32D extension, with 144 paired studies and zero maximum paired delta at 100/300/1,000 trials. No long-budget quality-equivalence claim is made.

Review-fix evidence (release blocked)

The review-fix report supersedes the original sampler measurements. 2026-10-03-review-quality.csv contains 500 per-seed comparisons across five complete configurations at 3k/10k, with trace hashes, binary hashes, frozen-reference description, normalizer and source provenance. Adjacent review-ask, review-native, and review-ordinal CSVs record flat history probes, finite before/after plus the final million-trial run, and dense versus disk insertion. Every CSV has a verified .csv.metadata.json sidecar. The selected 3k geomean is 1.015738: the required <=1.01 gate fails.

Historical history-switch retry evidence (superseded)

The history-switch report supersedes the review-fix estimator measurements. 2026-10-03-switch-quality.csv contains 100 per-seed comparisons (80 at 3k, 20 at 10k). Adjacent switch-ask, switch-native, switch-ordinal, and switch-identity CSVs record the default-selection and flat history probes, 100k/1M billing runs, dense/disk insertion, and 96 prefix identity studies. Each CSV has a verified metadata sidecar with source/binary/trace hashes, pins and protocol limitations. The historical 3k/10k geomeans are 1.164189/1.118446: both quality gates fail. Identity has zero differing proposal/value records; it is not a long-budget quality-equivalence claim.

Final 1,000-observation switch evidence (bounded opt-in)

The final report records the completed-only 1,000-observation model, now explicit opt-in rather than the default. 2026-10-04-final-1000-quality.csv contains 100 per-seed comparisons: 80 at 3k versus pinned 0.3.0 and 20 at 10k versus the benchmark-only full-history variant. Final geomeans are 1.036492/1.031570; both revised geomean gates failed; these are accepted opt-in tradeoffs. Rotated ellipsoid 20D is 1.245558 at 3k. The earlier 1.6% tradeoff must not be substituted for these measurements. Adjacent ask, native, ordinal, and identity CSVs have verified source/binary/trace/hash sidecars and record the flat-history, billing/memory, membership, and exact-prefix gates.