marin-community/verifier-unification-evidence

Verifier unification campaign evidence

This archive records the grading migration across verifyit, MarinSkyRL, Harbor, Evalchemy and lm-eval-harness. It includes source-policy findings, patches, independent reviews, controlled and archived-output comparisons, installed-byte checks, CI results, failed attempts and incomplete work. It is an evidence record, not a certification of every grader or a new source release.

Start with the per-grader index. It covers all 196 campaign ledger identities and links each recorded contract, motivation, historical patch identity and validation report. The machine-readable breadcrumbs preserve constraints, proof kinds and missing associations. The campaign ledger is the original status source; acceptance, architecture review and installed deployment are separate gates.

Code and scope

The related changes are verifyit PR6, SkyRL PR900, Harbor PR228, Evalchemy PR198 and the Marin port PR9697. External adapter adoption remains separate from the Marin port. Native grading stays the default; client opt-in flags select migrated routes.

Client publication identities for this snapshot are recorded in the final Sky base-class publication manifest, Harbor publication manifest and Evalchemy publication summary. Use these manifests rather than assuming that a historical proof commit is the latest branch head. Grader proofs retain the exact client/core pins they tested.

Frozen snapshots and patches are also archived. Their presence does not mean they were committed, deployed or accepted. In particular, Harbor's parked core-dependent client proposals and Evalchemy's unpublished core-dependent proposals were parked when the core boundary was locked. The later client-only cleanup uses published APIs. Do not substitute these proposals for shipped code.

The published core patch index covers all 15 modes plus shared spec, reward, execution, artifact-read and preparation boundaries. It distinguishes the shipped c391 snapshot from frozen uncommitted proposals and names regressions without inventing a dedicated cause record.

SQL behavior validation does not imply architecture completion. The published Sky SQLite helper still defines rounding, ordering/duplicate and protocol-acceptance policies and combines checks in client code. Core primitives execute those checks, but the full policy and grading composition are not yet centralized. The archived shared-runtime proposal was parked, not shipped.

Initial inventory and excluded populations

The 196-route migration ledger is not the initial source census. Read the TaskTrove mapping, all-row cohort inventory and converter/drift contracts for that separately excluded population. These are pinned source inventories and preparation investigations, not proof that every retained TaskTrove task was executed.

The lm-eval-harness mapping, source inventory and native eligibility inventory distinguish recognized configurations from actual evaluator execution. The native evaluator parity report is a separate validation record. A historical LiveBench coding patch manifest describes an incremental snapshot patch chain; it is not an independently applicable upstream-main patch. The harness integration PR1 is merged at e2c370ef, from tested head ea39286. This is separate from certification of all task configurations.

The coverage register, reconciled harness census and inline-declaration clarification explain the static denominator and unmapped populations. Earlier native/source inventories contain intermediate counts. These source inventories and historical execution counts must not be substituted for later route acceptance statuses.

Patches with direct finding-to-evidence links

Grader or boundary	What changed and why	Validation and limits
BIRD SQL	Replace the source-grader Script bridge with SQL/row capture and core Schema, Exact and ALL. Make read-only, finite, 100000-row and reference-first admission explicit.	Acceptance report and installed independent review. Three controlled native/enabled Trial pairs; no archived-model claim.
ACEBench	Compile normal and special source policies into Schema/Exact; use shared specificity-ordered first-fit and alternative aggregation. Preserve ordering rather than replace it with perfect matching.	Acceptance report and frozen independent approval. Three controlled Trial pairs and explicit finite/bounded restrictions.
BFCL	Replace grader delegation with source JSON views and Schema/first-fit. The initial review exposed trusted scalar-domain admission before output replacement; an actual failed Linux build exposed the incompatible enabled uv installer pin.	Independent v2 approval and paused partial index. Only 26 of 78 Trials completed, 13 pairs across five categories. Not accepted as a complete 13-category deployment proof.
Dolci stdio	Remove redundant normalization and raw debug serialization. Failure review exposed trusted text in unsupported-policy receipts; publish only safe policy labels, digests and redacted failures.	Frozen patch and independent review. Actual scored/wrong and partial-positive failure controls preserve status and stale-reward removal. No new Docker claim.
Core worker/import boundary	Clarify bounded worker scope and test large-result draining/privacy; migrate direct imports while preserving typed status transport. The runtime transport was retained after deadline/cleanup assessment.	Final index and independent review. Historical PATH-dependent failures remain recorded alongside the corrected test run.
Evalchemy client cleanup	Reuse the published JEE capture record and existing strict literal helper; remove intermediate raw preparation serialization. Keep source extraction and grading behavior unchanged.	Publication summary and independent source comparison. No new grading coverage or installed Olympiad acceptance.
Sky environment base	User review identified repeated opt-in flag reads and reward-error hooks. The shipped patch centralizes the base flag and Search/SearchCode reward boundary; task-specific tools and grading remain route-owned.	Final publication and test manifest and independent review. 25 search regressions and four IFEval factory controls; broader local coding failures require the missing sandbox and match the prior baseline. Exact-head CI37069773789 passed all six lanes, including the real NeMo sandbox; earlier green or pending CI snapshots remain historical. No new coverage.

The Harbor/core map names the exact historical commits for these patches. The final Sky route map covers all 46 Sky routes, including dormant, unsupported and unaccepted work. The full per-grader index preserves remaining Harbor and Evalchemy contracts and evidence. When no dedicated finding-to-commit association exists, the breadcrumb record says so; a native contract is not evidence of a specific bug.

Reading the evidence

A source-function differential checks preparation or grading semantics. A controlled native/enabled framework Trial checks an executed framework path with a controlled candidate. An archived-output comparison reuses retained model outputs; it does not imply a new model run. An installed-byte check binds files to immutable source. CI checks the repository head but does not certify all datasets. Read each report's proof kind and intentional deviations before comparing results.

Historical and failed attempts are retained. A later final/acceptance index identifies the successful proof and its immutable revisions. Earlier freeze versions, unsuccessful builds, timeout logs, environment failures and parked proposals must not be counted as successful runs. Some route directories are referenced historically without a resolved standalone report; these are explicitly listed in link-check.json.

Key reports and these indexes are loose files for browsing. Bulk traces, logs, source snapshots and data are packaged separately. Logical paths retain their original artifacts/, evidence/ and notes/ names; the packaging manifest maps them to archive members and hashes. A linked report is not a promise that every adjacent raw payload is publicly available. Missing or redacted material must remain explicit in that manifest.

Archive layout and sanitization

The archive maps 268,563 logical campaign files into 64 bulk shards plus 189 selected loose evidence reports. Generated navigation and packaging manifests are additional files. manifest.jsonl maps every original logical path to a loose file or archive member, with original and staged sizes/hashes. checksums.jsonl identifies upload-file bytes. late-capture-manifest.json records later captures, including final Sky CI; earlier pending-status snapshots remain historical.

The sanitization manifest records 39 distinct redaction paths and 51 logical files whose staged bytes differ from their originals, including normalized or rebuilt archive containers. Thirty-five unpacked cached files and ten cached archive members were excluded. Changes remove credential literals and signed URL query credentials; original source files were not edited. The structured parquet cleanup preserves all 7,500 rows and schema. No blanket removal of candidate text, prompts or trusted references was performed.

The final staged credential scan reports zero remaining strict findings. It classifies 211 broader assignment matches rather than treating every token-related word as a credential. A documented 57-character private-key fixture is intentional test data. This is a recorded scan result, not a guarantee about every possible secret format. Exact changes and non-byte-exact copies remain traceable through the manifests.

Environment exclusions, symlink dispositions and external evidence metadata distinguish rebuildable dependency environments, links and externally referenced material from retained proof. Original third-party source/data provenance and license notices remain authoritative; no blanket redistribution license is assigned to embedded material.