A local, supplied-evidence toolkit for examining source lineage, evidence independence, external presence and corrective capacity in AI-native information systems.
Status: Phase 3 private analytical core accepted through PR #34. Verification consolidation is in progress; its final four-environment gate and owner acceptance remain pending. Public auditing remains unavailable. The package remains installable as 0.1.0.dev0. The private core implements fifteen analytical families in twenty-nine active modules; nineteen protected modules retain their accepted bytes. Public audit calls raise NotImplementedError; CLI auditing refuses with exit 1 without reading dossiers or producing reports.
The planned v0.1 auditor consumes a caller-prepared local JSON dossier containing claims, evidence, origin and transformation records, evaluator relationships and correction history. It distinguishes supplied assertions, supported structural deductions and unresolved information. Different material counts and evidence roles remain distinct. Missing ancestry cannot become an independent source, and a structurally consistent dossier cannot certify its own authenticity.
The adopted scope provides a multidimensional evidence profile. It excludes a universal truth or integrity score, live crawling, hidden-lineage discovery, an LLM judge, automatic source sanctions and automatic correction. Source locators and extensions remain inert. Source Integrity Toolkit is independent of Recursive Integrity Toolkit, with no shared internals or runtime dependency.
The eighteen Phase 0 specifications remain approved at 7d2e5fcaff591641b5cefce00e71e88941dd1f95; PHASE_0_APPROVAL.md revision 1.1 preserves the authorized digest correction. PHASE_1_COMPLETION.md revision 1.0 was accepted with PR #8, merged at d7d73a790a2a627d19307c9dd48281eff3017023. Its historical pending header is read with that later acceptance event.
PHASE_2_PLAN.md and PHASE_2_COMPLETION.md were accepted through PR #18 at 3a9b75ab6ca4ed9d7c97207043a5c8f54c2e2547. The accepted preparation supports supplied values and UTF-8 bytes, bounded validation, references, time syntax and evidence navigation. The preparation entry points retain their original PC01 admission behavior.
The owner approved PHASE_3_PLAN.md revision 0.1 with 批准，启动 P3-W01; PR #19 merged at 80aa943f577f4a7deaeb8a0f3253c62d1263ca62. PHASE_3_PROGRESS.md records subsequent authority and evidence while prior reviewed documents remain byte-identical. W13 was accepted through PR #32, merged at 332e94a007ceb2a955b8e047f45284b431300995. W14 was accepted through PR #33, merged at 6dbca96f3314d537beed4ccb6202147bd9248dd9. W15 audits the completed component work and records its approved reason-code repair, history-verification optimization, finite CI budgets and remaining acceptance gates in PHASE_3_COMPLETION.md.
The private analysis accepts exact built-in values or supplied UTF-8 bytes and executes actual scoped owners, prerequisites, graph witnesses and immutable result delivery. It preserves supplied assertions, unresolved ancestry, qualified populations, correction history and missing evidence as separate facts. It has a fixed invocation budget and cooperative deadline. A complete supplied dossier can legitimately end with a safe resource interruption; the four original H7 examples currently stop under the unchanged work limits. Partial or interrupted execution cannot certify unperformed fields. These private entry points are implementation seams, with no stable public API or public report format promised.
All 228 field obligations have accepted owner-local core evidence. W14 added source-clause and counterexample checks. W15 identified and repaired the missing dimension_not_selected distinction for empty dimension selection, with exact paired value/UTF-8 tests and separate absent-subject/incomplete-role controls. The 228 historical rows remain unchanged and do not close public-envelope, rendering or release clauses. See phase3/obligation_coverage.json for the per-obligation evidence and remaining duties.
Use a complete Git checkout and a clean CPython 3.11 or 3.13 development environment. Install the existing pinned development requirements, then use the single current verification entry:
python -m pip install -r requirements-dev.txt
python -B tests/scaffold/test_ci_contract.py --local --evidence-dir ../sit-local-evidenceChoose a new evidence directory for each attempt. The entry verifies the actual tool versions, current module/workflow boundaries and frozen Phase 0 baseline, explicitly collects all five directories (scaffold, security, contract, unit and integration), runs the complete suite with third-party pytest plugin autoload disabled, and reconciles exact collected identities with JUnit. Successful subtest events are counted separately. It checks tracked bytes before/after and retains raw logs, failure records, distribution inventories and available installed-runtime evidence. Missing directories, an empty collection, failure/error/skip outcomes or incomplete evidence fail the command. The complete-suite process budget is 2,400 seconds.
This local command requires no SIT_PHASE_UNIT or fabricated GitHub event. It checks the local working tree and records its source hashes. It cannot attest hosted context, authorize a checkout or replace the four required hosted profiles. The retained preparation installation test asserts its complete witness locally, but its persistent W08 receipt is emitted only on an actual Actions run; local evidence states that limit. Use the resulting summary.json and raw logs to inspect the attempt, including a failed or timed-out attempt.
Dependency installation is a developer operation and may access the package index. The installed product has zero third-party runtime dependencies. Package construction, clean offline installation, all 48 installed module hashes, private preparation and analytical witnesses, and public refusals remain tested. The product never imports developer guards, catalogs or tests. Six former historical wrappers now contain their current assertions directly; normal test collection does not fetch or execute historical Git blobs. Some developer boundary tests intentionally use real temporary Git repositories. Building and using the installed package does not require that history.
Bare pytest retains its older scaffold-only configuration. Use the current entry above for complete local verification. Existing targeted pytest invocations are useful diagnostics and must be reported with their actual selected scope.
Current verification keeps Ubuntu 24.04 and Windows Server 2025, each with Python 3.11 and 3.13, immutable Action pins, read-only permissions, no stored checkout credentials, a 50-minute job budget and the 2,400-second suite budget. The seven existing workflow stages remain visible. Evidence reconciliation and artifact upload run with always(); the existing fourteen-day retention and sit-p3/evidence receipt destination remain intact.
CI resolves the actual GitHub event, base and checked-out head. A consolidation PR must use the same-repository verification/consolidation branch targeting main; a main push requires the exact SIT-Verification-Unit: VC footer and matching event head. These labels do not grant authority: the accepted Phase 3 anchor, approved consolidation plan, every post-anchor commit and actual index/worktree must also pass the current scope checks. Missing or inconsistent external context fails closed.
Risk routing uses every parent edge in the actual event change interval, including intermediate edits restored before the final head and merge sides. Only changes confined to the two exact non-executable consolidation records, VERIFICATION_CONSOLIDATION_COMPLETION.md and verification/consolidation_map.json, can use the explicit documentation scope. That scope still checks current authority, complete history, checkout, frozen bytes, module boundaries and workflow policy, and reports that no full suite was run. It does not acquire test tools or claim package/installed-runtime execution. Empty, unknown or mixed changes require the full scope. README changes also require the full scope because README is package metadata. File suffixes and candidate declarations cannot select the reduced route.
Full scope requires all five test directories, exact collection/JUnit identity reconciliation, separately observed subtests, unchanged tracked bytes, pinned installed tools, package inventories and both current-head installed receipts. Each attempt binds evidence to its actual head, event, run/attempt and matrix profile. Raw partial output and failure records remain evidence of failure. A successful documentation check or local check is never reported as full-matrix acceptance.
The accepted Phase 3 anchor is 2413a29b839b7e1de8f76a449762f031de19d52b, with the same tree as directly tested head 90684996eac568af6129973764fe40f3b666a15f. Run 36736103966, attempt 1 passed all four original profiles with 3,963 tests and 428 separate subtest events each. Those are historical Phase 3 results. The completed consolidation candidate still requires its own same-head four-environment run and raw-evidence audit in VC-06. Changes to executable verification, semantics, security, platform behavior, loading or packaging require that gate; strictly explanatory records use the explicit limited scope above. CI success does not automatically authorize acceptance, merge, a next implementation phase or release.
Current progress, collection changes and local evidence are in VERIFICATION_CONSOLIDATION_COMPLETION.md and the one-time consolidation map. The map is a review artifact and is never imported by current verification. Historical stage evidence remains in PHASE_3_COMPLETION.md, phase3/implementation_evidence.json and their immutable Git history. The complete pre-retirement test implementation is available at 3ac5128e4744e2d3708400e3cb1875a9cd4ed0e1; inspect any historical replay in an isolated checkout with its original tooling and label it as historical investigation. Earlier failed attempts and missing W01-W07 raw ZIPs and some W08 profile artifacts remain disclosed.
Windows Server component tests do not certify a future Windows 11/NTFS native adapter. Public audit/report output, CLI auditing, native file adapters and release remain unfinished. The product's 60-second cooperative deadline and existing work quotas are unchanged; a safe resource interruption is not certification of unperformed analysis.
Start with V0.1_PRODUCT_SPEC.md, DEFINITIONS_AND_UNITS.md, CLAIMS_EVIDENCE_AND_LINEAGE_SPEC.md and OBSERVABILITY_AND_REPORTING.md. THEORY_SOURCE_MAP.md, THEORY_TO_CODE_TRACEABILITY.md and VALIDATION_PLAN.md preserve source interpretation, responsibilities and future behavior tests. Historical catalogs and the approved plans remain unchanged; implementation evidence is tracked separately.
See CONTRIBUTING.md and SECURITY.md. Do not post private evidence, credentials or protected-source identities in this public repository. Use fictional development examples. Private vulnerability reporting remains subject to the existing unverified-channel limitation. Build/test tools and their original reviews remain in scaffold/toolchain_review.md and scaffold/ci_toolchain_review.md; the current scoped transition is recorded in phase3/ci_review.md.
Copyright 2026 Xiangyu Guo.
Original engineering repository materials are licensed under the Apache License, Version 2.0, in LICENSE, with attribution in NOTICE. This covers original engineering specifications, documentation, fictional examples, code, schemas and tests that the project has authority to license.
The grant does not automatically cover theory papers or their extracts, page images, translations or adaptations; third-party material; user-supplied evidence; confidential identity maps; or input-derived report content. Theory papers retain their own CC BY-NC-ND 4.0 notices. Excluded materials retain their own terms and permissions. Successful processing does not transfer rights.
LICENSING_NOTES.md controls these boundaries. Its historical root-license statement refers to Phase 0; accepted P1-W01 applied the license without changing that frozen specification. Upstream contribution/governance rules add no downstream conditions to Apache-2.0. No package-index release has been published by this work.