[Clear1nsight]

Change and scope

Adds a standalone, source-bound lossless evidence archive with create/verify/restore operations and an independent exact-restoration oracle. It conserves the stabilized archive, oracle and all original finite controls. Only research/lossless-evidence-archive is added (15 files); no automatic retention hook or application caller is changed.

The retained historical review records 27 runs / 1,179 files: 743,333,823 original logical bytes represented by 63,103,336 archive bytes (91.5108% representation reduction). This is a historical result, not a newly replayed payload. All originals remain retained: zero workspace bytes reclaimed. No scientific runtime speedup, C implementation, kappa gain or completion of the six-area program is claimed.

Task and overlap

Base: 2485805. Head: 3360154. Published Git tree: 6862a7448e904886f6f4d2f0e7614961a6de00e2, equal to the locally validated tree. TASK.json preserves the exact replay-runtime card, changed scope and partial outcome; the original same-fixture throughput acceptance gate is not claimed. Predecessor #6029 and local H25 source/independent-review owners are credited. This contributes a reproducible standalone archive boundary without editing the frozen predecessor package or replaying protected historical payloads. Companion #6316 is independent and not a dependency.

Validation

Linux, CPython 3.12, standard library, filesystem support for Linux renameat2 no-replace and user xattrs for adverse controls. From repository root:

python3 -B research/lossless-evidence-archive/verify.py
python3 -B -O research/lossless-evidence-archive/verify.py

Independent source review accepted adapter/source conservation. Fresh serialized parent-owned normal and optimized runs each passed all 98 original cases: 5 positive cases and 93 expected STOP cases. Controls include independent complete fixture restoration, exact paths/bytes/ordinary modes, separate restored inodes, collision/no-replace behavior, corruption, unsupported features and retained failure receipts/journal agreement. Final exported source also passed all 98 cases with source files/directories read-only, empty PATH and no Git checkout dependency. Package metadata and source pins passed.

Each registered command has a 180-second external cap; the adapter retains its 175-second alarm and 1 GiB address-space cap. VALIDATION.json records execution receipts/hashes, and retained reviews distinguish historical full-payload evidence from fresh finite fixture checks. Source/fixture conservation and expected failures remain explicit. Hosted CI has not yet completed at submission. Historical physical timing gaps and the administration/publication tail remain unknown, not zero.

Credit and next questions

Chris R / Clear1nsight; substantial OpenAI Codex assistance. Apache-2.0; predecessor, archive-author and independent-oracle credits retained in source/evidence manifests.

The supported restoration domain is ordinary files/directories and declared modes; links, special files, ACLs and source xattrs are excluded. Smallest useful follow-up is an independently budgeted complete restore of a newly authorized closed run tree in another environment. Any automatic retention integration first needs a separately reviewed writer-closure, selection and finalized-receipt contract. Neither follow-up is included here.

[Clear1nsight]

Publication handoff — hosted checks complete

• Result: standalone lossless archive/create/verify/restore with independent exact-restoration oracle. No automatic caller integration; PR6029 is unchanged. Companion Bound retained-log memory and preserve fast-exit output-cap checks #6316 is independent.
• Source binding: exact head 336015471e4f1bb95998f293aa76823fe11e33d5, tree 6862a7448e904886f6f4d2f0e7614961a6de00e2, parent 2485805a6d47f36ea83ddde6327d453ac82d44dc; all 15 committed files match the tested read-only export.
• Hosted validation: run 41136304437 passed plan and replay at this head. Shared suite: 710 tests, 24 platform skips. Selector chose only research/lossless-evidence-archive; CPython 3.12.15. Committed inputs remained unchanged.
• Coverage: all 98 original finite cases passed in both normal and optimized modes: 5 positive and 93 expected STOP per mode, including independent exact fixture restoration and adverse failure controls. Local normal/optimized and final read-only/empty-PATH export checks also passed. Protected historical payloads were not reread or replayed.
• Cost boundary: the recorded 91.5108% representation reduction belongs to the preserved historical 27-run/1,179-file result. Zero original files removed and zero workspace bytes reclaimed. No verifier throughput or scientific speedup claim. Historical timing gaps and final physical publication tail remain open; hosted execution is separate from the local allowance.
• Credit: Chris R / Clear1nsight, predecessor/archive/independent-oracle authors; substantial OpenAI Codex assistance, Apache-2.0.
• Remaining: independent full restore in another environment and any closed-writer automatic retention contract require separate work. This PR is open, unmerged; broader six-area work remains paused and incomplete.