[mark-e-deyoung]

Project-native authority: ghidra-tools DLE #4, mission-command plan, P0 accepted only for same-run consistency. DRAFT child stacked on current draft #546 exact head ec218499a73e65fac07adf04ed6e934929977992; no merge, release or model admission.

Mission outcome: take the existing public generated stripped x86-64 ELF before temporary deletion; directly rehash actual target bytes in a different Python module, launch two fresh native GDB processes per optimized branch-seed image (fixed RDI n=0 and n=1), attest entry+target breakpoint installation, entry actual PC, initial RDI7, intervention RDI n, ordered native target hit or authenticated clean exit. Check immutable SHA before/after replay and correlate generated entry and branch target RVAs from independent GNU oracle. Compare this native second measurement with the primary GDB per-island observations in the existing four-case P0 reducer, and fail closed on disagreement. New adversarial unit tests cover contradictory terminal, wrong PC, wrong RDI, event time order, corrupted/replaced ELF, signals and duplicates. No arbitrary user-file CLI; verifier is imported only inside the existing fixed synthetic source producer and has a bounded ELF sanity/header/size check.

Explicit independence limit: second measurement is a new debugger process and direct file hash but shares the public CI host, GDB, generated fixture source and source-known target address. It is not external peer grading, hostile stdout authentication, sealed blind evaluation, model generalization, an admitted arbitrary-binary executor or a release approval. Preserve original Ghidra/GNU/native matrix and seven-file P0 parent change rather than replacing it.

Authority: free public standard GitHub Actions existing job; $0; zero private GitHub-hosted minutes; no credentials, protected/private binaries, new corpus, privileged host, Sidecar admission, model spending, new runner, merge, release, or source/license custody transfer. Private ghidra-tools PR #5 project evidence contract remains draft/unaccepted. Qualification requires exact-head CI and a separate independent authorized reviewer.

[mark-e-deyoung]

Exact-head final CI reconciliation and bounded source-adversarial review (2026-10-25)

Authority: project-native mark-e-deyoung/ghidra-tools#4; experiment PR #550 is DRAFT/UNMERGED. This is a technical checkpoint, not external peer approval, product/schema acceptance or a request to merge.

Resolved final-head hold: At PR #550 source head 7c958b740a387e4eeea326de01ebeee8384b176f based on #546 ec218499a73e65fac07adf04ed6e934929977992, public Actions run 40210404402 initial job 121220220274 was CANCELLED during Ghidra/JDK dependency setup; it was NOT a tested-success final head. Subsequent retry job 121230190623 finished SUCCESS. Independently retrieved raw job log confirms checkout Merge 7c958b740a387e4eeea326de01ebeee8384b176f into ec218499a73e65fac07adf04ed6e934929977992, 144/144 tests PASS in 13 groups, and native Ghidra/GDB second replay markers for O2-seed3 and O2-seed2, each n=0 NATIVE_PC_WITNESS and n=1 NOT_OBSERVED_SAMPLE at exact image RVA 4200 / 0x1068. No failed steps on retry. Machine JSON is explicitly scoped P0_FOUR_CONTROLS_CONSISTENT_NOT_SEALED_BLIND, FRESH_GDB_PROCESSES_SAME_HOST_SAME_INSTRUMENTATION, source_receipt_sha256=d6c034f4130ab760b24c9b52818e17035cee1b5f1ee6823bd7733f7612dd8338. Do not interpret n=1 nonhit as unreachability or same-host replay as independent external attestation.

New bounded red-team finding, medium/low within fixed synthetic source, must correct before claiming general ELF profile integrity: scripts/reversing_benchmark_native_replay.py:replay_fixed_stripped claims to check a public x86-64 ELF but presently reads six header bytes only (\x7fELF\x02\x01). That proves ELF magic/class/data, not e_machine=EM_X86_64 (62), e_type=ET_EXEC (2) for the fixed -no-pie target, or header version. A 64-byte synthetic little-endian AArch64/RISC-V ELF header with that same prefix passes the existing pre-GDB architecture gate. This is not a claim that native CI actually executed the wrong architecture: in this public, source-known and independently compiled target, real x86-64 GDB and runtime observations provide additional constraints, and successful exact-head CI remains valid for its documented narrow sample. But the helper's claimed static identity check is overstated.

Reproduction: independently modeled the exact six-byte predicate, generated ELF64 little-endian headers varying e_machine 62/183/243, e_type 2/3/1, ELF version and truncation; original six-byte guard accepts the inappropriate variants while a minimum fixed-profile gate using stdlib struct.unpack_from('<H',header,18)==62, e_type==2, version==1 and length>=64 rejects them. 9/9 local review tests PASS, including a real gcc -fno-pie -no-pie compiled ELF positive. These are a disposable independent model of the guard rather than tests against the exact repository module or native GDB; this distinction is material.

Next small corrective action: require actual ELF header machine/type/version under the existing 2 MiB file bound and update the existing mocked ELF test fixtures to syntactically correct 64-byte ELF headers; fail closed before GDB. A GitHub code change needs a new exact-head CI, independent reviewer, and no automatic merge. Do not create a general ELF loader, service, new schema, or privileged runner for this.

Unchanged gates: $0 public standard runner; no private source/artifacts, no new Sidecar admission/keys/private Actions minutes; true independent peer, assistive-tech observation and private project-schema acceptance still pending.

[mark-e-deyoung]

Corrective verification closure — exact-head native CI (2026-10-25)

This resolves the narrow ELF-header source validation defect identified in comment #6400494317 without changing the scientific outcome or expanding the execution envelope.

Current PR #550 DRAFT HEAD: 9205cb6a5345b55d8cfc8fe6bd2cfad37fa1bcf9, still stacked on draft #546 ec218499a73e65fac07adf04ed6e934929977992. Two files corrected: scripts/reversing_benchmark_native_replay.py and tests/test_reversing_benchmark_native_replay.py. Before invoking GDB, the fixed public synthetic helper now checks 64-byte ELF header, 64-bit class, little endian, version 1, ET_EXEC=2, and EM_X86_64=62, consistent with its existing -no-pie compiler profile. Existing mocked fixtures were updated to minimal valid headers, and tests challenge wrong architecture (AArch64/RISC-V), wrong object type, wrong version, and truncation. No new tool, workflow permissions, runner family, binary input, credential or credential-exposure path was introduced.

Exact-head public free CI run #40214623536, job 121232973797: COMPLETED SUCCESS. Independently read raw job log, checkout Merge 9205cb6a5345b55d8cfc8fe6bd2cfad37fa1bcf9 into ec218499a73e65fac07adf04ed6e934929977992. Thirteen test groups total 147/147 PASS, with the native-replay group 14/14 PASS. Real pinned Ghidra/GDB O2 matrix still reports:

• O2-no-unwind n=0 NATIVE_PC_WITNESS, n=1 NOT_OBSERVED_SAMPLE, target image RVA 0x1068.
• O2-seed2-no-unwind n=0 NATIVE_PC_WITNESS, n=1 NOT_OBSERVED_SAMPLE, same target RVA.
• RE_BENCH_P0_NATIVE_FOUR_CASES_AUDIENCE_PROJECTIONS_PASS and source receipt SHA-256 d6c034f4130ab760b24c9b52818e17035cee1b5f1ee6823bd7733f7612dd8338 present in actual post-execution log.

Evidence class: SOURCE_AND_NATIVE_FIXED_SYNTHETIC_CI_VERIFIED, not strong independent same-host trust, blind source-held-out grading, universal reachability, actual screen-reader qualification, product schema acceptance, or merge. This technical implementation/review remains by the same assistant and does not satisfy the genuinely independent reviewer gate. The original final-head CI run #40210404402 was initially cancelled during dependency setup, then succeeded on retry; it was not evidence of a code defect.

Next bounded DLE baton: reconcile this exact head/CI into project-native mark-e-deyoung/ghidra-tools#4 through its authorized owner; request genuinely separate independent reviewer for #550 and #514; validate private candidate schema #5 with an independent semantic oracle before consequential promotion. Keep all upstream PRs DRAFT/UNMERGED, $0 free standard public Actions only, zero private hosted minutes, no private binaries, no new Sidecar grant.