{
  "actor": "research-problem-resolver",
  "contract": "research-completion-abi",
  "explanation": "Astra resolution-only decision, session run-b9c736bb16b1c5ea8f4a96a5. Observation/provenance: archived verify/results/research-completion-abi/objdump-2026-09-09-turnevents.txt SHA256 f95bbd2674aaa074feb70ebb13ec4877b3ceef3d2aafe5810b8b0ec337cfbb8b lines 211-213 contains lea candidate+8, lea stored+8 and call 0x0046f8c0. It jumps directly from 0x00825e19 to 0x00825e2e in the selected excerpt: the disputed test/branch is NOT archived there. No helper body exists in the four-file ABI results package. Surprise says true AL matches; checkpoint 26d1c4fa6bd58b6a442e0241 and handoff say false. Neither polarity is independently established by this archive. This is an evidence-completeness/conflicting-interpretation problem, not measured game behavior or a stale-build mechanism. Identity: assigned engine HEAD 7741d42fc5e4e761e6449bdaf0e4a61d00036a23 and RE HEAD 3bfde5a70d874a723e797a695bbd847fd82c0aa7 and canonical common dirs independently checked; both dirty inventories observed. Source-before manifests are in runtime/runs/run-b9c736bb16b1c5ea8f4a96a5.json; no fresh content binding substituted for old measurements. Input SHA256 970b7de729956a53094c7eb98aba4270aee98e2fed5daf0d39e290013c90c841 and matching ReVa MD5 are recorded provenance only, not freshly rehashed here. Campaign checkpoint rehashed the raw excerpt and index successfully. No live positive execution, runtime binary/input binding, or hooks-off neutrality measurement exists in this review. Read-only archive inspection has no game-side instrument effect; runtime neutrality remains unproved. CLAIM DISPOSITIONS, exact domains: (1) SURVIVES: the archived x86 excerpt loads two +8 addresses and calls the named helper, and shows stride 0x74 and a matched-element return; static excerpt domain only. (2) QUALIFIED: events.md:182-185 claim that EvDsc is not compared; suspended as an implementable dedup rule, not conclusively falsified without helper semantics. (3) QUALIFIED: new handoff/index prediction that description-only difference prevents dedup; hypothesis only until complete helper and caller control flow are recovered. (4) OVERTURNED: assertions that this archived excerpt proves true-AL or false-AL match polarity, or is sufficient raw evidence for a complete comparator; the relevant instructions are absent. Actual polarity remains unresolved. (5) QUALIFIED: wrapper writes nothing/no RNG; no local write is shown, but transitive helper effects are not established. (6) SURVIVES UNAFFECTED, not newly verified: separate recovered container strides, ownership/allocator-family requirements, get-or-create/prune hypotheses and original-dependency limits; this conflict does not invalidate their independent raw observations. (7) SURVIVES: research-replacement remains proposed, W1-only historical observations do not establish dedup branch coverage, and full record/ID/text/allocator effects plus original dependency declarations remain mandatory. ABI ownership-recovered and independent-cross-check criteria are unchanged and NOT passed. INVALIDATION: withdraw dedup polarity/semantic conclusions in campaign/research/research-completion-abi.md:43-48, recovered-static.md dedup interpretation, reva-2026-09-09-index.md:34-37,52-54, the original surprise summary, analyst checkpoint 26d1c4fa6bd58b6a442e0241, and findings/subsystems/events.md:182-185 as current implementation authority. Preserve raw bytes and historical files; invalidate interpretation, not the recorded call bytes. No registered evidence array or accepted pilot verdict is supplied here; controls clear this contract evidence/checkpoint and invalidate any verdict. Revised scope/dependencies: resolve to needs-revision solely to enable lead-scheduled analyst read-only evidence repair within existing owned results/handoff scope. No implementation or lab work is released. Before ordinary analyst resumption lead must reconcile checkpoint/owner/basis and lifecycle under campaign controls. Dedup implementation and semantic promotion remain blocked on complete source/input-bound static recovery and independent cross-check; runtime claims additionally require positive branch execution and neutral controls. No acceptance waiver, new original-helper policy, or architecture/source edit. Integrator must link this decision and regenerate projections after integration; do not paste another history into the board.",
  "id": "d-2ff30c9f5355116bea822924",
  "invalidated_checkpoint": "campaign/runtime/checkpoints/research-completion-abi-fd414ea958fd39c18cee1964.json",
  "invalidated_evidence": [],
  "model": "openai/gpt-6-astra",
  "probe": "Cheapest discriminating action: lead schedule the ABI analyst to rehash dumps/sots.exe against SHA256 970b7de729956a53094c7eb98aba4270aee98e2fed5daf0d39e290013c90c841, bind paired source manifests, checkpoint, then archive UNFILTERED objdump output for 0x00825d40..0x00825e65 and helper 0x0046f8c0 through all returns (follow any callees), with raw bytes, exact command, tool identity and exit status. Recover push/ECX conventions, test/branch polarity and short/long string semantics; mismatch or missing bytes stops interpretation. This is read-only local static work, not permission to mutate VM/Ghidra. Independent verifier reproduces the complete windows. Only after semantics and ownership are pinned should an explicitly assigned leased lab experiment compare same-bucket events equal in action/location/finite position/message/image, first identical descriptions then description-only difference, including short/long strings. Require positive comparator entry/return evidence, full pre/post records and IDs, exact binary/input/source identities and neutral controls; helper/branch analysis predicts outcomes but is not runtime measurement.",
  "role": "resolver",
  "schema": "sots-decision/1",
  "surprise": "s-8996365dab2cd6dc0e17bb9f",
  "timestamp": "2026-09-09T23:52:51.024636+00:00"
}