{ "schema_version": "operator-workflow/v1", "request": "Independently classify the failed b2048/C4 offered-concurrency-8 seed-2801 control and decide whether the optional C8 arm must stop and whether exact b2048/C4 qualification may continue.", "gate_state": "blocked", "targets": { "evaluated_model": "GLM-5.3-Flash served as glm53-flash-exl3-runtime-dcp1-candidate", "implementing_model": "GPT-6 Astra", "reviewer_model": "GPT-5.6 Sol", "recipe_scope": "DCP1, batch2048, max-num-seqs4, native context327680, offered concurrency8 diagnostic", "promotion_scope": "exact production C4 recipe only; C8 arm rejected" }, "tools_used": [ { "name": "read_only_file_inspection", "source_class": "local", "ok": true, "dry_run": true, "confirmed": true, "target": "retained campaign artifacts", "error": null }, { "name": "sha256sum", "source_class": "local", "ok": true, "dry_run": true, "confirmed": true, "target": "retained evidence and frozen plans", "error": null }, { "name": "workflow_packet_validate", "source_class": "mcp", "ok": false, "dry_run": true, "confirmed": false, "target": "c4-offered8-failure-independent-review.json", "error": "MCP/control-plane validator not available in this reviewer context; JSON syntax checked locally only." } ], "artifacts": [ { "kind": "finding_summary", "severity": "high", "classification": "confirmed", "summary": "The offered-concurrency-8 control failed its predeclared 100% strict/canary gate. Request 15 returned a complete HTTP response with finish=stop but omitted its own marker and emitted 3 requested code words plus 150 extra words. This is a genuine model response-contract failure for this offered-C8 population, not a missing-response or scorer-observability failure.", "references": [ "concurrency-b2048-reload-offered8-seed2801.json:4", "concurrency-b2048-reload-offered8-seed2801.json:14", "concurrency-b2048-reload-offered8-seed2801.json:20", "concurrency-b2048-reload-offered8-seed2801.json:30", "concurrency-b2048-reload-offered8-seed2801.json:35", "concurrency-b2048-reload-offered8-seed2801.json:38", "concurrency-b2048-reload-offered8-seed2801.json:41", "concurrency-b2048-reload-offered8-seed2801.json:44", "concurrency-b2048-reload-offered8-seed2801.json:51", "concurrency-b2048-reload-offered8-seed2801.json:57", "concurrency-b2048-reload-offered8-seed2801.json:292" ] }, { "kind": "scope_disposition", "severity": "high", "classification": "confirmed", "summary": "Stop the entire dependent C8 comparison now. Do not load C8, retry seed2801, substitute another seed, or compute a partial comparison. The frozen arm required 100% strict/canary compliance and both reserved pairs eligible; the general estimator says any failed cell stops dependent comparisons.", "references": [ "concurrency-arm-plan.json:21", "concurrency-arm-plan.json:22", "concurrency-arm-plan.json:28", "matched-performance-plan.json:63", "optimization-search-v5.json:67", "optimization-search-v5.json:76" ] }, { "kind": "promotion_scope", "severity": "medium", "classification": "confirmed", "summary": "This out-of-envelope offered-C8 failure does not by itself disqualify the unchanged b2048/C4 production recipe. The required production capacity gate is C4, and the optional optimization plan separates final C4 qualification from the concurrency axis. Continued exact-C4 agentic 18/18, SWE >=3/5, images, natural/64K diagnostics, routed policy, and real-client gates is defensible only after retaining this scope ruling. Promotion remains blocked until every applicable original gate passes.", "references": [ "campaign-plan.md:5", "campaign-plan.md:12", "campaign-plan.md:13", "campaign-plan.md:15", "campaign-plan.md:16", "optimization-search-v5.json:55", "optimization-search-v5.json:56" ] }, { "kind": "causal_limits", "severity": "medium", "classification": "refuted", "summary": "The evidence does not support attributing the semantic miss to overload, runtime failure, or cache leakage. Identity was stable; the owner remained running; OOM and kernel fault counters were zero; all requests received HTTP 200. Zero foreign markers only refutes observed cross-request marker leakage in this sample. Logs show 4 running and 4 waiting under offered8/max-seqs4, but that scheduling fact is not causal proof.", "references": [ "concurrency-b2048-reload-offered8-seed2801-owner.json:18", "concurrency-b2048-reload-offered8-seed2801-owner.json:24", "concurrency-b2048-reload-offered8-seed2801-owner.json:31", "concurrency-b2048-reload-offered8-seed2801-owner.json:32", "concurrency-b2048-reload-offered8-seed2801-kernel.json:2", "concurrency-b2048-reload-offered8-seed2801-kernel.json:3", "concurrency-b2048-reload-offered8-seed2801-identity.json:2", "concurrency-b2048-reload-offered8-seed2801-identity.json:12", "concurrency-c4-failure-owner-logs.txt:65", "concurrency-c4-failure-owner-logs.txt:138" ] }, { "kind": "residual_risk", "severity": "low", "classification": "plausible", "summary": "The cgroup peak equals its configured limit and memory.events.max is nonzero, so resource pressure cannot be excluded; these counters are cumulative and do not establish that pressure caused request 15. Retain this limitation without causal wording.", "references": [ "concurrency-b2048-reload-offered8-seed2801-owner.json:22", "concurrency-b2048-reload-offered8-seed2801-owner.json:27", "concurrency-b2048-reload-offered8-seed2801-owner.json:31" ] }, { "kind": "production_condition", "severity": "high", "classification": "plausible", "summary": "A C4 promotion is supportable only if routed admission is independently shown to enforce the declared production max_concurrency4 and no publication or harness claims support for offered concurrency8 or max concurrency8. The direct offered8 observation must remain a disclosed failed diagnostic.", "references": [ "campaign-plan.md:16", "concurrency-arm-plan.json:23", "concurrency-arm-plan.json:29" ] }, { "kind": "breakage_probes", "items": [ { "axis": "unreadable state/tools", "input_state": "owner, kernel, identity, or gate sidecar absent/unreadable", "wrong_result": "classify model-only failure or permit promotion without runtime/identity evidence", "result": "refuted; required sidecars were readable and hash-bound" }, { "axis": "malformed/boundary input", "input_state": "complete response missing own canary and violating exact 16-word contract", "wrong_result": "count request as eligible or calculate partial throughput", "result": "refuted; native artifact failed request and made aggregate performance ineligible" }, { "axis": "resource exhaustion/bounds", "input_state": "offered8 against max-seqs4 with queued requests and cgroup max contacts", "wrong_result": "infer overload caused the semantic miss or ignore timeout/resource bounds", "result": "plausible contributor only; no causal attribution, no OOM/Xid/crash" }, { "axis": "state drift/seams", "input_state": "container/image/revision/recipe changes before versus after cell", "wrong_result": "compare different identities as one control", "result": "refuted; before/after identity matched" } ] }, { "path": "concurrency-b2048-reload-offered8-seed2801.json", "sha256": "0e16e02e1e41609f74833ee0fd4bbcb4363573d74ba0b8f2fb279427458aca3b" }, { "path": "concurrency-b2048-reload-offered8-seed2801-gate.json", "sha256": "e2f5d64a8bc6712db6b608878d38c74b42ecdc3478b34cda83cae7c70fcf6a31" }, { "path": "concurrency-b2048-reload-offered8-seed2801-owner.json", "sha256": "9635bc628563afc693bd786e50ae9ee9dba2f08e5c75cb1c1a9588cfe83f432e" }, { "path": "concurrency-b2048-reload-offered8-seed2801-kernel.json", "sha256": "3e166338158867bbf1db7e228699bc37208227d1e6ac47bfa4c504523c1ced42" }, { "path": "concurrency-b2048-reload-offered8-seed2801-identity.json", "sha256": "c6845c7e9360e81fa945674eb25c31a6c20d714491f9e623b7eb4f39da180ccf" }, { "path": "concurrency-c4-failure-owner-logs.txt", "sha256": "0a63ebf359530e520094ec3e9620d0688529a9f5b5d5f13a469b20336413d013" }, { "path": "concurrency-arm-plan.json", "sha256": "50155c6002d3b75b05f4493908838e0f57faff223425673bc8e3221d42ea0d51" }, { "path": "optimization-search-v5.json", "sha256": "c73d0a3330b8099b1413bc34b854f3782532addf37b11fce155ee68cf4bfee34" }, { "path": "campaign-plan.md", "sha256": "c4c3cb9011e54bfdadbbcf56a9ade78c31e129ca4d4132e780bd1a728d30938d" } ], "advisory_priors": [ "Direct offered concurrency8 exceeded the declared production router concurrency4; this is scope context, not a causal explanation.", "The failed synthetic controlled-output cell measures deterministic instruction adherence and isolation, not general coding quality.", "Generating model did not grade itself; this review is by independent GPT-5.6 Sol of Astra-authored orchestration and GLM output." ], "recommendation": "needs_more_data", "human_gate_required": false, "promoted": false }