Isolated reasoning-limit probe

The frozen request packet, prepared October 25, 2026, defines two synthetic requests and retained runtime pins. Its historical authorization and execution fields remain unchanged. Current execution requires explicit --confirm, fresh runtime evidence and the operator checks below. Ignis remains held.

The earlier GLM diagnostic exhausted total completion caps of 16,384 and 32,768 tokens below the configured reasoning budget of 49,152. It therefore did not exercise numeric reasoning-budget enforcement. Prompt-level reasoning_effort=max is a separate control.

Offline preparation

The packet fixes total output at 512 tokens and reasoning at 64, with C1, temperature 0.6, top-p 0.95 and the retained MTP3 configuration. The first request is a small arithmetic puzzle with an independent answer oracle. One predeclared harder puzzle is permitted only if the first naturally finishes reasoning below 64 tokens and produces a coherent answer. Natural completion below the boundary leaves enforcement untested.

tests/test_reasoning_probe.py is the current hermetic regression check for the diagnostic runner. It writes a synthetic router configuration and captures the request body at an in-process endpoint, while checking the bounded deadline, immutable artifacts, operator-attested execution flags, and sanitized failure behavior. Run it with python scripts/run_tests.py tests/test_reasoning_probe.py -q. It does not prove engine enforcement or authorize running the packet.

Repeatable commands

First collect read-only runtime and resource evidence for the exact managed recipe container (this runs a fixed CPU tokenizer inspection, never inference):

anvil-serving eval reasoning-limit-inspect --registry recipes.toml --container candidate --plan docs/benchmarks/reasoning-limit-probe.json

The inspection requires the pinned GLM recipe, verifies its owner/digest and immutable container identity before and after capture, and returns source and template hashes plus tokenizer markers. Compare every source hash with the pinned source before traffic. It uses the engine's patched HF tokenizer path, with local files only. Raw command arguments, arbitrary environment variables and mount paths are excluded. RAM availability/pressure, swap activity, GPU usage, cgroup memory and engine running/waiting gauges accompany the capture.

Record router_config_sha256, the exact expected_model, and nonempty integer arrays start, end, forced_end in a private boundary.json. The command currently supports identical natural and forced closing markers only. Preview without network or writes:

anvil-serving eval reasoning-limit-probe --router-config router.toml --plan docs/benchmarks/reasoning-limit-probe.json --boundary-config boundary.json --output probe-evidence

After independently checking the fresh inspection and source parity, retain a private execution copy of the plan with execution_authorized=true and live_identity_verified=true. Record its hash together with the inspection and source-parity hashes in the operator's verification record. Confirmed execution requires that reviewed copy; it refuses the historical false/false packet. Run the command with --plan execution-plan.json --confirm. Those fields record the operator's verification; the request runner does not validate that separate verification record or claim automatic source attestation. Its executed status and zero exit code mean transport execution completed, not that the boundary or answer gate passed; inspect boundary, oracle, and response_anomalies.

The evidence directory must be new. Request starts, exact outbound JSON, bounded raw responses and terminal records are retained with owner-only permissions. No shared router configuration is changed: only an in-memory tier copy gets the 64-token budget and token-ID capture controls. The selected endpoint and its credential reference remain the same. This direct diagnostic does not share the running router's admission counters. It checks the exact served identity and zero running/waiting occupancy immediately before each request. This is an idle observation, not an exclusive scheduler lease; run only during a quiet window. Stop if occupancy or memory pressure makes the canary inappropriate.

Execution prerequisites

Notify the user before inspecting or sending anything to the running model. Then verify package/template bytes through the managed owner surface against the packet. A version string or image label alone does not prove byte parity.

The existing shared route hard-overrides the numeric budget. Sending a caller field does not bypass it. A separate, explicitly configured managed request path to the same endpoint must be prepared and reviewed before execution; the offline test does not install that path. Preserve the shared route, model, provider, sampling, history policy and MTP settings.

Capture includes finish reason, nullable native usage, prompt token IDs and full generated token IDs. The prompt opens the reasoning marker, so classification uses the concatenated prompt and output. Exactly 64 tokens after the complete opening marker followed by the complete closing marker is boundary-consistent evidence. The API has no forced-versus-natural provenance flag; even that result does not directly prove which kernel branch emitted the marker. Missing IDs, partial/repeated markers or a later close leave enforcement unestablished.

Run and interpretation

Allow at most two sequential requests, zero retries, 120 seconds per request, and 300 seconds including closure. Freeze both requests and source identities before starting. Capture RAM availability/pressure, GPU and cgroup evidence before and after; verify owned requests have drained. Stop on any packet stop condition, retaining the failure and unknown fields. Do not add more cases to make a failing probe pass. Repeat the inspection after the requests and compare identity/configuration, memory/OOM counters and engine running/waiting gauges. Client socket closure is separate from engine drain; the runner leaves the latter unknown until that owner-side evidence is supplied.

Report boundary enforcement and visible-answer correctness separately. A correct puzzle answer does not prove forced closure; a forced boundary does not guarantee a correct answer. Preserve marker counts and transition timing privately; publish only necessary sanitized evidence. No quality or performance conclusion follows from this two-request canary.

The pinned fork's forced-logit sampling path warrants specific inspection: upstream PR 58231 fixes a split-top-p rounding defect, while the inspected fork kernel retains the older comparison. The executed baseline probe subsequently measured a failed transition at the requested boundary. Its active kernel was not traced, so source applicability does not establish the request's root cause or explain the earlier runs below the 49,152 boundary.

The pinned patch trial then completed six of eight forced-logit kernel checks and three of four ordinary exact-parity controls. The fourth ordinary comparison failed at GLM vocabulary 154,856 and batch 64, stopping startup before weights or candidate inference. The two remaining forced cases were not run. The model boundary failure is not proven fixed; retain the ordinary difference and obtain an independent mask/probability reference before revising the gate.

The later prospective upstream-tolerance trial passed its native-count kernel gate and executed one candidate request. The earlier token-0 symptom did not recur, but closing markers appeared at zero-based positions 64 and 126 and the answer contained duplicated prose. The strict boundary and answer gates failed. Fresh worker dispatch receipts do not establish per-token provenance. Exact operational restoration passed; matched strict JSON smokes failed before and after with no established quality regression. The full reasoning-limit gate remains unresolved; earlier failed trials are not reclassified.

Only after this gate passes should a separate coding comparison hold total output at 16K and compare an 8K reasoning budget against the original control, with the same task oracle, two-attempt policy and sampling. Changing effort is a separate experiment. No Ignis run or promotion is part of this packet.