Three responses, zero valid actions

Question

Can the installed 4B model return and commit a valid investigation proposal within a 30-second request deadline and 512 generated tokens?

Hypothesis

All three fresh, bounded requests will meet the action-delivery criterion: a valid propose_investigation decision, local action committed, and cycle 1. A timely but invalid response does not pass.

Registration

Local protocol frozen at 2026-10-13T20:15:48.934+00:00 before dispatch. Retrospective public backfill on 14 October 2026. It was not publicly preregistered. The supported think=true deviation was recorded before dispatch.

Methods

Three fresh runs shared seed 1301, the same empty simulated world, ordinary context assembly and a sourced operator task to propose an investigation. Only transport identifiers differed. No assistant history, hidden evaluator answers or prior-run memory was supplied. Ordinary parsing and semantic/access validation were used. The operator task marks these open calibration runs contaminated under apparatus policy.

Baselines

No comparative model or memory baseline was executed. Nonresident versus resident loading was observed, not randomized; cache and loading effects confound timing comparisons. This is a delivery gate, not an effect estimate.

Evaluation

Success required strict JSON parsing, valid decision/access checks, a committed propose_investigation action, and cycle 1. All three scheduled requests remain in the denominator. There were no forecast probes or Brier scores.

Resource ceilings

Shared ceilings: 3 calls, 18000 input tokens, 1536 generated tokens, 120 active seconds. Each fresh run: one call, 6000 input-token reservation, 512 generated-token cap, 30-second deadline. Repair and warmup calls were disallowed.

Stopping conditions

The three-call sequence ended and was closed against redispatch. Pending repairs were finalized by the one-call budget with provider/network access disabled. No subsequent experiment was launched for publication.

Executing code

Executing source SHA-256 3136c9f9d4f38ddef060af6dedac72e1fcd4d434bf7d26e111ebd129b0694988. Runner SHA-256 5c83b97504e4aba461d46efa6f603bd6dcc054cd0d2c71cca545a5503cef4ec0. Package 0.1.0. This source digest is not a Git commit or public source release; archived original code remains private.

Configuration

Ollama 0.40.1; qwen3:4b, 4.0B/Q4_K_M, digest 359d7dd4bcdab3d86b87d73ac27966f4dbb9f5efdfcc75d34a8764a09474fae7. Temperature 0; provider seed 424242; num_ctx 8192; JSON mode; stream=false; supported think=true; keep_alive 10m; generated-token cap 512. Seeds do not guarantee identical inference.

Complete outcomes

Scheduled 3; dispatched 3; returned before deadline 3; timeouts 0; invalid JSON/empty ordinary content 3; valid decisions 0; committed actions 0; repairs 0; unexecuted slots 0. All three ended budget_exhausted at cycle 0. Each response reported done=true and done_reason=length. Exposed thinking lengths were 2245, 2448 and 2449 characters; these are not token counts. Observed: the cap arrived without an ordinary decision. Inferred: this configuration fails the action-delivery criterion; larger caps or other models remain separate questions.

Realized resources

Reported aggregate usage: 3 calls, 3474 input tokens, 1536 generated tokens. Controller active time 38.478565 seconds; guarded HTTP time 38.468848 seconds; execution window including checks/exports 38.717 seconds. Input estimates were 5040/request, distinct from 6000-token reservations and provider counts. Outstanding uncertain input/output reservations 0/0. Cost was not measured; no paid or remote inference was used in this calibration.

Limitations

Only three observations, one model/configuration, no valid actions, no memory comparison, no forecast outcome, and uncontrolled hardware contention. Separate thinking-token counts are unavailable. Cache/loading effects cannot establish generation-speed improvement. Partial replay validates saved-state reconstruction only. This does not establish LLM learning, consciousness, functional compatibility outside this envelope, or eventual uncapped completion time.

Reproducibility

The downloadable derived summary and per-request numeric table permit independent accounting and count checks. Generation rate is eval_count divided by eval_duration/1e9 when available. Historical verification reported 22 events per run, zero effects and zero provider calls in offline partial reconstruction; 45 tests passed before dispatch. These checks were not rerun for publication. Original contexts, reasoning, controller code and private database are withheld, so complete independent replay is unavailable. Running new inference requires separate research authorization, not this publishing command.

Revision record

14 October 2026, r1: initial public backfill from frozen local evidence. Private transcripts, reasoning, local paths, internal service logs and source archives omitted. Numeric artifacts are explicitly derived; original research evidence is unchanged.

Every scheduled request

Recorded method references

• Recorded API reference: Ollama chat
• Recorded thinking-control reference

Reviewed evidence & downloads

• Protocol, outcomes and accounting (JSON) — derived. Selected project-authored factual data; redistribution permitted.
• All three per-request outcomes (CSV) — derived. Selected project-authored factual data; redistribution permitted.

These are selected public derivatives; raw transcripts, internal configuration and private source are withheld. Hashes describe the actual downloads, not unmodified original evidence.

Artifact manifest and source provenance · Permanent revision URL

Substantive versions

• Revision 1