Persistent observations and a minimal deterministic controller

Question

How much calibration saving is explained by ordinary persistent observations and deterministic scientific execution, and what does the LLM add?

Hypothesis

At matched correctness, stored authentic observations can reduce revisit calibration. A deterministic procedural controller may achieve the same task without model inference. These predictions are tested, not assigned as new-fixture outcomes.

Registration

Local commit 565e8472c5275e4561efd3cffa78c2fd9fe21915; public r1 verified at 2026-10-14T20:00:38+00:00, before first model dispatch. Actual receipt retained; original public r1 unchanged.

Methods

Two independent blocks, three fresh apparatuses and six tasks A,B,C,A,C,B per block. Exactly the same fixtures across four conditions. Fresh stores per block/condition; fresh task contexts; all controllers restart after task three. Cyclic condition rotations balance each condition across serial positions three times. Same frozen scientific runtime and model configuration as earlier memory-new-fixtures-v1. Full generator, filters, memory restrictions and recovery are in the registration artifact. Neutral identifiers assigned independently of dynamics. Constructed outcome-changing revisits and sufficient-calibration guarantee restrict the tested distribution. All 23 rejected candidates and reasons retained; 7 in block one, 16 in block two. No model calls for fixture selection.

Baselines

LLM_PERSISTENT/LLM_EPISODIC choose actions with the same model. RULE_PERSISTENT/RULE_EPISODIC implement the documented schedule with zero inference. Persistent memory contains authentic observations only; episodic access is current-task only. Exact apparatus identity matching; estimates recomputed every task. Constant labels and historical-label-copying references included.

Evaluation

48 scheduled task executions, 12 per condition, 24 involving the LLM. Main unit six apparatuses in two blocks; dependent revisits reported separately. +1 correct, -1 incorrect, 0 explicit abstention; missing terminals and failures separate. Report calibrations, selected sources, scientific operations, retrievals, rejected requests and resources for every task. Compare matched-correct persistent/episodic and rule/model performance. No precise population estimate.

Resource ceilings

192 total model requests including auxiliary; 1600000 full logical input tokens including cached input; 98304 generated; 512 generated and 300 seconds per request; 14400 seconds actual loading/provider-request wall time. Eight decisions per model task, eighth finish only. Zero auxiliary calls, warmup, transport retries, downloads or paid inference. Rule/offline runs zero inference. Loading nested in request elapsed, counted once. Preparation/reporting outside inference budget.

Stopping conditions

Execution completed. Stop reason: None. No outcome-selected retries or fixture/prompt changes. Historical uncertain reservations retained separately. All 48 scheduled executions reached terminals; post-execution audit harness repairs did not reopen execution or consume model resources.

Executing code

Isolated worktree and new frozen runtime directory; production digest e668e42492ade662b3b3f2f2f5754cca54be45c01b94ef1fcff6fb603d87ad64. Same scientific operations, numerical semantics, sensor precision, controller identities, final-slot behavior and late-response fences. Sealed hashes included; private source withheld.

Configuration

qwen3.8:27b-q4_K_M, digest 25b843619e944cd0ae6069f94ff4e5e26a16e109ccbc0a66a0f05979ed70098e; seed 424242, temperature 0.7, top_p 0.8, top_k 20, min_p 0, context 8192, cap 512, think=false. Six-decimal sensors, full inclusive parameter intervals; calibration geometries x=0.25 and 0.3, six steps, strict target separation reduction >0.0001. Rule baseline uses the same public APIs and observation provenance.

Complete outcomes

{"LLM_PERSISTENT": {"correct": 12, "calibrations": 9, "requests": 74, "input_tokens": 245498, "generated_tokens": 9929, "provider_seconds": 4373.36456646584, "total_task_seconds": 4383.728188461624, "rejections": 12, "abstentions": 0, "failures": 0, "incomplete": 0}, "LLM_EPISODIC": {"correct": 12, "calibrations": 13, "requests": 63, "input_tokens": 188638, "generated_tokens": 8080, "provider_seconds": 3449.210198873654, "total_task_seconds": 3458.0717049548402, "rejections": 2, "abstentions": 0, "failures": 0, "incomplete": 0}, "RULE_PERSISTENT": {"correct": 12, "calibrations": 6, "requests": 0, "input_tokens": 0, "generated_tokens": 0, "provider_seconds": 0, "total_task_seconds": 1.9543900806456804, "rejections": 0, "abstentions": 0, "failures": 0, "incomplete": 0}, "RULE_EPISODIC": {"correct": 12, "calibrations": 12, "requests": 0, "input_tokens": 0, "generated_tokens": 0, "provider_seconds": 0, "total_task_seconds": 1.2024850007146597, "rejections": 0, "abstentions": 0, "failures": 0, "incomplete": 0}} Apparatus-level differences and first/revisit outcomes are in the summary; all 48 rows are retained. Rule matching or exceeding the model supports a simpler implementation for this task. No measured task capability is assigned to the LLM without an observed advantage. Persistent observations explain revisit calibration avoidance in both controller families. Rule persistence saved one calibration on every apparatus, three in each block. LLM savings per apparatus were 1,1,0 and 1,1,0, two per block. LLM first encounters used 9 versus 6 calibrations; revisits 0 versus 7. Rules used 6 versus 6 on first encounters and 0 versus 6 on revisits. At matched correctness the rule controllers matched labels, used fewer calibrations and required no inference. No measured capability in this task required the LLM. This establishes a simpler implementation here, not absence of LLM value elsewhere. Evidence qualification: all 24 rule terminals had exact full-estimate-interval target integration. All 23 committed LLM target integrations changed at least one printed endpoint: 21 were equivalent under the historical eight-ULP trace-copy relation and two differed beyond it; only 3 enclosed the full original interval exactly. One correct LLM_PERSISTENT terminal (block 1 task 6) had an estimate but no committed target integration after four rejections. Literal full-interval preservation was therefore not achieved by the model submissions. The registered label-correctness scores remain unchanged; correct labels are not treated as proof of complete scientific justification.

Realized resources

{"LLM_PERSISTENT": {"actual": {"calls": 74, "inference_seconds": 4373.36456646584, "input_tokens": 245498, "output_tokens": 9929}, "outstanding_reservations": {"calls": 0, "inference_seconds": 0, "input_tokens": 0, "output_tokens": 0}, "committed": {"calls": 74, "inference_seconds": 4373.36456646584, "input_tokens": 245498, "output_tokens": 9929}, "timing_seconds": {"scientific_tool_seconds": 0.06345437653362751, "calibration_execution_seconds": 0.007761374115943909, "storage_retrieval_seconds": 0.9846459897235036, "other_processing_seconds": 0.8010870544239879, "guard_seconds": 8.508355705998838, "provider_request_seconds": 4373.359261953272}, "loading_seconds": 8.740549912999999, "total_task_seconds": 4383.728188461624}, "LLM_EPISODIC": {"actual": {"calls": 63, "inference_seconds": 3449.210198873654, "input_tokens": 188638, "output_tokens": 8080}, "outstanding_reservations": {"calls": 0, "inference_seconds": 0, "input_tokens": 0, "output_tokens": 0}, "committed": {"calls": 63, "inference_seconds": 3449.210198873654, "input_tokens": 188638, "output_tokens": 8080}, "timing_seconds": {"scientific_tool_seconds": 0.047631832771003246, "calibration_execution_seconds": 0.015675289556384087, "storage_retrieval_seconds": 0.9015561770647764, "other_processing_seconds": 0.7177687212824821, "guard_seconds": 7.177610851824284, "provider_request_seconds": 3449.204110168852}, "loading_seconds": 0.475569666, "total_task_seconds": 3458.0717049548402}, "RULE_PERSISTENT": {"actual": {"calls": 0, "inference_seconds": 0, "input_tokens": 0, "output_tokens": 0}, "outstanding_reservations": {"calls": 0, "inference_seconds": 0, "input_tokens": 0, "output_tokens": 0}, "committed": {"calls": 0, "inference_seconds": 0, "input_tokens": 0, "output_tokens": 0}, "timing_seconds": {"scientific_tool_seconds": 0.03200953733175993, "calibration_execution_seconds": 0.004765873774886131, "storage_retrieval_seconds": 0.4191576289013028, "other_processing_seconds": 0.06125504896044731, "guard_seconds": 1.4327056193724275, "provider_request_seconds": 0}, "loading_seconds": 0, "total_task_seconds": 1.9543900806456804}, "RULE_EPISODIC": {"actual": {"calls": 0, "inference_seconds": 0, "input_tokens": 0, "output_tokens": 0}, "outstanding_reservations": {"calls": 0, "inference_seconds": 0, "input_tokens": 0, "output_tokens": 0}, "committed": {"calls": 0, "inference_seconds": 0, "input_tokens": 0, "output_tokens": 0}, "timing_seconds": {"scientific_tool_seconds": 0.028372046537697315, "calibration_execution_seconds": 0.008646498434245586, "storage_retrieval_seconds": 0.2919577034190297, "other_processing_seconds": 0.043726081028580666, "guard_seconds": 0.828641340136528, "provider_request_seconds": 0}, "loading_seconds": 0, "total_task_seconds": 1.2024850007146597}} Loading is nested in provider wall and is not added again. Exclusive timing categories and total wall are distinct. Hypothetical added calibration-cost surface in seconds/calibration is supplied, not a claim of energy, monetary cost or FLOPs. LLM persistence saved four calibrations but added 11 requests and 924.154 provider seconds (15.4 minutes), or 925.656 total task seconds. Rules saved six calibrations but took 1.954 versus 1.202 total task seconds, including guards. Hypothetical additional calibration-cost break-even at measured provider speed: about 231.414 seconds per calibration for LLM persistence, and 0.125318 seconds per calibration for rule persistence. These are additional assumed wall costs, not observed physical experiment costs. The two-axis cost surface varies added calibration seconds and a hypothetical multiplier on measured provider elapsed; it assumes unchanged behavior and predicts no model trajectory.

Limitations

Six apparatuses, two selected simulator blocks, stable simple dynamics, exact source identity matching, engineered changed outcomes and each legal calibration sufficient. Results are conditional on this constructed distribution. Matching rules show this task admits a simpler implementation; they do not establish absence of LLM value elsewhere. An LLM advantage requires faithful baseline and equivalent information. No consciousness inference. Token counts imply no FLOPs, energy, monetary cost or provider determinism. Execution order balances serial positions, but pairwise precedence is not fully counterbalanced: LLM_PERSISTENT preceded LLM_EPISODIC on 9 of 12 targets. Provider cache, load and timing variation remain possible confounds for time comparisons. Model interval-copy deviations and one unsupported terminal limit scientific-procedure compliance claims despite perfect label accuracy.

Reproducibility

{"verified_at": "2026-10-14T22:13:59.085+00:00", "provider_disabled": true, "reconstructed_llm_requests": 137, "accepted_replays_inert": 231, "scientific_facts_recomputed": 95, "scheduled_rows": 48, "restarts": 8, "baseline_policy_public_context_only": true, "source_and_evidence_seals": true, "accounting": true, "rule_provider_requests": 0} Request contexts reconstructed from pre-dispatch event-prefix states. Rule response procedure recomputed from identical public context. Scientific facts and categorical scoring independently recomputed; accepted actions replay without effects. Public factual derivatives support count checks; withheld source, hidden fixtures, journals and full traces prevent full public replay. Automatic closeout verification initially failed on a legacy old-order audit helper; a second diagnostic exposed an overbroad denial of public sensor category metadata. A separate provider-disabled reporting verifier used the already registered task-order adapter and allowed only the authentic public categories metadata. It verified archived event hash chains, reconstructed contexts and all rule responses, and replayed accepted operations without effects. No executing source, fixtures, prompts, scoring or inference changed. The original errors and corrected audit source are retained privately.

Revision record

2026-10-14, r2: all 48 scheduled outcomes and findings. Public r1 registration preserved. Future proposal: change the apparatus dynamics between visits under a registered change-point rule to test when observation reuse becomes stale; do not launch that study.

Reviewed evidence & downloads

• actions — derived. Selected project-authored factual data and policy description; redistribution permitted.
• break even surface — derived. Selected project-authored factual data and policy description; redistribution permitted.
• registration — derived. Selected project-authored factual data and policy description; redistribution permitted.
• rule policy — derived. Selected project-authored factual data and policy description; redistribution permitted.
• summary — derived. Selected project-authored factual data and policy description; redistribution permitted.
• tasks — derived. Selected project-authored factual data and policy description; redistribution permitted.
• tool contracts — derived. Selected project-authored factual data and policy description; 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 2