New-fixture replication of persistent observation memory

Question

Does persistent memory again reduce calibration experiments while maintaining correct decisions on previously unused apparatuses and target geometries?

Hypothesis

At matched correctness, authentic prior-task calibration observations reduce new calibrations on revisits. Calls, tokens and time are separate costs. Incorrect answers, premature abstentions and missing terminals cannot establish useful efficiency.

Registration

One trial committed locally as 5df9edff1f6c683af35a6e3b230602d982036c1c before inference. Fixtures, seed, selection filters, scoring, order, recovery policy, budgets and runtime hashes were frozen. This was local preregistration; this page is initial public availability, not backdated public registration.

Methods

Three new apparatuses with stable hidden dynamics, six new target geometries, sequence A,B,C,A,C,B. Every revisit changes outcome. Reproducible generator seed 2026101401 samples simulator seeds uniformly from integers 1000000 through 2147483646; the unchanged simulator samples signed coefficient magnitude uniformly from 0.06 to 0.18. Target x is sampled on a millionth-unit grid: first encounters 10000 through 240000; revisit band selected with equal probability, close 498000 through 499980 or wall 50 through 1950. First passing candidate per apparatus, cap 1000; all 15 rejected candidates and reasons retained privately. Filters require unused target geometry, calibration-target separation, sufficient calibration information, a full-interval threshold margin at least max(0.00001, ten enclosure widths), both outcomes possible under the public coefficient domain, and changed revisit outcomes. Both legal calibrations suffice, stronger than the minimum one. No model selection calls. Neutral apparatus metadata discloses no hidden parameters or outcomes. The frozen model chooses search, retrieval, calibration, source IDs, estimation, explicit integration arguments and terminal labels; deterministic assistance computes scientific operations. Fresh contexts per task, condition order alternated, both controllers restarted after task three.

Baselines

Separate initially empty memories: PERSISTENT may search and retrieve prior authentic observations; EPISODIC_ONLY accesses current-task evidence only. Constant OCCURS and DOES_NOT_OCCUR each score 3/6 correct, 3 wrong, utility 0 on the balanced six targets. Historical-label copying scores 0/3 correct on revisits, utility -3; undefined on first encounters. First-encounter labels are not supplied to the model as historical memory.

Evaluation

All twelve scheduled task executions are reported in the CSV, including explicit errors, abstentions, incomplete tasks and execution failures. Strict target separation reduction >0.0001 after six simulation steps. Utility +1 correct, -1 incorrect, 0 explicit abstention; absent terminals are separately counted. Experiment savings use matched-correct pairs only. Authentic reuse requires prior environment observations linked through model-selected retrieval and parameter estimation. Full estimated interval and registered target integration support are audited. Three apparatuses are the main experimental units; revisits are dependent.

Resource ceilings

Twelve task executions, eight decisions each including rejected requests. Eighth decision reserved for finish; no task replacements or outcome-selected retries. 96 total inference calls including auxiliary; 800000 full logical input tokens including cached tokens; 512 generated tokens per request and 49152 total; 300 seconds per request; 7200 actual model-loading/provider-request seconds. Preparation/reporting excluded; loading counted once. Durable conservative reservations precede dispatch. Unknown usage remains reserved; canceled dispatches stay charged. No auxiliary inference or transport retries.

Stopping conditions

Execution status completed. All scheduled tasks reached terminal outcomes within the allowance. Actual interruption recoveries: 0. The committed recovery policy permits same-allowance checkpoint recovery only with journal proof of at-most-once accepted execution and immutable late-response fencing. Uncertain execution fences stop. Historical usage uncertainty remains held. No subsequent experiment was launched.

Executing code

Successful package source anchor 894ee3bc3ca96f95dc0a96c97d0e2fe006d834c4; production digest e668e42492ade662b3b3f2f2f5754cca54be45c01b94ef1fcff6fb603d87ad64. Exact original source hashes were verified, and imports used an isolated experiment directory in a separate worktree. Concurrent shared-checkout changes cannot change those imports. A fixture-validation adapter changed only required order to A,B,C,A,C,B. Model prompts, schemas, controller-assigned identities, scientific tools, numerical semantics, malformed-argument behavior, memory policies and decision limits were fixed. Private source is 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, generated cap 512, think=false, nonstreaming. Same two calibration geometries x=0.25 and x=0.3, exact registered symmetric initial positions, six-decimal undistorted sensors and full inclusive parameter intervals. Model identity and provider idleness checked before each dispatch.

Complete outcomes

Both conditions: 6/6 correct, utility +6; zero incorrect decisions, abstentions, incomplete tasks or execution failures. PERSISTENT used five calibrations versus seven: two matched-correct experiments saved across all six scheduled pairs. First encounters used five versus three; revisits zero versus four. Five authentic prior calibration observations were retrieved and used across three revisits without new calibration. PERSISTENT had eight rejected requests, all missing scientific argument fields; EPISODIC_ONLY had four, comprising three missing-field requests and one flow/action mismatch. Rejections executed no scientific operation. Each condition finished five tasks before decision eight and one at decision eight. All twelve terminals had evidence-linked full-interval target integration support. Apparatus-level totals: A used two versus three calibrations, B one versus two, C two versus two. Experiment savings replicated within this small registered sample. The earlier known-fixture amended recovery separately obtained 6/6 correct per condition and three savings; it is not pooled with this prospective sequence.

Realized resources

PERSISTENT 40 calls, 141615 input tokens, 5487 generated tokens, 2174.238 request seconds; EPISODIC_ONLY 35 calls, 105752 input tokens, 4755 generated tokens, 1724.672 request seconds. Loading seconds, already included in request elapsed: {"PERSISTENT": 11.616148126, "EPISODIC_ONLY": 0.281446706}. New-trial actual, outstanding and conservative committed usage are provided separately from unchanged historical accounting in the derived JSON. Historical reservations remain 34852 input tokens, 1024 generated tokens and 600 seconds; these are not released by the new result. Persistent overhead was five calls, 35863 input tokens, 732 generated tokens and 449.566 provider-request seconds (about 7.5 minutes). New trial totals: 75 calls, 247367 input tokens, 10242 generated tokens and 3898.910 provider-request seconds; no new outstanding reservations. All costs include rejected requests.

Limitations

Only three apparatuses in one selected sequence, narrow deterministic dynamics, fixed model and deterministic scientific assistance. Revisit observations are dependent. Offline outcome balancing, identifiability and precision filters restrict the sampled population; metadata alone can admit both outcomes, but no claim is made that all broader fixtures are feasible. Seeded stochastic inference does not imply deterministic responses. Source-linked reuse is an auditable behavioral measure, not proof of internal causal reliance. This small replication does not establish broad generalization or autonomous research. Private evidence is withheld, limiting full public replay.

Reproducibility

25 focused offline tests passed before dispatch, covering generator reproducibility, runtime hashes, boundaries, isolation, budgets, restart, at-most-once replay, crash recovery and immutable late-response fences. Provider-disabled replay reconstructed 75 requests and recomputed 12 authentic calibrations and scientific facts. Both task-three restarts, independent categorical scoring, resource settlement and request/response/action associations were verified. Public factual derivatives support count checks; full private prompts, responses, fixtures, journals and source are not released. All 63 accepted decisions replayed without invoking an operation again. An independent provider-disabled audit matched all 75 prompt and generated token counts to retained backend logs. The result evidence was sealed after process exit; the initial pre-flush log seal and its finalization note are retained privately.

Revision record

2026-10-14, r1: initial public-safe report of one prospectively registered replication on new fixtures. Derived JSON and CSV omit private model text, hidden seeds, trajectories, source, journals, service logs and local infrastructure. CSV elapsed values rounded to six decimals. Publication uses static selected outputs and no inference. Proposed next step: a separately authorized, larger preregistered apparatus sample to estimate how often calibration savings survive inference overhead.

Reviewed evidence & downloads

• Derived counts, baselines and accounting — derived. Selected project-authored factual data; redistribution permitted.
• All twelve scheduled task outcomes — derived. Selected project-authored numerical results; 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