Experiment record / persistent-agents
Memory saved calibration experiments, not inference
Both conditions answered all six known-fixture tasks correctly after an explicit recovery amendment. Persistent observation memory used three calibrations versus six, but required 37 inference dispatches versus 32. Canceled-call uncertainty and every earlier outcome remain preserved.
Question
Does persistent authentic observation memory reduce new calibration experiments, inference, both or neither across three first encounters and three apparatus revisits?
Hypothesis
At matched correctness, previous-task observations will reduce calibration acquisition on revisits. Fewer experiments without correct paired terminals do not count as a saving. Inference calls, tokens and elapsed time are separate efficiency measures.
Registration
The original sequence was committed locally before inference; it was not publicly preregistered. Recovery was explicitly authorized and committed locally before task6 continuation. A further committed source-preservation supplement used the exact registered snapshot after concurrent publication wording changed shared files. This is initial public reporting on 14 October 2026, not backdated public registration.
Methods
Six tasks used three fixed apparatuses in A,B,C,B,A,C order. Each had a stable hidden coefficient in a public two-particle dynamics family, two legal six-step calibration geometries and a physically distinct sealed target. Deterministic tools estimated full admissible parameter intervals from authentic rounded endpoints and integrated explicit public dynamics. The model selected operations, evidence IDs, arguments and terminal labels. Observations persisted automatically with provenance. Contexts were fresh between tasks; both controllers restarted after task3; condition execution order alternated. Controller-owned logical IDs were separate from transport IDs.
Baselines
PERSISTENT could search and retrieve prior-task authentic observations; EPISODIC_ONLY could access current-task observations only. Stores were separate and started empty. No prior-trial memories, previous-task transcript, evaluator answers or automatic scientific pipeline entered the model contexts. These are known corrected fixtures, not an independent generalization test.
Evaluation
Terminal categories were OCCURS, DOES_NOT_OCCUR or INSUFFICIENT_EVIDENCE for strict target separation reduction greater than 0.0001 after six steps. Utility was +1 correct, -1 wrong and 0 explicit abstention; incomplete and failed executions had separate counts. All twelve scheduled cells remain reported. Full-interval integration support, acquisition/reuse and matched-correct savings were checked separately.
Resource ceilings
Original fresh ceilings: 96 total inference calls, 800000 input tokens, 49152 generated tokens, 512 generated tokens and 300 seconds per request, 7200 seconds of actual model-loading/provider-request elapsed. Each task had eight decisions, with decision8 reserved for finish. Recovery allowed at most twelve additional calls, including auxiliary calls: four remaining decisions for EPISODIC_ONLY and eight for PERSISTENT. Preparation, reporting and deterministic audits were outside inference elapsed; loading counted once within request wall time.
Stopping conditions
The original run stopped when the chat interruption canceled call60 without a persisted response, leaving ten of twelve executions complete. Call60 remained spent with its full uncertain reservation. The authorized journal supersession retained usage uncertainty and blocked late delivery before persistence or execution. Recovery ran only task6, in EPISODIC_ONLY then PERSISTENT order. A source guard stopped after five settled recovery calls; an immutable source snapshot continued the remaining PERSISTENT slots without changing conditions or concurrent work. Both conditions finished before decision8. No transport retries, auxiliary inference or subsequent trial occurred.
Executing code
Original identity registration cb32e92; explicit recovery registration 894ee3b; frozen-source supplement 17792b4. Executing production digest e668e42492ade662b3b3f2f2f5754cca54be45c01b94ef1fcff6fb603d87ad64; frozen production anchor 894ee3bc3ca96f95dc0a96c97d0e2fe006d834c4. Private original code and journals remain retained; hashes are provenance, not a public source release.
Configuration
qwen3.8:27b-q4_K_M, digest 25b843619e944cd0ae6069f94ff4e5e26a16e109ccbc0a66a0f05979ed70098e; seed424242, temperature0.7, top_p0.8, top_k20, min_p0, context8192, output cap512, think=false, nonstreaming. Public simulator rules, six-decimal sensors, calibration choices and targets stayed fixed. Seeds do not guarantee identical stochastic outputs.
Complete outcomes
Scheduled twelve; completed twelve; correct twelve; wrong zero; abstentions zero; final incomplete/failure counts zero; utility +6 per condition. Each terminal had linked full-interval target integration support. First encounters used three calibrations in each condition; revisits used zero in PERSISTENT and three in EPISODIC_ONLY. Matched-correct saving: three experiments across six pairs. Persistent memory retrieved three relevant prior observations and reused them without new calibration on three tasks. Both conditions executed six searches, six estimates and six integrations; PERSISTENT additionally executed three retrievals. Rejected model decisions: seven persistent, one episodic; malformed scientific argument requests: six persistent, zero episodic; duplicate-ID rejections: zero. Rejections executed no operation. The historical canceled provider request remains separately counted; successful task recovery does not erase it.
Realized resources
PERSISTENT:37 dispatches,118128 known input tokens,5354 known generated tokens,2163.862 known request seconds. EPISODIC_ONLY:32 dispatches,91200 known input tokens,3930 known generated tokens,1713.866 known request seconds. Fresh aggregate:69 dispatches,209328 known input,9284 known generated,3877.728 known seconds; held call60 uncertainty15593 input/512 generated/300 seconds. Conservative fresh commitment:69 calls,224921 input,9796 generated,4177.728 seconds. Recovery added9 calls,32515 input,1403 generated and518.523 seconds; new held uncertainty zero. The older reservation19259 input/512 generated/300 seconds remains separately held. Historical plus fresh totals are in the derived JSON. Resource limits were not reset or enlarged.
Limitations
One known-fixture sequence, three apparatuses, fixed weights and deterministic scientific assistance. The observation-memory effect does not establish autonomous research, internal causal reliance, broad retrieval competence, learning or consciousness. Recovery and stochastic choices limit attribution to the ID repair. Unknown canceled-call tokens/time remain unsettled, so exact whole-run resource settlement is unavailable. Original interrupted and source-stopped analyses remain preserved separately; their cells were not silently removed. At-most-once invocation is supported across tested journal boundaries; exactly-once external side effects are not claimed.
Reproducibility
Twenty-one focused identity/recovery checks passed, including crash rollback, immutable supersession, explicit transport identity, late responses at multiple boundaries, recovery refusal after response/execution fences, caps, scientific validation and a provider-disabled task6 fixture continuation. Offline replay reconstructed all69 dispatch contexts, recomputed9 authentic calibrations and scientific facts, verified both original task3 restarts and checked60 accepted replays without effects. Late delivery was simulated offline; no actual call60 response was recovered. Tasks1–5 and original observations were preserved exactly. Derived JSON/CSV permit count and accounting checks. Private contexts, source, raw model transcripts and databases are withheld, preventing full public independent replay.
Revision record
14 October 2026, r1: initial public-safe report of a known-fixture comparison completed through explicit amended recovery. Earlier invalid, corrected, interrupted and source-stopped local evidence is retained. Selected artifacts are derived factual tables; CSV time values are rounded to six decimals. Raw conversations, model rationale, private source/journals, local paths and service logs are omitted. Publication invokes no model and does not reopen the experiment.
Reviewed evidence & downloads
- Derived outcomes, stages 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