Experiment record / persistent-agents
Coherent continuity contract: corrected repeat
Prospective corrected repeat on the same known fixtures, with coherent continuity descriptions and complete-context validation; 64 scheduled executions, 16 using the 27B model.
Question
When should a controller reuse an observation, and when must it acquire fresh evidence because the apparatus may have changed?
Hypothesis
Under explicit continuity contracts, a change-aware deterministic rule may suffice. Changing the environment alone does not demonstrate an LLM advantage. Primary comparison is LLM_PERSISTENT against RULE_CHANGE_AWARE.
Registration
Corrected repeat of scientifically invalidated observation-validity-v1 revision 2, preserved at https://chronomata.ai/experiments/chronomata--observation-validity-v1/revisions/r2/. New local preregistration commit 51a74ff949e2dfc81f68614e9d810c668b4005b4. This public r1 is verified before inference. No prior model responses, tool facts, feedback or memories enter new stores.
Methods
Exactly the predecessor's eight apparatuses, hidden schedules, sixteen targets, calibrations, four policies and condition ordering. No new fixture generation or filtering. Two separately seeded blocks and constructed diagnostic strata. A,B,C,D,B,D,A,C per block; restart after four first encounters. Corrected repeat on known fixtures, not independent generalization. Only continuity descriptions, directly necessary validation and new-run bookkeeping change.
Baselines
All four policies use identical public APIs, observations, targets, precision, interval semantics and memory restrictions. Policies receive only public context, never the backend, coefficient, hidden schedule, evaluator truth, case-specific answers or privileged estimate cache. Recompute estimates each task. Fresh stores per condition/block, fresh model context per task, no earlier observations or conversations. LLM_PERSISTENT: verified installed 27B selects memory actions, calibration, estimation source IDs, integration references and final category/citation itself. Both contracts are supplied in every task. Only selected authentic observations carry cross-task evidence. RULE_CHANGE_AWARE: use current-task observations; otherwise search exact apparatus metadata only in certified mode, retrieve lexically sorted observations whose certified version equals the current token. In uncertified mode acquire current-task evidence. If none usable, execute the first legal calibration in lexical experiment-ID order. Estimate from all selected applicable observations, integrate exact target by reference, and finish from explicit certified event status, citing that integration. Unresolved estimate/integration: acquire the next untried lexical calibration only if at least four decisions remain for calibration/estimation/integration/finish; otherwise abstain. No hidden access or inference. RULE_ALWAYS_FRESH: acquire current-task calibration and use the same estimation/integration/finish procedure, without relying on previous-task evidence. RULE_BLIND_CACHE: search/retrieve by apparatus identity and ignore continuity. Calibrate only when no identity-matched observation exists, then use the same tools. Intentionally unsafe. It is not the competent baseline for an LLM-advantage claim. The common predeclared unresolved-tool rule applies. The four Williams rows (LLM, CHANGE, BLIND, FRESH), (CHANGE, FRESH, LLM, BLIND), (FRESH, BLIND, CHANGE, LLM), (BLIND, LLM, FRESH, CHANGE) repeat four times across sixteen targets. Every policy occupies every position four times; each policy precedes each other policy on 8/16 targets. Dispatch serially. Restart all controllers after task four in each block, preserving exact checkpoint and allowed observation IDs.
Evaluation
All 64 scheduled rows; sixteen per policy, only sixteen using LLM. Eight apparatuses are experimental units. Label accuracy and +1/-1/0 label utility separate from supported-correct classification, which requires cited committed exact-target integration of the full unaltered interval and publicly applicable source provenance. Track errors, abstentions, incomplete tasks, unsupported answers, missing/invalid citations, reuse validity, fresh/redundant calibration, calls/rejections and costs. Private unchanged truth does not certify stale evidence. No precise population claims.
Resource ceilings
128 inference calls including auxiliary; 1100000 full logical input tokens, 65536 generated; 512 generated and 300 seconds/request; 10800 seconds actual loading/provider-request elapsed. Eight model decisions/task, eighth finish only. Loading nested, counted once. No auxiliary inference, downloads, paid/cloud calls or interference. Rules/offline verification use zero inference. Preparation/reporting outside inference budget.
Stopping conditions
Stop entire partial study on unknown usage or uncertain operation. Canceled dispatch spent, unknown reservations retained; immutable late-response tombstone before any response/execution fence. No redispatch, adaptive prompts, outcome-selected retries or replacement fixtures. Registered checkpoint restart after task four. Accepted operations replay inertly.
Executing code
Isolated sealed runtime 68ea974ee935c5a31e8ed61505df1ef751adc6235cfef8780dde42f60c959658. Numerical components byte-identical to the predecessor. Source and model-visible input diffs retained privately, with reviewed descriptive changes and component hashes published. Policy/backend/scoring components verified structurally unchanged.
Configuration
Verified installed qwen3.8:27b-q4_K_M, digest 25b843619e944cd0ae6069f94ff4e5e26a16e109ccbc0a66a0f05979ed70098e; seed 424242, temperature .7, top_p .8, top_k 20, min_p 0, context 8192, cap 512, think=false. Six-decimal observations, six steps; calibration x=.25/.3 distinct from sealed targets. Target strict separation reduction >.0001. Both policies receive identical public contracts and API capabilities.
Complete outcomes
Not executed at public registration. Forty-three focused offline tests passed, including exact legacy-claim injection before dispatch. Ninety-eight complete scripted requests and corrected forensic views of all 85 predecessor requests validated without inference. Three rule policies completed simulator-only feasibility: change-aware 16/16 supported-correct with 14 calibrations, fresh 16/16 with 16, blind 12/16 correct labels and 10/16 supported-correct with 8. These are offline checks, not model outcomes.
Realized resources
Zero new inference at registration. Invalidated predecessor: 85 calls, 292363 input tokens, 10043 generated tokens, 4321.216771 provider seconds. Corrected consumption will be reported separately and with an explicit combined cost total. Earlier uncertainty retained separately: 34852 input tokens, 1024 generated tokens, 600 seconds; nothing released.
Limitations
Known-fixture corrected repeat, eight constructed apparatuses, one model and stochastic decoding. Changed model answers cannot by themselves establish that wording caused a difference. Equal policy exposure alone is not proof of coherence. Public continuity determines source applicability; privately unchanged truth supplies no certificate. Label accuracy is not evidence of sound source use. Private source, hidden schedules, real model transcripts and logs remain withheld, limiting independent reproduction. No population, energy, monetary or consciousness inference.
Reproducibility
Canonical field descriptions, explicit schemas, immutable observation provenance, full-interval/exact-target recomputation, complete request reconstruction and accepted-operation replay checks. Scripted provider-disabled request artifacts are labeled and include system, user and response schema. They are not actual model conversations. Every actual archived request will be checked against the same frozen validator after execution.
Revision record
2026-10-14 r1: prospective public registration of one corrected repeat. The predecessor remains invalidated. No subsequent task is launched.
Recorded method references
Reviewed evidence & downloads
- consistency table — derived. Selected project-authored contracts, factual records and provider-disabled scripted examples; redistribution permitted. No real model conversation, hidden schedule or private source included.
- consistency table — derived. Selected project-authored contracts, factual records and provider-disabled scripted examples; redistribution permitted. No real model conversation, hidden schedule or private source included.
- contracts — derived. Selected project-authored contracts, factual records and provider-disabled scripted examples; redistribution permitted. No real model conversation, hidden schedule or private source included.
- input changes — derived. Selected project-authored contracts, factual records and provider-disabled scripted examples; redistribution permitted. No real model conversation, hidden schedule or private source included.
- policies — derived. Selected project-authored contracts, factual records and provider-disabled scripted examples; redistribution permitted. No real model conversation, hidden schedule or private source included.
- registration — derived. Selected project-authored contracts, factual records and provider-disabled scripted examples; redistribution permitted. No real model conversation, hidden schedule or private source included.
- scripted complete requests — derived. Selected project-authored contracts, factual records and provider-disabled scripted examples; redistribution permitted. No real model conversation, hidden schedule or private source included.
- source compatibility — derived. Selected project-authored contracts, factual records and provider-disabled scripted examples; redistribution permitted. No real model conversation, hidden schedule or private source included.
- summary — derived. Selected project-authored contracts, factual records and provider-disabled scripted examples; redistribution permitted. No real model conversation, hidden schedule or private source included.
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