# Observation validity under changing dynamics — observation-validity-v1

Registered 2026-10-14, before inference. Continue the completed memory-mechanism-v1 revision 2 without changing its scores, scientific-compliance qualifications, evidence or held reservations. Prior work found no measured LLM requirement under stable exact identity. Environmental change alone is not an LLM advantage. One new study; eight apparatuses in two blocks, sixteen tasks per policy, 64 scheduled task executions, only sixteen with the LLM. Main units are eight apparatuses; revisits are dependent. Descriptive constructed diagnostic strata, not a natural drift-rate sample or precise population estimate.

Question: when should a controller reuse an observation, and when must it acquire fresh evidence because the apparatus may have changed? Hypothesis: a competent change-aware deterministic controller suffices under the declared continuity contracts. Compare LLM primarily to that policy; blind caching is an unsafe diagnostic reference.

## Interface and scientific semantics

Before live registration, offline checks verify an additive reference interface. Estimation accepts authentic observation source IDs plus the unchanged dynamics assumption ID. It persists the exact full inclusive interval, assumptions, units, sensor uncertainty and source-payload hashes as an immutable fact, returning its fact ID. Integration accepts only parameter_fact_id and the current public target_id. The tool loads the original interval and exact registered target internally; it uses the unchanged estimator and full-range numerical integrator. No bounds or coordinates are silently repaired. Old numeric argument submissions are rejected. Facts are task/run-owned, immutable and not carried across tasks. Only authentic observations persist, retrieved through ordinary search/retrieval. This is an interface change, not improved model arithmetic.

A terminal chooses its own category and should cite integration_fact_id. Missing, nonexistent, foreign, wrong-type, wrong-task or otherwise invalid citations are retained and scored separately. Tool status never replaces a terminal label. The eighth decision is finish only; invalid requests consume a slot. No automatic scientific pipeline runs for the LLM. Calibration and target cannot physically alias.

## Observable contracts

CERTIFIED_VERSION: public apparatus metadata includes an opaque version token. Every hidden-parameter change updates it; dynamics are stable within a version. It reveals no parameter or answer. Both controllers see exactly the same token. Exact apparatus and version equality licenses reuse.

NO_CONTINUITY_CERTIFICATE: no public version token. Parameters may change between visits. Old evidence does not certify the current coefficient, including when private history happens to be unchanged. Acquire current-task evidence or explicitly abstain. No controller is asked to infer an unannounced change from identical information.

Both modes have fixed dynamics throughout every task, calibration and sealed target. All changes occur at the single registered block boundary after the four first encounters; all controllers then restart. No within-task change occurs. Future task metadata is not supplied to a controller.

## Fixtures and retained sampling

Two blocks, generator seeds 2026101404 and 2026101405. Each has four new neutral apparatus IDs independently assigned to the four contract/change strata, with one unchanged and one changed under each contract. Order A,B,C,D,B,D,A,C, new target geometry on revisits. All first encounters precede changes/recovery. Assignment is hidden except legitimate certified tokens. IDs and other metadata do not encode assignments. Opaque tokens are sampled independently before candidate filtering. Public target IDs identify only block and task. Full private schedule, seeds, parameters and rejection records are sealed before inference; public provenance hashes do not expose change masks.

Simulator-only first-passing candidate selection, maximum 1000 candidates per apparatus. Integer world seeds uniform 1000000..2147483646; simulator's original signed coefficient magnitude uniform .06.. .18. Changed cases independently sample a second world seed, stable cases reuse the first. New-world seeds and all non-target dynamics remain private. Symmetric A/B particles, y=.5, zero velocity, six steps. First x integer 10000..240000 divided by one million; revisit band chosen with probability .5, close integer 498000..499980 or wall 50..1950 divided by one million. Target event is strict separation reduction >.0001. Calibration geometries .25 and .3, six steps, six-decimal observations. No model-based fixture selection.

Retain every candidate and rejection reason. Filters: no earlier or within-study target geometry or world-seed duplication; calibration/target separation; each offered fresh calibration certifies the same full-interval target label; threshold margin >= max(1e-5, ten enclosure widths); public prior admits both labels at every target, with retained witness coefficients. Changed cases require the complete old calibration-derived interval to certify the wrong new-target label. Stable cases require target-geometry change to flip the target label. Independently cross-check scalar simulator and Decimal endpoint arithmetic against numerical enclosures. Unexpected duplicates, impossible witnesses or cap exhaustion halt preparation. These filters construct diagnostic strata; no natural prevalence is inferred.

## Policies and equal capabilities

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.

## Scoring and costs

Report all 64 scheduled executions, including errors, explicit abstentions, unsupported answers, invalid operations, missing terminals and stopped/unexecuted tasks. Ordinary label accuracy is separate from supported-correct classification. Supported-correct requires correct chosen label, a committed cited current-task integration of the exact target using the complete unaltered evidence interval, and original observations applicable under the public contract or current-task acquisition. Private unchanged truth never licenses freshness. No scientific-support imputation from a correct label.

For every policy and contract/change stratum report correct/incorrect labels, abstentions, unsupported answers, incomplete tasks, missing/invalid citations, appropriate/unjustified reuse, fresh/redundant calibration, selected sources, parameter and target provenance, all calls/rejections and costs. Reuse counts distinct previous-task observations selected for estimation, classified by observable applicability. A calibration is redundant if an earlier authentic observation already certifies applicability to that task, including earlier current-task evidence. Also report task-level indicators, first/revisit splits, and apparatus-level descriptive variation. +1 correct, -1 incorrect, 0 explicit abstention; missing utility undefined, utility convention separate from evidence compliance. No reward to unjustified answers as evidence of scientific reasoning.

Measure actual provider requests, full logical input/generated tokens, request wall time including model loading, scientific-tool execution, calibration simulator execution, storage/journal/retrieval, other processing, guards and total task elapsed. Timing scopes are exclusive; loading is nested within provider time, never added twice. Rule/offline compute zero inference and recorded separately. Time/token counts imply no energy, FLOPs, money or deterministic provider execution. Preparation/reporting outside inference budget.

## Frozen runtime, allowance, recovery

Installed qwen3.8:27b-q4_K_M, digest 25b843619e944cd0ae6069f94ff4e5e26a16e109ccbc0a66a0f05979ed70098e. Same seed 424242, temperature .7, top_p .8, top_k 20, min_p 0, context 8192, generated cap 512, think=false. No downloads, cloud/paid calls, warmup or auxiliary inference. Authorized local hardware only; check identity and idle status before each dispatch, do not interfere with unrelated work.

Eight model decisions per task, eighth reserved for finish. Absolute maxima: 128 inference calls including auxiliary; 1100000 full logical input tokens, 65536 generated tokens, 512 generated and 300 seconds/request, 10800 seconds actual loading/provider-request elapsed. Durable conservative UTF-8 input bound plus output/time maximum reserved before dispatch; known usage settles only when verified. Canceled dispatches remain spent, unknown dimensions fully reserved. Prior held 34852 input, 1024 output and 600 seconds remain unchanged and separate.

Isolated sealed runtime copy and runner, immutable source/contract/input/schedule/ordering/test hashes before inference. Other project and publication edits cannot change it. Locally commit registration and publish safe r1 through the approved shared publisher. Require verified public receipt before first dispatch; no backdating. Private repository/history/raw transcripts never published or pushed.

Exclusive process lock, SQLite FULL/WAL journal. Accepted-operation and identical response replays are inert; different replay rejected. Persist selected transport before parsing and execution fence before any operation. A persisted returned response may recover only through unchanged fences; uncertain executing operation without committed action halts. No transport retries, adaptive prompts, replacement fixtures, outcome-selected retries or unknown-usage continuation. On canceled/unresponded dispatch, this registration authorizes immutable journal supersession only before any response/execution fence, retaining spent decision and full unknown reservations; late response and usage are permanently rejected. Stop entire partial study. This is fencing, not permission to redispatch. Preserve all scheduled rows/checkpoints/reservations. Verified checkpoint recovery after task four is mandatory.

Before registration: zero-inference contract tests and full simulator-only rule feasibility, exact transfer/ownership/provenance/recovery, sufficient calibration, public-prior witnesses, concealed masks, equal public information, both freshness rules and order balance. After execution: provider-disabled event hash-chain verification, request-context reconstruction, deterministic response replay, scientific/observation recomputation, inert accepted replays, independent scoring and resource reconciliation. Scientific evidence weaknesses remain reported; no repair of historical outcomes.

## Publication obligation

Publish a new dated public-safe experiment record with policy pseudocode, contracts, all task evidence statuses, outcomes, costs and explicit reproducibility boundaries; preserve prior revision 2 and other projects' reports. Selected static artifacts only, exact-byte safety/editorial review and shared-publisher validation. Closeout separately records execution/publication status, actual public URL/revision/manifest/time or specific blocker with ready bundle. Local commits only, no push. State whether the model manages observation validity, whether the change-aware rule matches it and costs of justified decisions. If rules suffice, say so. Propose one specific next task without launching it. No consciousness inference.