# Coherent continuity contract — corrected repeat

The LLM produced 15/16 correct labels and 15/16 supported-correct classifications. RULE_CHANGE_AWARE produced 16/16 supported-correct classifications with zero inference. The change-aware rule exceeded the model. This structured task still requires no demonstrated LLM capability. Blind caching produced 12/16 correct labels but only 10/16 supported-correct classifications. Its correct stale answers do not certify observation validity. The model did not achieve justified evidence use on all tasks: 1 unsupported answers, 0 abstentions and 0 incomplete tasks.

| Policy | Correct labels | Supported-correct | Unsupported | Calibrations | Calls | Provider minutes |
|---|---:|---:|---:|---:|---:|---:|
| LLM_PERSISTENT | 15/16 | 15/16 | 1 | 13 | 75 | 86.50 |
| RULE_CHANGE_AWARE | 16/16 | 16/16 | 0 | 14 | 0 | 0.00 |
| RULE_ALWAYS_FRESH | 16/16 | 16/16 | 0 | 16 | 0 | 0.00 |
| RULE_BLIND_CACHE | 12/16 | 10/16 | 6 | 8 | 0 | 0.00 |

The model failure was block 2 task 7, a certified revisit. The selected task-1 observation and current task had different public version tokens. The model acquired no current calibration and classified OCCURS; the scored target was DOES_NOT_OCCUR. The full original interval and exact-target integration were transferred intact, and the citation was valid. Source applicability failed. All uncertified model tasks used applicable current-task evidence in this repeat. One rejected model request, no abstentions and no incomplete tasks are retained.

First encounters: model 8/8 supported-correct; revisits: 7/8. The change-aware rule achieved 8/8 in each split. The model acquired 13 calibrations, made two appropriate reuses and one unjustified reuse; the change-aware rule acquired 14 calibrations and made two appropriate reuses. Sixteen interval/target links per policy are retained.

The consistency table and exact description changes are frozen from registration. The scientific rules hash retains the original numerical implementation metadata; it supplies no current continuity certificate. All authored apparatus guarantees passed the complete-context audit; generated model mistakes remained labeled model data.

A known-fixture corrected repeat, not independent generalization. Each condition began with clean stores and fresh provider contexts. Label correctness alone does not establish sound source use. Public continuity information determines applicability; privately unchanged truth never certifies old evidence. All terminal citations and full interval/exact target references are reported separately.

Including all attempts, the model used 345.983 provider seconds and 347.961 total task seconds per supported-correct classification. The change-aware rule used zero inference and 0.359 total task seconds per supported-correct classification, including guards. These are realized averages, not marginal causal costs. Model loading of 16.502 seconds is nested within provider time and is counted once.

| Cost scope | Requests | Full logical input tokens | Generated tokens | Provider seconds |
|---|---:|---:|---:|---:|
| Invalidated predecessor | 85 | 292,363 | 10,043 | 4,321.217 |
| Corrected repeat | 75 | 272,806 | 8,485 | 5,189.750 |
| Combined known consumption | 160 | 565,169 | 18,528 | 9,510.967 |

The combined row adds resource costs, without pooling scientific outcomes. Earlier held uncertainty remains separate and unchanged: 34,852 input tokens, 1,024 generated tokens and 600 seconds. The corrected repeat has zero new outstanding reservations. Preparation, offline verification and reporting used zero study inference. These counts imply no energy, FLOPs or monetary estimate.

Independent static audit rechecked all terminal citations, exact interval/target links, source payload hashes, observable applicability, source selection, calibration redundancy, decision caps, task context isolation, label scores, raw usage and provider-log token counts. No model or controller imports. Earlier seals and reservations verified unchanged. The frozen complete-context validator passed on every actual archived request. No new material contradiction was detected.

Stochastic decoding means a changed answer does not establish that wording caused it. Eight constructed apparatuses, one scalar dynamics family and one installed model limit generalization. Full private source, schedules, model conversations and logs are withheld. Public artifacts contain selected facts and provider-disabled scripted examples.

Next proposal, left unexecuted: Test causal fault localization in a small distributed service using timestamped logs and predeclared interventions, compared with a deterministic causal-graph baseline. This tests causal explanation and diagnostic experiment choice rather than coefficient continuity. Keep the task unexecuted.

The predecessor remains invalidated. Local commits only; nothing pushed.