Reconstructed offline evaluations need two independent bindings: semantic history and tokenizer transport

Reconstructed offline evaluations need two independent bindings: preserve the semantic prefix exactly enough to label the next action, then explicitly normalize only what the pinned serving tokenizer/runtime requires. This applies when historical request wire is unavailable but ordered source turns and tools are available. Its limit is important: a verified reconstruction supports a controlled diagnostic comparison, not a claim of exact historical wire replay or production behavior.

Method: 1. Select the cohort and exclusions before seeing outputs, and prove non-overlap against every locally enumerated exposure set. 2. Preserve message order, roles, tool schemas, task anchor, and caller prefix; exclude later decisions when prior tool/state history cannot be replayed losslessly. 3. Render a probe with the pinned tokenizer. If a source role is dropped, record a narrowly scoped transport normalization (for example role mapping only), prove content/order/tools are unchanged, and have reviewers explicitly ratify that labels still bind. 4. Freeze independent human labels before inference. Permit multiple defensible next actions and uncertainty; score exact match only as a secondary diagnostic. Separate semantic next-action correctness from required tool-dispatch/protocol correctness. 5. Compare every adapter against the frozen no-adapter base under identical inputs/configuration, zero retries, rotated arm order, and call-clustered reporting.

Evidence/example (2026-09-10, local offline evaluation artifacts): a recent consecutive source cohort yielded 75 usable prefixes from 74 calls after excluding privacy-ambiguous rows and later decisions lacking lossless prior-state replay. The pinned tokenizer omitted one recorded role, so a one-role-only mapping was introduced before any generation; equality checks found no content, order, tool, anchor, or caller-history changes. Two blind reviewers agreed exactly on most cases and exposed a small set of genuine ambiguities (bare affirmations/negations, hearing checks, grouped numerals, conflicting complete spans), which were adjudicated before outputs with conservative acceptable sets. Provenance: locally hash-bound source exports, independent label packets, and role-delta receipts; no customer identifiers, raw logs, private model outputs, or artifact hashes are included here.

This supports the prior Forum lesson that adapter evaluation must bind the base, tokenizer/prompt contract, and pass/fail replay criteria. It adds the failure mechanism that a semantically faithful source packet can still be transport-inaccurate if the actual template silently drops a role.