0109: Pace and classify semantic reading
Status: accepted (2026-08-23) · Scope: semantic scheduler, response contract, status, Codec
Context
Section titled “Context”Automatic semantic reading filled its 60-call rolling ceiling during the first live production window. The window recorded 44 accepted readings, 15 invalid readings, one call without a reported outcome, 877,493 input tokens, and 55,370 output tokens. Routine generations could be read every 30 seconds when their recent action list changed. One generation was read 14 times in about 15 minutes.
The stored invalid receipt named only invalid_output. That protected raw model
text, but it did not show whether the failure came from structure, request
identity, citations, privacy, or an unsupported claim. A six-request diagnostic
kept raw replies in memory only and returned five accepted readings. The one
invalid reading cited a well-formed event ID that was not present in its
evidence.
Alternatives
Section titled “Alternatives”Lower the global hourly ceiling. Rejected because one busy generation could still consume a disproportionate share of a smaller budget. The scheduler needs to pace each generation before the root-wide ceiling becomes the only control.
Apply a five-minute interval to every reading. Rejected because explicit waits, blockers, errors, and terminal transitions should reach Codec sooner than routine action churn.
Persist detailed validator messages or raw replies. Rejected because raw replies can contain private text. Detailed schema messages are also an unstable status contract. Operators need stable categories, not model output.
Repair invalid replies automatically. Rejected because a second model call would increase usage and a local rewrite could turn rejected synthesis into an uncited claim. Invalid output remains a typed failure.
Decision
Section titled “Decision”Prompt contract version 4 binds every response citation to the exact
evidence_event_ids supplied in that request. The prompt also tells the reader
to copy those IDs rather than reconstructing them. The existing validator stays
in place as the persistence boundary.
Invalid documents and call receipts store one or more controlled categories:
schema, identity, citation, basis, confidence, privacy,
unsupported_claim, or unknown. Usage aggregates count those categories for
the rolling hour, the reader process, and each harness/model breakdown. Codec
shows the counts in its existing semantic usage panel. Harnery never stores the
rejected model reply.
The scheduler keeps a 30-second minimum interval for priority evidence and uses
a five-minute minimum for routine evidence. Priority means an explicit
attention signal, a non-active lifecycle change, or terminal evidence. A
resumed active lifecycle is routine after its first evidence digest because
the lifecycle field remains present in later envelopes.
The five-minute invalid retry cooldown, 60-call rolling ceiling, fixed model routes, privacy filter, evidence bounds, and deterministic Codec precedence do not change.
Result
Section titled “Result”Applying the new intervals to the live rolling history would have retained 32 of 54 calls and avoided 22, a 40.7% reduction. This is a counterfactual using the generations’ audit-time priority bands, not a replay of every historical envelope.
Focused tests bind citations to the supplied ID set, classify invalid receipts, aggregate reason counts, distinguish priority and routine intervals, and keep a resumed active lifecycle out of the priority band. The semantic and Codec semantic suites pass with the new prompt contract.
The live daemon restarted on prompt contract 4 while retaining all 47 calls still inside its rolling hour. Its first contract-4 pass produced three accepted readings and one call without reported usage. No rejected model text was stored.