Expert #12: bounded request-classification substrate #26

Closed
lost-rob0t wants to merge 0 commits from rage/12-request-classifier-red into main
lost-rob0t commented 2026-08-31 08:11:09 +00:00 (Migrated from github.com)

Consumes #12 under #8.

Canonical ARADR: research/LLM-LOG-RESEARCH-036-request-classification-expert.org.
Implementation evidence: evidence/LLM-LOG-EVIDENCE-017-request-classifier-red-green.org.

Executed RED: exact head f5e15c2f1fbd8800d767ab8eb701f353ab151668, hosted run 33371681379. Existing #10 Common Lisp/Tek9/SWI contracts passed before the new classifier contract failed on the missing multi-label behavior; Python expert-policy and Home Manager remained GREEN.

Implementation keeps the accepted architecture: Python remains capture/transport; Common Lisp owns typed validation, bounded Tek9 materialization/persistence and SWI-Prolog supervision/result validation; SWI-Prolog owns the declared request-classification rules. No Python rule engine, no arbitrary Prolog callable term, no whole-corpus load, and classification metadata grants no authority.

The first bounded GREEN surface persists exact user-message source evidence separately from durable derived assertions, validates rule/expert/evidence provenance, bounds results, and preserves the old #10 event_transport fixture path. Existing Tek9 public point lookup/write/transaction APIs were sufficient; no Tek9 repository mutation is included. No dotfiles consumer mutation is required by this slice.

Do not close #12 solely from this PR: ambiguity coverage and bounded indexed query acceptance by request/user-message/provider/model/time remain follow-up #12 work under the canonical design.

Merge only an exact current head that is GREEN and mergeable; zero reviews/comments are not approval. PR #5 remains a separate transaction.

Consumes #12 under #8. Canonical ARADR: `research/LLM-LOG-RESEARCH-036-request-classification-expert.org`. Implementation evidence: `evidence/LLM-LOG-EVIDENCE-017-request-classifier-red-green.org`. Executed RED: exact head `f5e15c2f1fbd8800d767ab8eb701f353ab151668`, hosted run 33371681379. Existing #10 Common Lisp/Tek9/SWI contracts passed before the new classifier contract failed on the missing multi-label behavior; Python expert-policy and Home Manager remained GREEN. Implementation keeps the accepted architecture: Python remains capture/transport; Common Lisp owns typed validation, bounded Tek9 materialization/persistence and SWI-Prolog supervision/result validation; SWI-Prolog owns the declared request-classification rules. No Python rule engine, no arbitrary Prolog callable term, no whole-corpus load, and classification metadata grants no authority. The first bounded GREEN surface persists exact user-message source evidence separately from durable derived assertions, validates rule/expert/evidence provenance, bounds results, and preserves the old #10 event_transport fixture path. Existing Tek9 public point lookup/write/transaction APIs were sufficient; no Tek9 repository mutation is included. No dotfiles consumer mutation is required by this slice. Do not close #12 solely from this PR: ambiguity coverage and bounded indexed query acceptance by request/user-message/provider/model/time remain follow-up #12 work under the canonical design. Merge only an exact current head that is GREEN and mergeable; zero reviews/comments are not approval. PR #5 remains a separate transaction.
nsaspy changed title from WIP: Expert #12: bounded request-classification substrate to Expert #12: bounded request-classification substrate 2026-09-04 14:35:41 +00:00

Pull request closed

Sign in to join this conversation.
No description provided.