From 5949234e249c3fdb4db1d69e977c436d38defeaa Mon Sep 17 00:00:00 2001 From: KN Date: Tue, 9 Jun 2026 04:33:55 +0300 Subject: [PATCH] chore(html-kit): close T21 + T25 + D01, add 18_ debate trigger spec MIME-Version: 1.0 Content-Type: text/plain; charset=UTF-8 Content-Transfer-Encoding: 8bit Board moves (16_session-backlog-board.html): - T21 (telemetry stack decision) — blocked → done. Laminar chosen (self-host gRPC on lan.ds.ke, cloud fallback). omp session parser + back-emit helpers shipped via insight-kit B1/B2. - T25 (multi-agent debate orch) — trigger spec landed; card points at 18_t25-debate-trigger-spec.md. Stays in `next` for implementation claim. - T24 (agent-as-judge + Harbor) — blocked → next. D01 (score gate semantic hardening) shipped via B1, unblocking the Harbor RL signal coupling. - D01 (score.py semantic hardening) — discovered → done. Pointer flipped to eval/README.md §score. New: 18_t25-debate-trigger-spec.md - 5 OR-composed primitives: score floor <0.7, critic-severity cluster (≥2 HIGH/CRIT), 3-run stagnation (Δ<0.05 & score<0.85), cite-graph contradiction via T29 supports/refutes edges, manual --debate override. - Asymmetric cost bias: false-positive (debate on easy run, ~5x) cheaper than false-negative (single-pass ships a bad claim into RL signal). - Env tuning hooks: T25_SCORE_FLOOR, T25_CRITIC_THRESHOLD, T25_STAGNATION_{WINDOW,DELTA,CEILING}, T25_FORCE_DEBATE. - Forward-pointer to insight_kit.platform.orchestration.debate (not yet implemented). Index (00_index.html): card 18 added under cards 16-17. Co-Authored-By: Claude Opus 4.7 --- .html-kit/00_index.html | 5 ++ .html-kit/16_session-backlog-board.html | 20 +++--- .html-kit/18_t25-debate-trigger-spec.md | 95 +++++++++++++++++++++++++ 3 files changed, 109 insertions(+), 11 deletions(-) create mode 100644 .html-kit/18_t25-debate-trigger-spec.md diff --git a/.html-kit/00_index.html b/.html-kit/00_index.html index 8702eb1..2af93f2 100644 --- a/.html-kit/00_index.html +++ b/.html-kit/00_index.html @@ -848,6 +848,11 @@

Three days of work, laid flatLaminar coverage — 2 of 7

Opus audit of the T17 Laminar wiring. Phase wall-times + artifacts captured; everything inside the 535s omp black box (per-turn LLM calls, ~30 DuckDB queries, persona switches) is opaque. lmnr-claude-code-proxy verdict, 1-hour quick wins, 1-week session-JSON post-parse path.

+
+
18 · T25 debate trigger spec
+

When to escalate to opus + N-builder debate

+

Cost-gated escalation trigger for multi-agent debate. Five primitives OR-composed: score floor (<0.7), critic-severity cluster (≥2 HIGH/CRIT), 3-run stagnation (Δ<0.05 & score<0.85), cite-graph contradiction, manual override. Asymmetric bias: false-positive (~5x cost) cheaper than false-negative (bad claim into RL signal).

+