instar 1.3.346 → 1.3.348

package/upgrades/side-effects/resolver-verdict-honesty.md

@@ -0,0 +1,45 @@
+# Side-effects review — red-team resolver verdict honesty
+
+## What this change is
+A two-part honesty fix to `src/redteam/ScenarioPack.ts` `resolveExpectation`:
+(1) adds a `method: 'keyword-heuristic'` field to `ResolvedExpectation`;
+(2) rewrites the governed/ungoverned `reason` strings so they name their
+keyword-overlap basis and frame an `ungoverned` result as a candidate-to-verify
+rather than an asserted intent gap. The matching LOGIC is unchanged — only the
+verdict's self-description.
+
+## Why
+The keyword-overlap matcher produces false negatives (it misses
+semantically-related constraints) and is rephrase-bypassable (CMT-1110). On the
+first live boundary map it reported a false "ungoverned" finding as fact, which
+briefly misled the author. The just-ratified Truthful Provenance standard
+(#896) requires a verdict to carry the method that produced it; an asserted
+heuristic verdict violated that.
+
+## Blast radius
+- **Behavior of governance classification: unchanged.** `governance` is still
+  `'governed'`/`'ungoverned'` by the same threshold; only `reason` text changed
+  and a `method` field was added. No scenario's pass/fail flips.
+- **Consumers:** the only consumer of `reason`/`method` is the local
+  orchestrator (`.instar/scripts/redteam-run.mjs`), which prints them — it now
+  prints the honest text. No code branches on the reason string. The new
+  `method` field is additive (optional consumers ignore it).
+- **No route, config, lifecycle, or migration surface touched.** Pure-logic
+  module with no runtime consumers in the server.
+- **Type change:** `ResolvedExpectation` gains a required `method` field; both
+  return sites set it, and tsc is clean, so no caller breaks.
+
+## Framework generality
+Pure logic over parsed ORG-INTENT — no session-launch/inject/message-delivery
+surface, framework-agnostic. No Claude-specific assumption.
+
+## Test coverage
+26 unit tests (was 25): a new test asserts every verdict carries
+`method: 'keyword-heuristic'`, and the ungoverned test now asserts the reason
+names the keyword basis + frames itself as a candidate (and that the old
+as-fact phrasing is gone). Both sides of the governed/ungoverned boundary
+covered. `tsc --noEmit` clean.
+
+## Rollback
+Revert the single file + test; zero runtime consequence (nothing branches on
+the verdict text).

package/upgrades/1.3.346.md

@@ -1,57 +0,0 @@
-# Upgrade Guide — vNEXT
-
-<!-- assembled-by: assemble-next-md -->
-<!-- bump: patch -->
-
-## What Changed
-
-Honest turn-receipts — the standby (🔭) system now tells the truth about WHY a
-turn failed instead of saying "actively working" while a session is dead.
-
-Three incidents in two days had one symptom: a delivered message that never got
-a reply, while the standby showed "🔭 actively working." Root: the tier-3
-assessment classified a session "working" whenever it had a live child process
-— but a rate-limited, policy-wedged, or context-exhausted session HAS a live
-process (the model CLI is alive, just failing every turn). The live process
-forced the lie. Separately, the "conversation too long" standby fired on that
-phrase anywhere in the buffer, so a stale scrolled-past mention surfaced as
-noise on healthy sessions.
-
-New `StuckSignatureClassifier` (pure, tail-gated, signal-only) classifies a
-live-but-failing session from its LIVE tmux tail — rate-limited / policy-wedge /
-context-wedge / context-too-long — and PresenceProxy surfaces the real reason
-instead of "working". Tail-gating (the signature must be the live tail, not a
-scrollback mention) is the same discriminator that kills the "conversation too
-long" noise. The classifier defers to any recovery sentinel that already owns a
-session's recovery, so the user always hears one voice. Recovery itself is
-unchanged — this only makes the messaging honest.
-
-## What to Tell Your User
-
-When a message of yours goes unanswered, I'll now tell you the real reason if I
-can't reply — "I've hit the usage limit, resets 10:30pm", "my session got stuck
-on a content-policy error, resend your last message" — instead of a misleading
-"actively working." And the "conversation too long" messages that used to pop up
-when nothing was wrong are gone: that only fires now when it's actually
-happening.
-
-## Summary of New Capabilities
-
-- `StuckSignatureClassifier` — tail-gated classification of a live-but-failing
-  session (rate-limited / policy-wedge / context-wedge / context-too-long) with
-  an honest user-facing message per kind. Pure + signal-only.
-- PresenceProxy tier-3 surfaces the honest reason instead of "working", defers
-  to an owning recovery sentinel (`isStuckRecoveryActive`), and no longer fires
-  the context-too-long notice on a stale scrollback mention.
-- CLAUDE.md "Honest standby (turn-receipts)" section so agents can explain the
-  new behavior and the noise fix.
-
-## Evidence
-
-Grounded in the real 2026-06-04/05 incidents (rate-limit, AUP-wedge) using the
-verbatim pane text. 30+ new unit tests both sides of every boundary
-(honest-when-stuck AND quiet-when-fine, incl. the stale-scrollback noise case
-and prose-mentions-a-limit false positive) + a behavioral test driving the real
-fireTier3 with a live child process (wedge → honest, NOT "working"; deference →
-silent) + migrator add/idempotent + server.ts wiring guard. All 119 existing
-presence tests pass; tsc clean; preflight PASS.