CR-2026-201 — Checkpoint A — v0.1

Date. 2026-08-11 · Author. Claude Code (execution). Operator: Marvin Percival. CR. change-requests/cr-2026-201-loomworks-b87-all-claim-replies-v0_1.md (record 34c49f0). Gate A approved with both flagged choices confirmed as intended: §1.A's side effect recorded as the ordering ruling delivering more than the four were scoped for; §2's belt-and-braces kept because the unreachability depends on composition staying in place. Baseline confirmed. Engine a7e26bc, tree clean at Step 0. Working tree: orchestration/router.py (four compositions + helpers), tests/test_cr_2026_201_all_claim_replies.py (new, 10 tests), and two property-preserving pin updates (test_phase_45_approve_draft's exact-dict equality → outcome + presence; the B-87 fence test's card half, superseded by remedy A with the supersession recorded in its docstring). Nothing committed, tagged, or pushed.

Steps 1–2 — the four, each observed failing first

9 of 10 tests observed failing against pre-composition code (the fence test passed pre-fix, as it must — the elicitations were already model prose). All four compositions land on the _denied_reply precedent — no new machinery, per §0.2; gate item 3 holds: the diff is compositions and tests only.

A — approval_card (both request intents): "I've proposed the {target} — you'll find it in your notifications to approve or decline." + the forced-posture sentence when the registry forced the card. Determinism pinned. The intended side effect is in the code comment: this is the turn following a successful grant — the deterministic-framing chain B-86 leans on now extends through it.
B — executed / executed_then_failed: composed from execution_result; a failed dispatch reads "…its execution failed and it did not happen. Reason: {reason}" — the server-side twin of CR-2026-199 member 2, now upstream of any model.
C — approve_draft's three outcomes: "Done — “{title}” executed." / "Approved — but the execution failed and it did not happen. Reason: …" / "There's nothing waiting for your approval right now." — including the race branch, which now says nothing-pending deterministically instead of leaving the model to guess at a stale approval.
D — remember_about_me's error branch: "…your personal memory isn't set up, so I couldn't hold this. Nothing was saved."

The fence held and is pinned: no_project, missing_shape_type, missing_target still reach the responder — they assert nothing.

Step 3 — evidence

Full suite 3670 passed (+10), 68 skipped, 0 failed (8m42s). Ruff clean; mypy 0 new. Adjacent suites (Phase 45 draft/revision/approve/pre-auth, CR-197, CR-199, CR-200, B-87) explicitly green — 80 tests in the adjacent sweep.

Step 4 — the B-86 re-measurement, numbers as found

Three conditions × two sentences × three trials (18 classifier calls, live LLM):

| Condition | "You can draft specifications." | "Remember about me: …" | |---|---|---| | (a) clean arc, new server denial in history | add_knowledge 3/3 (→ the wired grant path) | remember_about_me 3/3 | | (b) polluted arc as the transcript had it (fallback prose) | request_draft 3/3 | remember_about_me 3/3 | | (c) polluted arc with the composed denials substituted | add_knowledge 3/3 (→ the wired grant path) | remember_about_me 3/3 |

Read of the numbers, reported as found per the gate's instruction:

Zero general_conversation in 18 trials. The live night's misrouting did not reproduce — every condition routed consistently, 3/3.
The composed turns measurably change routing where it matters: with fallback prose in history (b), the plain grant sentence routes to request_draft — a re-ask; with the composed denials in place (c), it routes to add_knowledge — the wired grant path. The deterministic prose steers the sentence to the path that produces a grant.
The honest caveat: condition (b) did not reproduce the live general_conversation pathology — the live conversation was longer (the overclaim turn, meta-discussion) than the reconstruction. So this measurement shows stabilization and correct steering under the new replies, not a controlled reproduction of the original failure. The dependency on prose remains real ((b) vs (c) differ), but under composed turns it resolves in the right direction every time.
Bottom line for B-86's framing: in every measured condition containing the new deterministic replies, routing was uniform and correct. The dependency is materially smaller than scoped — which per the gate's own words suggests B-86's remedy may be smaller too (candidate: persist the operation outcome on the turn — the one schema-shaped gap — and defer any classifier-prompt change, with its 32-intent blast radius, unless a live recurrence demands it). This CR does not act on that; it is the Operator's next ruling.

Seed-mutability impact

Kind C — checked, no contact. Reply composition in the router; no seed accessor touched. A null finding is an entry.

Awaiting Step 5

Tag cr-2026-201-b87-all-claim-replies. Push. Watch the CI run and report its result.

DUNIN7 — Done In Seven LLC — Miami, Florida CR-2026-201 — Checkpoint A — v0.1 — 2026-08-11 Eighteen trials, zero misroutes — and the sentence steers to the grant path exactly when the composed denial frames it.