Loomworks — B-86's classifier dependency, re-measured against the deterministic replies — v0.1

Date. 2026-08-11 · Author. Claude Code (read-only; pure classifier probes, no writes). Operator: Marvin Percival. Fence. Read-only. Engine read and probed at 64610b4. Classifier model as production resolves it: claude-haiku-4-5-20251001, temperature 0. Why. The Operator's instruction: one live recurrence (E0127) is a trigger, not a measurement; before the 32-intent classifier change is accepted alongside the persist-the-outcome remedy, re-measure the dependency against the current deterministic replies — the deferral's premise was a dependency measured smaller than scoped.

Method

E0127's turns survive intact in conversation_turns — the live-failure history was replayed verbatim, not approximated. Probes call classify_intent with the production-resolved model and key, 3 trials per cell (temperature 0; every cell returned 3/3 identical).

Histories: H0 = none · H1 = the denial exchange only (draft ask + the CR-2026-200/B-87 composed denial, verbatim) · H2 = the full live history (the ask_about_past_input exchange, then the denial exchange — the state the misroute fired in).
Messages: M1 = you can draft specifications (the live sentence, lowercase, unpunctuated) · M2 = You can draft specifications. (the denial's own suggested canonical form).
Context: with E0127's real engagement summary + project id, and without either.

The measurement

| History | Message | In-project context | No context | |---|---|---|---| | H0 none | M1 | general_conversation 0.95 | — | | H0 none | M2 | general_conversation 0.95 | — | | H1 denial-adjacent | M1 | remember_about_me 0.92 | general_conversation 0.95 | | H1 denial-adjacent | M2 | general_conversation 0.95 | general_conversation 0.95 | | H2 live-full | M1 | general_conversation 0.95 | general_conversation 0.95 | | H2 live-full | M2 | general_conversation 0.95 | general_conversation 0.95 |

Every cell 3/3 stable. One wired cell in ten. (remember_about_me and add_knowledge are the two routes that run extract_delegation_grant on the raw message post-CR-2026-200; general_conversation routes to the responder with no grant recognition at all. The one wired cell does mint correctly today — CR-200 reads the Operator's own words, so the classifier's paraphrase, which reappeared in these probes as "Can draft specifications for this project.", no longer matters there.)

What the measurement says

The live misroute is deterministic, not a flake. Under the exact live history the grant sentence routes general_conversation at 0.95, three of three, with and without context. E0127 will reproduce every time.
The deterministic denial steers routing only in one narrow condition: denial immediately prior AND the lowercase form AND in-project context. One additional preceding exchange (H2) defeats it; leaving the project context defeats it.
The denial's own suggested sentence never routed to a wired path in any probed condition. The composed denial suggests "you can draft specifications"; an Operator typing it as a normal sentence — capitalized, with a period — measured general_conversation 0.95 in every history state, including denial-adjacent. The system's own recommended words do not do what the system says they will, in their most natural typed form. This is a new, harder member of the family than the deferral scoped: not history-pollution fragility but failure of the canonical happy path.
The deferral's premise is gone. The dependency measured larger than scoped, not smaller: deterministic misrouting in nine of ten cells. Whether the classifier change now rises to join the persist-the-outcome remedy is the Operator's ruling; this measurement is what it was waiting on.
One direction visible in the code, reported not decided: extract_delegation_grant is deterministic and history-independent, but it only runs if classification first lands on one of two intents. Running raw-message grant recognition ahead of intent routing (the way redirect_orient_recall_misroute already post-corrects the classifier structurally) would have caught E0127's sentence in every probed cell without touching the 32-intent taxonomy. Blast radius: the recognizer's own false-positive surface, already deliberately conservative — far smaller than a classifier-prompt change. A candidate for weighing alongside, not instead of, the persisted-outcome history work.

Interaction with the disclaimer scoping (filed alongside): a misrouted grant sentence today gets chat prose as its reply — E0127's second half. Until routing is fixed, the disclaimer is the layer that makes that failure visible; the two filings are one program seen from two ends, the same way the B-86/B-87 pair was.

DUNIN7 — Done In Seven LLC — Miami, Florida — v0.1 — 2026-08-11