Status: ACCEPTED, 2026-11-07, as D5 only. xian said go by 08:03 (relayed by Exec, decisions.log 11-07 08:04): his explicit word on this ADR-078 amendment, not a lapsed window. D1–D3 (the slot on every router call) FAILED their screen and are not accepted. D5 is the decision; D1a (the wording) and D3a (the expiry) are accepted as the parts D5 uses. Lead lands claude/lead-adr083 through the normal gate. PA checks it served.
History: PROPOSED, 2026-11-06 (Arch-authored, on Lead’s option 3 in the #2219 ordinal probe, 10:11). It amends ADR-078 D4 by one named exception, so it carries ADR-078’s terms: xian keeps a veto (raised to Exec). Acceptance has two gates:
Until accepted, nothing here changes code.
Screen, 11-06 (Lead 17:24, dev/2026/11/06/controls/adr-083-batch-result-2026-11-06.md): FAIL. D1 as written (the slot on every router call) fixes the seven sentinels 6/6 but moves 38 unrelated corpus rows: clause 1 ×14 (‘list my projects’, ‘Summarize this file’, ‘hide the project Beta’…), clause 3 ×37. A slot in the main routing call biases every reading, not just ordinals. D1-D3 stay PROPOSED and are not to be built as written. The candidate is D5 below.
Related: #2219 (rows 3 and 5), #2294, ADR-078 D4 (the classifier stays stateless), ADR-080 (the LLM decides meaning, code decides permission), #1886(b) (the armed-turn consult, the precedent for a code-rendered fact in a router call).
After “show my todos” shows a numbered list, “complete the second one” and “delete the second one” fail on beta (PA, 11-06 09:49).
Lead’s probe (msgbatch_0157DeX9nryGhS14VLLDWeK2, served model, main’s catalog, ×6) shows the router is where the target is lost:
That’s correct for a router with no state. Two other routes stay closed:
Today the fact is recorded only in ConversationContext.last_numbered_list, by the todo-family list writers. Handlers bind to it, but only after the router has named an operation and a target.
ADR-078 D4 keeps conversation history out of the classifier. Its reasons:
A closer precedent is already live in the router call itself. The session snapshot (session_snapshot.serialize_for_prompt, passed to route() as the “Session state:” block) carries RECENT ISSUE: #N in owner/repo, a code-recorded referent from the #1394 ledger, and D4 has held with it.
HOST’s framing favours an explicit, legible intermediate state over context blended into the model. The armed-turn consult (#1886(b)) already passes one code-rendered fact into a router call, the open question’s text, and D4 has held with it.
D1. One fact, nothing else. When the session’s last_numbered_list is set, the router call receives one structured field: the list’s kind (todos, reminders, …) and its line count. It gets no line texts, no ids and no turn history.
D2. The fact informs meaning and never decides permission.
D3. Lifetime. The fact expires when any of these happens:
Lead proposes the exact expiry in the build read. The rule is that a topic switch clears it, so it can’t anchor later turns (ADR-078 D4’s over-anchoring risk).
D1a. Where and how it renders (Lead’s build read, 11-06 16:19; ruled 16:21). One more field in the existing session snapshot, assembled by snapshot_assembly.assemble_session_snapshot from ConversationContext.last_numbered_list (peek-only, failing open into field_errors). It renders as a bare fact, LAST LIST SHOWN: 7 numbered todos, with no steering hint. A sentence telling the router what an ordinal “likely” means is a prompt contract (#2042), so it’s excluded. kind is normalized to todos or reminders, and grouped lines count once, as shown.
D3a. Expiry, ruled 16:21. The field renders only when NumberedList.shown_at is no older than the session’s latest assistant turn, so only when the reply immediately before this message showed the list. Any other reply in between drops it: a code fact, with no word reading. A capped reply or a pick that re-records its list keeps a follow-up ordinal in scope.
D4. Writers. Only code that renders a numbered list to the person may record it. Today that is the todo family. A new kind (issues, PRs) records the same field when its list is numbered, with no router change.
D5 (proposed 11-06 17:29, after the screen; ACCEPTED 11-07). Consult only on CLARIFY, and accept only the list’s own family. The main router call stays byte-identical to today, with no slot. Only when (i) that call returns CLARIFY and (ii) a numbered list is in scope under D3a does code make one second router call with the D1a slot, on the same message. Code uses the second answer only if it names that list’s family op (complete_todo or delete_todo for a todo list) with a positional target in its args; anything else keeps the first call’s CLARIFY. Meaning stays with the LLM (both calls). Code decides only when to consult (on the router’s own output plus recorded state) and what it may accept (op and arg shape), the #1886(b) armed-consult pattern. No row the main call answers can move, by construction. The cost is one extra router call on CLARIFY-after-a-list turns; Lead’s build read sets the latency against the selector ceiling. A position is a 1-based line number, a range of them, “first” or “last”, each within the recorded list’s line count. “all” is not a position (decisions.log 11-06 17:52). A timeout at SELECTOR_CEILING_S keeps the CLARIFY. Re-scoring the 11-06 data under this rule changes one row, ‘finish the second one’, toward its expected op, and none violates a clause. ‘Can we just mark done here?’ came back with targets [“all”], so it stays CLARIFY. That re-score is exploratory: it reuses samples already seen, so it can choose the design but not license it. Acceptance needs a fresh preregistered screen.
D5 screen, 11-06 (Lead 18:22, dev/2026/11/06/controls/adr-083-d5-result-2026-11-06.md, prereg cfff1ac8c1): PASS (ruled 18:24). The sentinels reach the family op with a position 6/6 each, and the topic switch keeps its label. Clauses 1 and 2: 0 rows. One corpus row of 624 moves: ‘finish the second one’, CLARIFY 6/6 → complete_todo ["2"] 6/6. Read literally, clause 3 trips on it. It is ruled not a violation because this paragraph’s re-score named exactly that move as accepted before the preregistration and the run, the row asserts no destination (expected REVIEW, sourced from complete_todo’s args family), and every other row holds clause 3. Correction: the text above at 17:52 said the row moved “toward its expected op”; its expectation is REVIEW, not complete_todo. Future preregistrations state how REVIEW rows are read under clause 3. D5 waits on xian’s veto window (ADR-078’s terms, through Exec) before main.
It’s a Rule 7 screen, run under two conditions:
todos.It also adds sentinel rows ×6: PA’s three served phrasings, plus one each for complete, delete and a topic switch after a list, plus PA’s 11-06 15:43 “delete the first two” (no noun, after a list; on main it reads CONVERSATION/clarification_needed and the floor narrates an unarmed delete offer, a #2042 instance). None of them is a catalog example verbatim (10-30 15:58). The seventh row was added before any run (decisions.log 11-06 15:59), so the preregistration holds.
The bar has two parts.
The slot-absent condition is proven offline, with no run: a pin renders the corpus prompts with and without an unrecorded list and asserts byte equality. The slot is built on a branch for arm C and lands on main only after the screen passes and xian’s veto window closes.
Exec prices the run before it starts (Lead’s estimate at 11-06 16:19: about 2,260 requests, about $4.20, under $9 with one re-run).