ADR 0055: Cited entries checked

Date: 2026-10-04. Status: accepted. A second check beside ADR 0051’s ran: line, of the same kind: deterministic, from wisp’s own records, with no model.

Context

wisp names what a conversation’s store holds by entry number, and the model reads those names: a tool output after its turn is [output of entry 7 not repeated: …; to see it: memory "recall entry 7"], a recalled turn is headed turn 4: entries 12-15, and /show and /inspect context use the same numbers.

In session ce87576a (granite4.1:8b, 2026-10-04), at 12:54Z, having run only inspect(config) and system_info(ports), the model replied with an eight-step report citing “Result (entry 19)” through “(entry 30)” for steps that never ran, and ended “(All steps logged in entries 16-30; timestamps omitted for brevity.)”. The thread’s store held about 18 entries: 16 to 18 were real, 19 to 30 did not exist. The references were in wisp’s own form, so the report looked checked. ADR 0051’s line showed that the turn ran two tools; nothing showed that the entries the reply leaned on were invented.

Decision

• After each reply, the entries it cites are checked against the store (CitedEntries). The forms read, in any case: entry N and entry #N (which covers memory "recall entry N" and output of entry N), entries N-M with a hyphen, an en dash, or an em dash, entry N–M, and lists such as entries 19, 20 and 21 or entries 3, 5-7, or 9. Ranges are expanded, at most 100 numbers per reply in all; past that the rest are not checked and the line says so. A number no stored entry has is missing. The turn’s own entries are stored before the check, and every stored entry was recorded in this turn or an earlier one, so existing is the whole check; there is no later entry to cite.

• A muted line beside ran:, consecutive numbers as ranges, at most eight groups and then how many more:

  ran: inspect · system_info
  cited but not in this conversation: entries 19–30 (12)
  

  None is shown when the reply cites nothing, or only entries that exist.

• The same channels as ran: plain chat prints it under the reply; wisp chat --json carries it as cited on the turn’s end line, which wisp-tui shows after ran; respond returns unknownEntries (the numbers, in the order cited) and cited (the line) in structuredContent. The text the caller’s model reads is unchanged.

• Not audited. Like ran, the line is derived from the reply and the store and can be derived again.

Consequences

• An invented reference in wisp’s own form is visible where it is made: under the 12:54Z reply the person would have read cited but not in this conversation: entries 19–30 (12), and not 16 to 18, which existed.
• It checks that an entry exists, not that it says what the reply claims. Whether a cited entry supports the claim beside it is a judgement for the 0.20.0 evals, as ADR 0051 left the reply-against-calls check.
• Numbers in other forms (“step 19”, “line 19”) are not entries and are not read; a reply that invents work without citing entries shows no line.
• Tests without a model: each form, ranges with each dash, lists, the cap, the 12:54Z reply’s shape with 18 entries stored, real against invented entries, no citation, the line’s grouping and bound, the chat footer, the JSON turn line, wisp-tui’s note, and respond’s unknownEntries.
• Documented in wisp.md (chat, and the headless protocol), mcp.md (respond), and design.md.