This is the record of an abcd-managed repository

abcdBetav0.16.2 · 2026-11-06

Development record

Widen options, never recommend

principlewiden-options-never-recommend

The rule. At a conjectural question — one whose answer is a judgement the human owns, not a fact the record can settle — an abcd surface widens the space and stays out of it: it presents candidate construals (at most three, a ceiling not a target), always includes the null answer ("none of these" / "can't articulate yet"), and recommends nothing. Ordering, phrasing, and defaults must not smuggle a preference; the tool's contribution is that the human sees more of the space than they would have generated alone.

Why. An elicitation tool that recommends converts a question into a confirmation: anchoring does the deciding and the human does the assenting, and the record then carries the tool's frame with a human signature on it. The value of machine participation at conjectural questions is coverage, not selection — the same division of labour the record already runs elsewhere: the itd-84 validator proposes routing and the human adopts it (verifier-selects-gates-decide), and the interview's closing summary pre-formats selection grounds without selecting. Widening is what a machine can do without owning the judgement; recommending is owning it.

Bounds.

  • Applies at conjectural questions. Where the record settles an answer (a fact, an existing decision, a lint result), stating it plainly is not a recommendation — it is reporting, and hedging it would be false balance.
  • The ceiling is three construals because widening degrades into burying: a long menu is a recommendation by exhaustion. Fewer is always acceptable.
  • The null answer is load-bearing, never decorative: a hold (per the itd-142 hold register) is a first-class outcome, not a failure of the round.
  • Presentation neutrality is part of the rule: no starred defaults, no "(recommended)" labels, no ordering by the tool's own scoring at conjectural questions.

Promotion. Underway (staged rung). The cold-reading output contract carries a supply-regime field, and ingest checks each reading's output against what its regime licensed it to produce: an evaluative output containing an ordering, a score, or a single named recommendation is refused; an explicative output in which a surfaced claim arrives already dispositioned is refused; a registrative output attaching a proposed resolution to a detection is refused. The check ships enforced per the ruling, with a reserved degradation path: where a signature proves noisy in practice it records the regime and flags for researcher review instead, and the claimed property weakens from enforced to observed, said out loud per loud-staging. (Enforced-at-birth is in tension with this principle's own promotion rule — calibrated before it gates; the tension is recorded as standing, and the ruled design governs the instrument meanwhile.)

mentions2