enhance(#2229): add three-way claim disposition (admit/refute/abstain) to /gsd-explore research pass (#2543)
Closes #2229. Each claim surfaced by /gsd-explore's research pass is dispositioned admit, refute, or abstain, with abstentions routed to a visible ledger instead of being smoothed into confident prose. Refute and abstain are separated by whether the disagreeing source is authoritative for that claim; a strong prior is never authoritative alone. Two guards ride with it: conflict-abstention, and a tier floor that presents a would-be admit as an abstain when the researcher's resolved tier is the budget tier or cannot be determined. To make that floor enforceable, resolve-model now emits the effective tier (--pick tier). It was already computed above the resolve_model_ids omit gate but was unreachable from a workflow, which left the floor inert on every non-Claude install - the model id is blank under omit and runtime-substituted where a tier map exists, and the profile defaults to balanced. The tier signal mirrors every resolution step that can change which tier runs, including the model_policy preset, and reports unknown rather than guessing. Output is additive; model, profile and effort are unchanged. Two residuals are disclosed in the workflow rather than papered over: a raw-model-id model_overrides pin reports unknown and is floored (fails closed), and a model_profile_overrides entry repointing a tier at another tier's model can under-report (fails open, and predates this change). Admin merge used only to satisfy the missing secondary reviewer on a single-maintainer PR. No CI failure and no conflict were bypassed: 38 checks green, remote runner 32255/32255 on both Node lanes.
This commit is contained in:
@@ -404,6 +404,12 @@ The SINGLE source of truth for "did the phase SPEC supply section X (with at lea
|
||||
### MVP Mode
|
||||
Phase-level planning **enrichment** layered on top of the default tracer-first decomposition (see Tracer Bullet): it frames the phase goal as a User Story and, on Phase 1 of a new project, emits a Walking Skeleton. Vertical slicing itself is now the default, so MVP Mode no longer *turns it on* — it adds the user-story framing + skeleton. Resolved at workflow init via the precedence chain: `--mvp` CLI flag → ROADMAP.md `**Mode:** mvp` field → `workflow.mvp_mode` config → false. All-or-nothing per phase (PRD #2826 Q1). Surfaced as `MVP_MODE=true|false` to the planner, executor, verifier, and discovery surfaces (progress, stats, graphify). Canonical parser: `roadmap.cjs` `**Mode:**` field; canonical resolution chain documented in `workflows/plan-phase.md`. Concept index: `references/mvp-concepts.md`.
|
||||
|
||||
### Claim Disposition
|
||||
Three-way verdict every research claim carries before it may be shared, on the `/gsd-explore` Step-3 research pass (#2229): **admit** (survived a prompted-to-refute pass AND grounded in a source authoritative for that claim → stated with the source), **refute** (a source authoritative for that claim contradicts it → corrected, with the source), **abstain** (unverifiable, no primary support, a non-authoritative disagreement, a source-vs-prior conflict, or an untagged return → routed to the Unresolved Ledger, never smoothed into prose). Authority for the claim's own subject — not surprise — is the refute/abstain discriminator; a "strong prior" is never authoritative alone and can only produce an abstain. Two guards: **conflict-abstention** (a source-vs-prior conflict is never a silent pick-a-side) and the **tier floor** (every would-be admit is presented as an abstain when the researcher's resolved **tier** — `gsd_run query resolve-model <agent> --pick tier`, computed above the `resolve_model_ids: "omit"` gate in `resolveModelInternal` so it stays readable when the model id is blank or runtime-substituted — is `haiku` (over-defers to whatever source it was handed) or is `unknown`/`inherit`/empty (could not be determined, treated as potentially-cheap, never as verified-adequate); refute and abstain are unaffected. `--pick profile` defaults to `balanced` when `config.model_profile` is unset — it is not a tier signal. Disclosed residual, two cases: a per-agent `model_overrides` pin to a raw model id carries no tier, so it reports `unknown` and is floored — deliberate over-flooring in the safe direction; and a `model_profile_overrides.<runtime>.<tier>` entry that repoints a tier at another tier's model reports the asked-for tier, not the answering model's tier, so the floor stays silent on a cheap model — the one direction that fails open). A **prompt-level** judgment on this ideation surface: it deliberately does NOT call the verify-time `probe-core` disposition, which sits on the verifier↔predicate rail (ADR-857) and is out of altitude here. Claims-side analogue of the #1154 honest verifier (abstain-and-flag on the non-inferable; ADR-550 D4 — never a silent pass). Canonical text: `workflows/explore.md`. See Unresolved Ledger.
|
||||
|
||||
### Unresolved Ledger
|
||||
The named output bucket a `/gsd-explore` research claim lands in when its Claim Disposition is **abstain** (#2229). Presented *side by side* with the admitted claims rather than folded into the narrative, each entry carrying its abstention reason (`unverifiable` | `source-vs-prior conflict` | `non-authoritative source` | `tier-floor: unearned confidence` | `untagged — disposition not reported`). Empty sections are suppressed; an all-unresolved outcome is stated in one line, because that is itself the signal. Exists to make the absence of grounding *visible* — the failure mode it replaces is a plausible-but-ungrounded claim entering the conversation as confident prose (the class the research pass targets: recent / version-drift facts the model is measurably overconfident on). See Claim Disposition.
|
||||
|
||||
### User Story
|
||||
Phase-goal format under MVP Mode: `As a [role], I want to [capability], so that [outcome].` Required regex shape: `/^As a .+, I want to .+, so that .+\.$/`. Used as the framing input by `gsd-planner` (emits as bolded `## Phase Goal` header in PLAN.md) and as the verification target by `gsd-verifier` (the `[outcome]` clause is the goal-backward verification anchor). Authored interactively by `/gsd-mvp-phase`, validated by SPIDR Splitting when too large.
|
||||
|
||||
|
||||
Reference in New Issue
Block a user