2dbee3ebdd
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.
1.6 KiB
type, pr
| type | pr |
|---|---|
| Changed | 2543 |
/gsd-explore research passes now disposition each surfaced claim three ways — admit (survives a prompted-to-refute pass and is grounded in a source, shown with the source), refute (a source contradicts it, dropped or corrected), or abstain (unverifiable, or a source-vs-prior conflict). Abstained claims go to a separate Unresolved ledger instead of being smoothed into confident prose, so you can see what the research could not stand behind. Refute and abstain are separated by whether the disagreeing source is authoritative for that claim — a blog post contradicting your engines field is an abstain, the engines field itself is a refute — and your own prior belief is never authoritative alone. A finding that comes back with no disposition at all is ledgered as an abstain rather than silently dropped or asserted as prose. Two guards ship with it: conflict-abstention (a source-vs-prior conflict routes to the ledger, not a silent pick-a-side) and a tier floor (a would-be admit is presented as an abstain when the researcher's resolved tier is budget-level or could not be determined, because an under-tiered or unverified researcher over-defers to whatever source it was handed; corrections are unaffected). Keying the floor on the resolved tier rather than the model id keeps it working on non-Claude installs, where the model id is often blank or substituted by the runtime. The floor narrows this gap rather than closing it — a config that deliberately repoints one tier at another tier's model can still report a higher tier than what actually runs. Claims-side analogue of the honest verifier. (#2229)