04 / Specs
Selected requirements from Merlin’s SpecKits. Each excerpt names the criterion it comes from. Read the behavior the system is designed to uphold.
005 / Evidence and grounding
Observations must remain attributable, so a reader can distinguish a supported claim from an assertion.
Recorded file paths, symbols, and op-result keys are included in the snapshot.
Source: specs/005-evidence-provenance-grounding/spec.md
Selected acceptance excerpt · 27 September 2026
006 / Analysis
Rebuttal has explicit outcomes. Confidence alone does not turn a candidate into an established finding.
Rebuttal result is one of `survived`, `refuted`, or `inconclusive`.
Source: specs/006-analysis-v2-state-machine/spec.md
Selected acceptance excerpt · 27 September 2026
003 / Mutation lifecycle
A staged change and a changed workspace are different states. The transition matters.
The real workspace is changed only by commit.
Source: specs/003-mutation-lifecycle-and-staging/spec.md
Selected acceptance excerpt · 27 September 2026
004 / Final authorization
The same rule illustrated by the operating record on the Proof page. Incomplete work can be reported honestly; it cannot be called complete.
After model authorization, `ALLOW_FINAL` and `SEND_TO_EPISTEMIC_FINAL` are deterministically rejected whenever `task_summary` reports pending or in-progress work through either its exact counts or total-count fields. This veto is post-authorizer only: `ALLOW_INCOMPLETE_STATUS_FINAL` remains available for an honest incomplete report.
Source: specs/004-final-authorization/spec.md
Selected acceptance excerpt · 27 September 2026
020 / This website
The marketing site follows the same standard: say where the boundary is, including the option to cross it.
`/run-it-yourself` states plainly that cloud escalation exists for long context, is config-gated, and can be switched off.
Source: specs/020-marketing-website/spec.md
Selected acceptance excerpt · 27 September 2026
How to read this collection
These are selected acceptance excerpts, reviewed on 27 September 2026, rather than the full internal SpecKit corpus. They explain the intended behavior. The Proof page separately shows an observed decision.
