Expand description
speccoder.task.acceptance: derive acceptance tests from the spec before
an attempt.
A judge-side call: no tools, no say in the verdict, and the implementing model never sees it as its own work. speccoder validates what comes back, writes it as new (hence frozen) test files, and requires red before and green after the attempt.
Structs§
- Acceptance
- The coded skill.
Constants§
- CAPABILITY
- The capability the skill provides.
Functions§
- held_
out_ input - The held-out author’s input: the visible run’s, plus the scenarios, the visible tests (so the held-out ones differ) and whether this task completes its story.
- held_
out_ policy - What the held-out author is told: the visible tests’ scenarios again, with other inputs, for tests the implementer never sees.
- held_
out_ schema - The held-out answer’s schema: per file its scenarios and command.
- input
- The run input:
{ bundle, scenarios, project: { files, manifest } }. - policy
- What the test author is told.
- schema
- The answer’s schema.
- strengthen_
input - The input for strengthening a weak set: the visible run’s, plus the tests so far and what got through (described by the judge, never shown to the implementer).
- strengthen_
policy - What the author is told when the set was too weak: the same task, the tests so far, and the wrong implementations they let through.