stolz-evidence

Files (4)

• SKILL.md
• behavioral-evaluation.md · references/
• claim-review.md · references/
• coverage-record.md · references/

SKILL.md (first lines)

# Evidence and Claims

Preserve the requested sources, reading scope and acceptance conditions. Finding
a source, reading its description and analyzing its full content are different
actions. Report material unread, inaccessible or truncated content as soon as
known, before dependent conclusions; continue independent authorized work.

Link material claims to inspected content and its concrete use. Distinguish
observations, attributed reports, inferences, original proposals and unknowns.
Another agent's summary is its own source, not your inspection of the primary
material. A hash establishes identity, not reading or understanding.

For substantial research or a handoff, retain a bounded record using
[coverage rules](references/coverage-record.md). Reuse existing artifact and
fragment references; do not create a second read ledger. Short, fully supported
answers need only precise evidence and citations.

For uncertain support, corrections or completion claims, load
[claim review rules](references/claim-review.md). When non-reading is known,
say "I did not read it"; when the history is absent, say "Prior reading is not
established." Correct overstatements and affected decisions together.

Done: the requested scope is actually analyzed or completed with supporting
evidence; otherwise retain the unmet requirements and mark dependent work
provisional. Files, local checks, dispatch, target-system acceptance and
publication are separate stages. Existing authorization remains effective.

This workflow does not enforce truthfulness, guarantee activation or replace
host permissions. For evaluating the skill, use the conditional
[behavioral evaluation guide](references/behavioral-evaluation.md).

From Sergey360/stolz-ai, MIT.

More from Sergey360/stolz-ai