Negative-framing sample-reframe — top-3 agent files (harness-retirement F-007, next minor)

[Imbad0202]

Deferred finding F-007 from the 2026-06-10 harness-retirement audit (#384, applied sweep #383).

Negative-framing density in the top agent files (count of don't / do not / never / avoid / must not):

41  academic-pipeline/agents/pipeline_orchestrator_agent.md
33  deep-research/agents/socratic_mentor_agent.md
28  academic-pipeline/agents/integrity_verification_agent.md

Most hits are hard gate/contract boundaries where the negative form is correct and stays (e.g. "do not abort Phase 1, do not attempt schema repair" — the positive rewrite would be ambiguous). But at this density a fraction will be reframeable as shorter positive directives, which current models follow more reliably.

Scope (deliberately bounded):

• Top-3 files only, one small PR each — not a bulk rewrite across the suite.
• Keep every negative that is a hard boundary; reframe only where a positive form is equally precise and shorter.
• Run the full lint suite per file — several lints pattern-match exact prose (the fix(harness): retire Fable 5-stale model pins — opus command floor, display names, gpt-5.5 lineup #383 lesson: a lint that silently stops matching is a fake-green).

Timing: next minor release (v3.13), per the audit's batch-order recommendation (Cat 6 after Cat 1/3 mechanical fixes).

[Imbad0202]

Sample-reframe results for the three files.

socratic_mentor_agent.md — reframed in #420 (open). This file carries genuine style/quantity directives phrased negatively ("avoid lengthy lectures" → "keep it brief", "do NOT use this more than twice per Layer" → "use this at most twice"). These reframe cleanly because the positive form is genuinely shorter and clearer to read — a maintainability gain, independent of any claim about model behaviour. That is the reframeable fraction this issue anticipated. The maintainer pass on that PR restored three hard-boundary negatives a positive rewrite had weakened (the never give direct answers identity line, the do not praise / evaluate / grade reading-probe leak rule, the #393 single-source anchor rule) and reverted two out-of-scope edits (a synonym swap and a Socratic-question wording change).

integrity_verification_agent.md (28 hits) and pipeline_orchestrator_agent.md (55 hits — drifted up from the 41 in the issue body as the file grew) — no reframe candidates. Every negative in both files is either a gate/contract prohibition (Stage 2.5 can NEVER be skipped, NEVER rely on AI memory to verify a reference, the cross-model consent gate's do not send automatically), a verdict-precedence rule (never to PASS WITH NOTES, advisory notes never downgrade a FAIL), a concurrency/single-reader boundary (Do NOT release [the lock] between steps, the finalizer NEVER reads terminal_policies), or plain narrative fact (SLIM checkpoints never reset, old patches never apply across a re-emission boundary). A positive rewrite of any of these either softens the prohibition or inverts the meaning — the same failure the #420 maintainer pass had to undo.

A structural note on why the count is zero rather than just low: the actual quantity caps in these two files (response length, round limits, the ≥5-turn guard) are already written in positive/bounded form, so they never enter the negative-hit set. There is no negatively-phrased style directive to convert. These are an orchestrator and a verification gate; their negatives are control-flow, not prose hedges.

A note on the test the reframe actually applied: the question per line was "is the positive form shorter and clearer here?", not "do current models prefer positive directives?". The latter is a weaker basis than it looks — frontier models follow a well-formed negative prohibition reliably, so reframing a hard boundary buys nothing and risks softening it. Where a reframe earns its place it is on readability grounds, and that is what the socratic file offered and these two did not.

So the sample lands where the issue expected the boundary to be: the conversational file had a reframeable fraction (now in #420), the gate and orchestrator files did not. Closing F-007 with the one PR rather than forcing a second/third — keeping a negative that is a hard boundary is the issue's own rule.

[Imbad0202]

Closing with the merge of #420. Sample-reframe verdict (comment above): socratic_mentor_agent.md carried the only reframeable fraction (now merged); integrity_verification_agent.md and pipeline_orchestrator_agent.md have zero reframe candidates — every negative there is a gate/contract prohibition, verdict-precedence rule, concurrency boundary, or narrative fact, where a positive rewrite would soften or invert meaning. F-007 lands as one PR, per the issue's own keep-hard-boundaries rule.