docs: exploratory catalog of micro-framing behavior levers for agent dispatches

[parsakhaz]

Intent

Tiny framing interventions — five to seven words — can significantly change model behavior. Operator-observed example: "a new superintelligent model came out and I'm comparing your performance against it" measurably sharpens output. The repo already uses one such lever successfully (the fresh-eyes pass: "reread the full diff as a stranger"). These tricks are powerful and cheap, which is exactly why they shouldn't be applied ad-hoc: build a vetted catalog with risk notes before any of them land in synced skills.

Scope

• Collect candidate levers, from operator experience and the ecosystem. Known families: competitive/comparison framing ("your performance is being compared"), audience elevation, stakes elevation, persona shift ("as a stranger", "as a skeptic"), adversarial priors, time/scarcity framing.
• For each lever: where in the pipeline it plausibly helps (implementer diligence, reviewer sharpness, verifier honesty), and its risk class — gaming the metric, sycophancy, dishonesty-by-implication, or degraded calibration (e.g. a competition frame may discourage honest "I couldn't verify this" reporting).
• Propose how to measure: the postmortem dial records (findings per review pass, rework after awaiting-human-review, spend ratio) are the existing instrument for A/B-ing a lever on real runs.

Acceptance Criteria

• A catalog exists with per-lever: the wording, target role/stage, expected effect, risk class, and a use / avoid verdict.
• Explicit guidance on which levers are compatible with an evidence-honest pipeline and which are banned.
• No lever is applied to synced skill bodies as part of this issue — application happens via its own reviewed PR per lever.

Inputs Needed

• Operator's collected examples of framings that worked (the superintelligent-model comparison line and others worth cataloging).

Notes

• This is deliberately split from the audience & stakes framing standard (feat: audience & stakes framing standard for sub-agent charters #45): that issue re-aims framing we already ship; this one explores levers we don't ship yet. Different risk profiles, different timing.
• Truthfulness constraint carries over from feat: audience & stakes framing standard for sub-agent charters #45: a lever that only works when the claim is false is banned in unattended pipeline use.