conviction-signalling

Conviction Signalling

Conviction ("the wormhole") is the harness's honest channel for instinct: a generator that believes in a non-obvious branch can say so, and the system protects that branch from early pruning — without ever letting belief decide acceptance. Effect on exploration, never on acceptance: that line is the invariant the design rides on.

The signal

One line, machine-parsed (regex in pqa/signals.py; byte-identical copy in hooks/precipitate_capture.py — change both together):

conviction: high, basis: <one sentence naming the non-obvious reason this works>

• Levels: high | medium | low. Only high protects from early pruning (Conviction.protects_from_pruning).
• basis is mandatory and must name a mechanism: "burst absorption beats rejection for this producer mix" calibrates; "feels clean" does not.
• The branch digest carries the same conviction/basis fields, so the orchestrator, the capture hook, and the engine all see one story.

What conviction buys — and what it never buys

Stage	Effect of conviction: high
collision (early pruning)	protected — cannot be dropped for "looking weak" before attack completes
adversary	attacked HARDER, not softer — calibration needs the stress
verification	nothing; the verifier does not read it
collapse	nothing; BranchResult.conviction is telemetry, and scripts/check_invariant.py statically forbids it from entering _rank_key
taxonomy / calibration	everything — signals joined to outcomes are the moat

Protocol

1. Flag only when real. Use high only if you would defend the branch under attack. Faked conviction is exposed by calibration and the harness learns to discount you — the signal's entire value is its honesty.
2. Record at capture: pqa.memory.record_signal(conn, session, level, basis, branch=...) — the precipitate-capture hook does this automatically for subagent transcripts that contain the conviction line.
3. Back-fill outcomes after collapse: update_signal_outcome(conn, signal_id, survived=..., verified=..., won=...). A finished run with outcome-less signals is a calibration loop with no data — never leave them unfilled.
4. Read the calibration before trusting the instinct. pqa-self-reflector joins signals to outcomes and emits P(win | conviction=high) vs P(win | no signal). That number — not the feeling — is the empirical answer to "do hunches mean anything here?"
5. Treat high-conviction deaths as the prize. "High conviction, failed verification" is the single most valuable row the system records; the failure taxonomy carries conviction per row — tag it faithfully.

Worked example

b2 (scout) digests back: conviction: high, basis: callers can shed load better than any limiter because only they know what's droppable.

• Collision: another branch looks cleaner early; b2 is protected from pruning, so the attack proceeds on all branches.
• Adversary hits b2 hardest (per protocol). Finding: "assumes every caller implements shedding; the legacy cron client cannot."
• Verifier: b2 fails one integration test on the legacy-client path — verified: false.
• Collapse: b2 is out; conviction never consulted. b0 wins.
• Records: signal row (high, basis verbatim); outcome survived=true, verified=false, won=false; failure row carries conviction: high.

Five runs later the calibration table shows high-conviction branches winning at 2× the no-signal base rate in this repo — the instinct is real here, and protecting it next spawn is justified by data. Or it shows 0.5× — and the harness starts discounting. Either answer is the system working as designed.

Anti-patterns

• Conviction as argument. Quoting the basis to the judge at collapse is a contract violation; collapse ranks evidence only.
• Reflex-high. Flagging every branch high destroys pruning protection (everything protected = nothing protected) and ruins your calibration curve.
• Vibes basis. A basis that names no mechanism can only be counted, never calibrated.
• Orphan signals. Signals without back-filled outcomes are dead weight that biases every future read toward "unknown."
• Softening the death. Recording a high-conviction failure as "almost worked" is exactly the instinct-vs-reality divergence the signal exists to measure.