Route compare probe discipline

Job

Prevent false localization when a circuit edit changes ops.bin. residual_probe and trusted eval_circuit derive test vectors from the op stream. If the edited route changes op count, op order, nonce tail, or targets, then the baseline and candidate are usually tested on different Fiat-Shamir samples. A dirty candidate screen can be an actual bug, a sample-shift exposure of pre-existing dirt, or both.

Use this skill before claiming a specific call, row, carry, codec window, or fold override is semantically dirty.

Hard rules

• Same-vector compare is a localization tool, not a submission gate.
• Never claim BEAT, a clean route, or a submit-ready package from route-compare alone.
• Promotion still requires fresh build, full 9024 trusted/residual zero classical/phase/ancilla garbage, then official benchmark score below frontier.
• If the baseline is dirty under fixed vectors, report the compare as relative evidence only: it can falsify a blamed call but cannot prove the candidate clean.
• Rebuild ops.bin before every residual/eval screen; do not trust stale streams.

When to use

Use when any of these are true:

• A per-call env gate appears clean for some calls and dirty for adjacent calls.
• residual_probe changed after adding/removing only a few ops.
• A local toy proof passes but full-circuit residual gets mixed classical/phase dirt.
• The candidate changes op hash, op count, or tail nonce placement.
• You need to decide whether to localize inside the edit or first clean the baseline.

Probe pattern

Create or reuse a diagnostic binary that:

1. Builds baseline ops and candidate ops in the same source tree.
2. Clears only the candidate env knobs for the baseline build.
3. Sets exactly the candidate env knobs for the candidate build.
4. Derives one deterministic input vector set independent of candidate hash, or deliberately uses baseline_fs / candidate_fs modes and labels the result.
5. Runs both routes on the same point-add inputs.
6. Reports:
   • baseline classical/phase/ancilla failures,
   • candidate classical/phase/ancilla failures,
   • output diff count between baseline and candidate,
   • phase diff batches,
   • first useful divergence mask.

Recommended command shape:

ROUTE_COMPARE_SHOTS=512 \
ROUTE_COMPARE_CANDIDATE_ENV='TLM_FFG_CY0_PARK_CALLS=182' \
cargo run --release --bin route_compare_probe

For values containing commas, separate env edits with semicolons:

ROUTE_COMPARE_CANDIDATE_ENV='TLM_FFG_CY0_PARK_CALLS=178,180,181;TLM_FOLD_CALL_CODE_OVERRIDES=248:5,265:5' \
ROUTE_COMPARE_SHOTS=2048 \
cargo run --release --bin route_compare_probe

Interpretation

• output_diff=0 and candidate failures match baseline failures:
  • The edit is not locally blamed by this sample.
  • If residual was dirty, suspect op-hash sample drift, baseline dirt, or later composition.
• Candidate has new classical failures vs a clean baseline:
  • Value bug. Localize inside the edit before widening.
• Candidate has phase-only failures:
  • Classify phase source before nonce search. Do not submit.
• Candidate has ancilla-only failures:
  • Lifetime/reclaim bug. Check host was truly zero at loan and restored at reclaim.
• Baseline is dirty:
  • Move the probe to a cleaner checkout or treat the finding as falsification only.

Report line

## HH:MM route-compare localization
Candidate=<name> baseline=<tree/hash> candidate_env=<env>.
Same-vector compare=<shots> base=<c/p/a> cand=<c/p/a> output_diff=<n> phase_diff_batches=<n>.
Interpretation=<local-edit-falsified|value-bug|phase-bug|baseline-dirty>.
Gate=no winner/no submit; next=<bounded action>.

Park conditions

• Same-vector compare is dirty and the failing route has no clean local toy proof.
• The baseline itself is dirty and no cleaner baseline can be produced quickly.
• The route saves peak qubits but Toffoli headroom cannot beat the refreshed frontier.
• A 9024 trusted/residual run is not zero classical/phase/ancilla.