Dashboard โ€บ Three-Axis Probes

the combiner ยท operational

๐Ÿšช The Gate

One promote/skip decision over the three probes. A tunable knob vetoes. Not a fourth axis โ€” it measures nothing itself. Three probes, one gate.

Cost-ladder execution

1. parameterless  (static, FREE โ€” no ClickHouse)   โ”€โ”€ FAIL โ”€โ–ถ STOP (tunable-knob veto)
        โ”‚ PASS
2. agnostic       (static no-dispatch + oracle)     โ”€โ”€ FAIL โ”€โ–ถ STOP
        โ”‚ PASS / PENDING
3. orthogonal     (full bigblack grid, EXPENSIVE)   โ”€โ”€ consulted only now
        โ”‚
   โ–ถ promote = parameterless.PASS AND agnostic.PASS AND orthogonal.PASS
2026-06-12 โ€” the gate now speaks CAMPAIGN. The orthogonal axis accepts two evidence dialects: the shared_data probe summary (min_regime_pass_rate) AND campaign/spike per-cell records via campaign_summary_adapter.py (worst-valid-cell over โ€–ฯโ€– + panel-Rยฒ + h_norm, stored labels re-derived from the raw numbers โ€” zero disagreements on the 2026-05-27 spike). The earlier PENDING was a dialect gap, not an evidence gap. Correction (same day, operator-caught): the first adapter version SKIPPED the cross-candidate (sibling/Finish) records and briefly showed PROMOTE for deferred candidates โ€” the sibling stage is now mandatory, with asymmetric demotion (promoted features won their disposition; deferred siblings are demoted by links to them).

Live demo verdicts

candidate implparamagnosticorthogonaldecision
bar_petrosian_fd ยท bar_katz_fd (in production, #509)PASSPASSPASSPROMOTE โ€” matches the real disposition (208 cells ร— 7 epochs; siblings resolved in their favor)
bar_svd_entropyPASSPASSFAILBLOCKED โ€” sibling stage: โ€–ฯโ€– 0.9745 vs bar_petrosian_fd (now in panel) โ€” the deferral reason, mechanically reproduced
bar_hjorth_mobility / _complexityPASSPASSFAILBLOCKED โ€” mutual duplicates (โ€–ฯโ€– 0.9985) + 0.92โ€“0.93 vs katz (now in panel)
add_antropy (hardcoded order=3)FAILSKIPPEDSKIPPEDBLOCKED โ€” tunable knob vetoes
antropy.higuchi_fd (hidden kmax=10 default)FAILSKIPPEDSKIPPEDBLOCKED โ€” introspection teeth-check (hidden library default caught)

Verdict strength (honest)

verdictstrengthwhy
BLOCKEDvery strongevery blocking path is a hard citable fact (a literal at file:line, a dispatch branch, a ฯ>0.95 partner)
PENDINGby designnames EXACTLY which evidence is still owed โ€” never inflates confidence
PROMOTEstrong, conditional"passed everything we can currently measure" โ€” known soft spots (orthogonality S4 circularity, agnostic oracle coverage) are surfaced, not hidden

Design rationale: PROBE-ARCHITECTURE.md

โ† AgnosticForward โ†’