axis 2 · static · operational
Does this feature contain secret hand-tuned numbers? "No magic numbers, no tuned constants, no if n≥500 dispatch" (msg 428). FREE — no market data — so it runs first and short-circuits before any ClickHouse cost.
AST-scan every literal, estimator kwarg and size-dispatch; classify against constants_registry.toml:
| class | meaning | promotion |
|---|---|---|
derived | forced by math (log K, ξ Var→2/5) | PASS |
project_wide | ONE global value for every feature (the _q1000 precedent) | PASS |
feature_tunable | a knob the candidate picked for itself | BLOCKED |
| inline literal / unregistered | hardcoded order=3 or unknown constant | BLOCKED |
bar_svd_entropy pinned to project-wide SVD_ORDER → PASS. The same feature in add_antropy hardcoding order=3, delay=1 → FAIL. Same math, different discipline — the exact near-promotion this prevents.Accumulates every tested candidate's knobs in candidate_knobs.jsonl (append-only). When ≥4 candidates across families share the same knob at the same value, that's a convention, not a magic number — promotable to a project-wide constant, rescuing every user at once.
| knob | verdict | members / families |
|---|---|---|
ordinal_embedding_m = 3 | REUSE (ORDINAL_M, #513) | 9 / 1 |
ordinal_embedding_tau = 1 | REUSE (ORDINAL_TAU, #513) | 9 / 1 |
mfdfa_detrend_order = 1 | PROMOTE-NEW | 16 / 2 |
takens_embedding_tau = 1 | PROMOTE-NEW | 4 / 1 |
--verdicts issues a concrete Axis-2 verdict for EVERY candidate from its declared knob set — no implementation needed. A FAIL goes to the orthogonality-only, no-promotion track (its Axis-2 fate is already known), so promotion compute is never wasted.
2026-06-12: the CECP trio came back from the 210-cell orthogonality campaign — #19 CECP velocity PASS (max‖ρ‖ 0.666 → PROMOTABLE NOW, implementation loop running) · #20 Binary CECP WATCH (0.937) · #16 Statistical Complexity BAN (0.969 vs bar_petrosian_fd). 2026-06-12 correction (operator challenge upheld): the 8 MF-DFA "CONDITIONAL" verdicts were under-declared — formula research shows 8–16+ handpicked knobs each (q-grid, scale schedule, fit range + variant machinery) → all 8 corrected to FAIL; a fail-closed family-manifest completeness gate now blocks under-declared rows. Evidence: parameterless_inventory_verdicts.json + the 2026-06-10 campaign spoke.
Deep dives: AXIS-2-PARAMETERLESS.md · AXIS-2-PROMOTION-SWEEP.md