Iteration 7 — Q6: can TAU=0.85 be derived, or must it stay policy? RESOLVED — policy, not derived

Read-only evaluation · no ClickHouse query needed — synthesised from the committed SPIKE-2026-06-16 replication (read-only --readonly=2 access present) · matrix-evaluation-for-orthogonality loop · PR #543

Question (backlog Q6)

TAU = 0.85 is the redundancy cutoff on the Spearman max|ρ| "orthogonal-today" metric (≤0.85 orthogonal · 0.85–0.95 WATCH · >0.95 BAN). Is 0.85 a derived, calibrated number, or is it policy? Required trust basis: a ground-truth anchor (the only basis that can back a redundancy cutoff).

Method (read-only, zero new compute)

Synthesised from the committed SPIKE-2026-06-16-RECONFIRMATION.md §4 (the empirical redundancy distribution on fresh crypto data) + STATUS-GROUNDING-AND-THRESHOLDS.md (threshold trust-basis ledger) + calibrate_prcp.py (the matched-null derivation harness). No new measurement → leakage_guard.py N/A.

Findings — every claim with its trust basis

Verdict

RESOLVED — TAU=0.85 cannot be derived from the data as it stands; it stays POLICY.

• 0.85 WATCH band = un-nulled policy → treat it as a transparent hold / re-probe disposition, not a calibrated verdict. It is, in the campaign's words, "the fuzziest number in the stack."
• Keep the 0.95 BAN — it is ground-truth-anchored and stands.
• The only path to a principled cutoff is a ground-truth anchor, specifically an information-loss-budget method: define the maximum acceptable information loss from dropping a redundant feature, then derive the cutoff from that budget — not from a density void (which does not exist on this scale). Building that derivation is out of the evaluator's write-surfaces (R3) → operator.

Anti-inference guard (R7)

This does not derive a new TAU and does not change the 0.95 BAN. It certifies that 0.85 is policy (un-nulled) and names the only derivation that would make it principled. No metric is promoted to a gate; the Spearman max|ρ| policy lines are unchanged and remain distinct from any descriptive Pearson values reported elsewhere.

Read-only & resource-cap proof (this firing)

What's next

• Operator action (only if a hard derived TAU is required): implement the information-loss-budget derivation — pick a tolerated information-loss level, derive the redundancy cutoff that respects it, and replace the policy 0.85 with that derived value. Until then, 0.85 remains an explicit policy hold and the 0.95 BAN remains the only anchored line.

SSoT: SPIKE-2026-06-16-RECONFIRMATION.md §4, STATUS-GROUNDING-AND-THRESHOLDS.md, calibrate_prcp.py · audit twin 2026-05-26-forward-orthogonality-prediction/ · read-only evaluator · do NOT merge PR #543