project_arm_promotion_bar_corrected

The promotion bar recorded in project_model_persona_matrix — "200+ priced picks with units AND positive CLV → register as a served tier" — is not a valid bar and will promote noise. Measured 2026-08-28 on 27,018 graded WNBA picks (2026-05-01 → 08-28, pulled per-day from /clv/picks; note bet_odds is AMERICAN, not decimal).

Three corrections the bar misses:

1. The baseline is +1.79%, not zero. Every graded WNBA pick anyone made this season returned +1.79% ROI. Resampling a formula's picks from that global pool at the same price bucket is the honest null — it needs no vig assumption and holds price constant, so it isolates selection. "Positive units" clears zero, not +1.79%.
2. 55 formula-arms have been measured (27 with n≥200). Bonferroni at 0.05 is p<0.0019. Zero arms clear it. A best-of-N winner is what a 27-way search produces from noise.
3. Positive CLV is not evidence. Construction bias is ~+1.1¢ (entry at best-of-books, close from Pinnacle-first). Everything sits 0.93–1.39¢. See project_two_month_retro_2026_08.

Measured (ROI / same-price-random / lift / permutation p):

form_wnba is not a second model — do not propose promoting it. Its own header: "props ranked by the gap between recent form (L10 average) and the line, price-agnostic" — i.e. the line_edges z-score minus the price gate and the suppressors. Generate() returns empty; picks are POSTed by sport-agents/core/lens_screen.py purely so the board's Form tab gets its own ledger. Its lift is +2.4pp at p=0.108 and its ROI decays 11.3% → 5.3% → 3.9% (May-Jun / Jul / Aug).

The deepseek arm reads the picker's own output. model_matrix.py:234 persona prompt: market_value : both_agree — line_edge_quality AND de-vig point the same side. 43% of its picks are identical to form_wnba's (same date/player/prop/line/side). Same signal, second opinion attached. CLV 1.06¢ [0.83, 1.29] = no closing-line edge.

The fade arm is the control that kills the whole family: it is instructed to bet AGAINST recent form and returns +4.5%, indistinguishable from the projection arm's +4.2%. When both directions pay the same, direction is not what pays.

Corrected bar — all four required before serving any arm:

1. Lift over same-price random resampling, not over zero.
2. p below the multiplicity-corrected threshold for every arm measured.
3. CLV above the ~1.1¢ construction bias, with slate-clustered CIs.
4. Pre-registered forward test (formula, league, start date, sample target, decision rule written down BEFORE the window opens) — a retrospective search cannot promote itself.

Only live candidate: deepseek-v4__market_value — largest n, biggest lift, only sub-0.01 p. Forward-test it, do not deploy it. Its deterministic twin both_agree (Go, no LLM cost) tests the same hypothesis and sits at n=27 — raise its volume so it has a real sample by NBA.

Analysis scripts: session scratchpad perm.py / pull3.py. Related: feedback_model_edge_calibration, project_decision_layer_audit, project_brier_vs_vegas_closed.