allocation.POOL_TGT (TE x0.9007, WR x1.0503) was calibrated on the preseason board path and D-250 found it wrong in-season. At HEAD, after G-SRCWK (LATEST_SKIP_FINAL_WEEKS=2), is it still right on the board path, still wrong on the walk-forward (in-season) path, and which way does each endpoint call it?

Owner lane: serving

Question

allocation.POOL_TGT (TE x0.9007, WR x1.0503) was calibrated on the preseason board path and D-250 found it wrong in-season. At HEAD, after G-SRCWK (LATEST_SKIP_FINAL_WEEKS=2), is it still right on the board path, still wrong on the walk-forward (in-season) path, and which way does each endpoint call it?

Prediction

Frozen in prereg/2026-09-23T1008Z-h174-pool-tgt-both-paths.md: WF ON TE targets share ~-9%, OFF ~+1%; TE points MAE OFF-ON in [-0.10, 0.00]; all-rows MAE within +-0.01. BOARD ON TE targets within +-3% of truth, OFF ~+8%. Falsifier of H-174 as stated: WF ON TE bias inside +-3%, or OFF worsens TE points MAE with an interval excluding 0.

Finding and verdict

TESTED -- PATH-DEPENDENT. H-174's claim that the error POOL_TGT corrects 'is not there' is TRUE on the in-season path (TE targets share OFF +1.1% [-0.6,+2.7], ON -9.3% [-10.9,-7.8]) and FALSE on the board path at HEAD (OFF +10.9% [+4.1,+18.0], ON -0.6%). By point MAE ON wins TE on both paths (WF +0.081, board +0.140 OFF-ON) and loses WR (-0.052, -0.037); all-rows WF nil (-0.0008), board +0.018 [95 +0.004,+0.034]. The TE MAE win fails a shuffled-label control (+0.217 under noise). Run status per frozen controls: VACUOUS (C2 bar), diagnosed post hoc as a zero-line counting error in the bar.

Reasoning

Kept the held claim: the run was complete and unrecorded; recording it closes the claim without spending a fit, and the path split is a serving-layer question (what the board has vs what the backtest has).

Next action

(a) Find what latest_rows feeds the TE allocation that rows_for does not -- the +10.9% board over-allocation is the actual defect the constant hides. (b) Board TE rec_tds +47% is a separate board-path error worth its own row. (c) Rerun C2 with the restated bar only if a rerun is otherwise needed.