Review of 2026-09-23T1240Z-h174-pool-tgt-both-paths

Owner lane: Not recorded

Reproduction

REFUSED. Verified byte-for-byte: source record 63eef5b8..., protocol 32acff06..., runner f5087b18..., result out.json d88912bb.... Not bound or changed: (1) input_artifacts is [] and the as-run snapshot h174b.duckdb is not on disk (attempt 3's byte-cmp to the live warehouse is a log claim, not a retained artifact); (2) the board arm reads data/injuries_live.jsonl at run time, an input no artifact binds; (3) HEAD is now 675864ee, 70 playerweek/ files differ from the recorded bb8f7057 (allocation.py, playermodel.py, share.py hashes still match the recorded ones; arms.py does not: 5bb12400 changed early_season only, outside rows_for); (4) warehouse has since ingested 2026 week 3. Consequently the non-exact reproduction cannot separate warehouse change, code change and overlay change for the board shift; only the overlay's own effect (live vs off, same snapshot, same code) is isolated.

Finding and verdict

WEAKENED

Review scope

Two halves with opposite outcomes. (A) The WALK-FORWARD half REPRODUCES (non-exact) and is UPHELD: TE targets share ON -9.34% [95 -10.90,-7.83] (record -9.32), OFF +1.06% [-0.66,+2.69] (record +1.08); WR targets ON +4.33% (record +4.33); TE points MAE OFF-ON +0.0810 [99.6 +0.0626,+0.1011] (record +0.0813); all-rows -0.0009 (record -0.0008); shuffled-label TE OFF-ON +0.2169 (record +0.2170). So on the in-season path the constant shaves ~9% off every TE's target share and adds ~4% to WR's, and the TE points-MAE preference for ON is the same MAE-vs-mean property the record already says it is. (B) The BOARD half does NOT reproduce and the record's headline 'PATH-DEPENDENT: FALSE on the board path at HEAD (OFF +10.9% [+4.1,+18.0], ON -0.6%)' is WEAKENED to unestablished: on a fresh snapshot with the frozen runner text unchanged the board arm gives TE targets ON -5.77% [-14.98,+4.44], OFF +4.73% [-4.98,+15.58] (record ON -0.58, OFF +10.89 [+4.10,+17.99]); WR targets ON flips sign +2.24 -> -4.01; TE rec_tds ON +47.0 -> +38.2; TE points MAE OFF-ON +0.1405 -> +0.1119. Cause found by exploratory attack (attack 2): the board rig calls PM.latest_rows(con) with live_designations=True, which overlays the LIVE data/injuries_live.jsonl (a file outside the snapshot, unbound) onto the historical board men; with CURRENT_SEASON pinned to L=2025/2024 it applied 102 of 120 and 80 of 91 designations to the 2025 and 2024 board men. Switching only that overlay off (live_designations=False, applied 0) moves the board TE targets to ON -8.44% [-16.22,-0.15], OFF +1.98% [-6.34,+10.77] -- the walk-forward numbers (-9.3 / +1.1), inside the board interval. So the board arm's absolute share bias is a function of an unpinned run-date file, and with that pinned off the board and in-season paths are not separable at this n (2 seasons, 64 team clusters). The record's mechanism ('the board and in-season paths hand the allocation different TE inputs; the constant cancels an 11% board over-allocation') is not supported: the 11% is not stable across runs, and the run-to-run instability is a stated, measured input, not football. serving_consequence half survives: after the roll to the in-season path the WF measurement (reproduced) says the constant biases TE targets ~-9% and WR ~+4%.

Next action

If the board-path question is still wanted: pin live_designations=False (or a dated jsonl copy bound as an input) in the board rig, refit on 3+ board seasons (L=2022-2025), report the board-minus-WF TE share difference with its own interval, and bind snapshot + jsonl as input_artifacts. The WF finding needs no further measurement.