mtgarb · buy globally → sell AU

v1 vs v2 — side-by-sideOperator gut-check. v1 = existing engine (point-estimate ROI rank); v2 = inference rebuild (tiered: deterministic buylist + Bayesian regression log-price posterior). Where they disagree most. Read large absolute deltas as "v2 thinks v1's $X estimate is off by Δ". Shadow-window discipline retired 2026-05-16 (operator amendment) — p50 + Δ visible again; width column kept as calibration health metric.
game:allmagicpokemonsort:abs Δv2 P(profit)Tier 1 firstv2 E[U]
Auditslatest per (agent, slice). Operator-evidence only — engine + portfolio do not read.
agentslicerecommendationfindingsn_predn_pairedconfwhen
comparable-sales-enumeratorprediction 1aaccd4dPROCEED_DEEP_DIVE121200.8211h agoSummary:
6 high-similarity (>=0.85) AU sold comparables for Fifth Dawn #114 nonfoil over 180d, ranging A$24.18 to A$47.12. Top-5 by similarity median ~A$45.53 (range A$24.18-A$47.12). Two LP-condition exact matches around A$24-30 (881d50a9, bda00fe0), three LP/MP exact matches A$45-47. Operator landed cost A$31.63 sits between the soft-floor LP cluster and the harder MP/no-#-but-named-set cluster. Cross-printing references cluster A$36-52 and support a A$40+ ceiling. EU CM EUR 16.57 (~A$27.34 floor), TCG USD 30.67 (~A$47.54 ceiling). 17 cross-printing rows and 1 SLD foil / 1 signed / 1 proxy excluded - verdict: deep dive justified.
anomaly-watchermagic tier=0 1dINVESTIGATE_AND_DERATE(high)3000.9011h agoSummary:
RED. Three HIGH-severity findings. (1) eBay AU listings 20% of 7d baseline (58 vs 294) - likely scraper outage or missing EBAY_APP_ID/CERT on VPS. (2)(3) Pokemon foil cm_eur 100% NULL for pre-2014 AND 2014-2020 eras (n=243 combined predictions) - matches documented foil-column inversion bug, landed costs understated 3-5x. Operator next: SSH to VPS to verify EBAY_APP_ID/CERT and re-run eBay scrape; concurrently derate or suppress Pokemon foil picks in pre-2020 eras until column-mapping fix lands.
calibration-auditormagic tier=3 90dINVESTIGATE_AND_DERATE(high)74413225740.6511h agoSummary:
Model exhibits two simultaneous systematic biases: (a) severe undershoot in the price_tier:300_plus slice (median +158% residual, coverage 22.7%) where the model both under-predicts AND has bands far too narrow — high-confidence evidence of pairing/cross-printing contamination in the source data, mirrored by the Mystic Remora alternating-realised pattern in recent_samples; (b) consistent overshoot across modern_2003_2014, recent_2015_2020, and sub-A$100 price tiers (-27% to -55% median residuals) suggesting expensive comparables are leaking into cheap-printing predictions. The prob_profit calibration line is non-monotonic (decile [0.2-0.3) collapses to 4.6% actual vs 26.8% predicted) and unsuitable for position sizing. Vintage_pre_2003 (n=10,032) is near-balanced (-8.3%) and the dominant slice masking issues in aggregate. Recommend INVESTIGATE_AND_DERATE: investigate pairing/printing-resolution in the realised-outcome join (cross-printing leak is the leading hypothesis per CLAUDE.md priors), and operator-side derate Tier 3 picks in price_tier:300_plus and the modern_2003_2014 era until investigation completes. No retraining or prior changes per pre-reg shadow integrity.
No scan_run selected.

agent	slice	recommendation	findings	n_pred	n_paired	conf	when
comparable-sales-enumerator`prediction 1aaccd4d`PROCEED_DEEP_DIVE121200.8211h ago Summary: 6 high-similarity (>=0.85) AU sold comparables for Fifth Dawn #114 nonfoil over 180d, ranging A$24.18 to A$47.12. Top-5 by similarity median ~A$45.53 (range A$24.18-A$47.12). Two LP-condition exact matches around A$24-30 (881d50a9, bda00fe0), three LP/MP exact matches A$45-47. Operator landed cost A$31.63 sits between the soft-floor LP cluster and the harder MP/no-#-but-named-set cluster. Cross-printing references cluster A$36-52 and support a A$40+ ceiling. EU CM EUR 16.57 (~A$27.34 floor), TCG USD 30.67 (~A$47.54 ceiling). 17 cross-printing rows and 1 SLD foil / 1 signed / 1 proxy excluded - verdict: deep dive justified.
anomaly-watcher`magic tier=0 1d`INVESTIGATE_AND_DERATE(high)3000.9011h ago Summary: RED. Three HIGH-severity findings. (1) eBay AU listings 20% of 7d baseline (58 vs 294) - likely scraper outage or missing EBAY_APP_ID/CERT on VPS. (2)(3) Pokemon foil cm_eur 100% NULL for pre-2014 AND 2014-2020 eras (n=243 combined predictions) - matches documented foil-column inversion bug, landed costs understated 3-5x. Operator next: SSH to VPS to verify EBAY_APP_ID/CERT and re-run eBay scrape; concurrently derate or suppress Pokemon foil picks in pre-2020 eras until column-mapping fix lands.
calibration-auditor`magic tier=3 90d`INVESTIGATE_AND_DERATE(high)74413225740.6511h ago Summary: Model exhibits two simultaneous systematic biases: (a) severe undershoot in the price_tier:300_plus slice (median +158% residual, coverage 22.7%) where the model both under-predicts AND has bands far too narrow — high-confidence evidence of pairing/cross-printing contamination in the source data, mirrored by the Mystic Remora alternating-realised pattern in recent_samples; (b) consistent overshoot across modern_2003_2014, recent_2015_2020, and sub-A$100 price tiers (-27% to -55% median residuals) suggesting expensive comparables are leaking into cheap-printing predictions. The prob_profit calibration line is non-monotonic (decile [0.2-0.3) collapses to 4.6% actual vs 26.8% predicted) and unsuitable for position sizing. Vintage_pre_2003 (n=10,032) is near-balanced (-8.3%) and the dominant slice masking issues in aggregate. Recommend INVESTIGATE_AND_DERATE: investigate pairing/printing-resolution in the realised-outcome join (cross-printing leak is the leading hypothesis per CLAUDE.md priors), and operator-side derate Tier 3 picks in price_tier:300_plus and the modern_2003_2014 era until investigation completes. No retraining or prior changes per pre-reg shadow integrity.