a ixj(@sdZddlZddlZddlZddlZddlZddlZddlmZej ddddl m Z edZ edZedZd d Zd d Zd dZddZddZedkreedS)aOne-shot reminder: review xhigh, and re-check a Poly-US-only rule. Scheduled for 2026-09-13 (about five weeks after the 2026-08-09 freeze). Emails a live summary read straight out of the forward-test log, then the Windows task that fires it is a one-off, so nothing recurs. WHY A SCHEDULED TASK AND NOT A CHAT REMINDER: Claude's own cron jobs are session-only and expire after 7 days, so they cannot carry five weeks. This runs off Task Scheduler on the box, which survives reboots and closed sessions. WHAT TO CHECK, AND THE TRAPS * xhigh was frozen 2026-08-09 (Kalshi-only, mid>=0.20, CLV in [0.05,0.12)). It needs ~480 bets. Anything below that is directional only. * Its discovery numbers (+15.1%, z+1.33) came from post-2026-07-05 data that was ALREADY the xlow forward-test window -- selection, not validation. Only fixtures after the freeze count. Expect the effect to shrink. * Poly-US-only was swept on 2026-08-09 and NOT frozen: the price-floor effect reproduced (+16.1% at minP>=0.20) but train and test disagreed in sign, and persistence hurt. Deliberately deferred to avoid a second frozen rule doubling the multiple-comparison exposure. * xhigh and xlow OVERLAP -- both fire on Kalshi CLV in the same band. If both look good, that is largely the same bets, not two pieces of evidence. * Better than freezing a Poly-US rule: split xhigh's OWN realised bets by venue. That answers "is the edge venue-dependent" for free. N)PathzE:\soccer\main) send_emailz3E:\soccer\research\forward_test_unique_book_log.csvz(E:\soccer\research\pin_fresh_capture.csvz*E:\soccer\research\pin_accuracy_cache.jsoncCsJz.ddtttddD}|r*|dniWStyDiYS0dS)NcSsg|]}|qSr).0rrr'E:\soccer\main\reminder_xhigh_review.py ,zlatest..utf-8encoding)csv DictReaderopenLOG Exception)rowsrrrlatest*s  rc Cs(z t|WSttfy"YdS0dS)N)float TypeError ValueError)xrrr_f2s rc Csi}ts|Sttttdd}t||d<tdd}|D]`}|d|d|d|d f}| d r||d  |d | d rD||d  |d qDdd| D}t|pd}t||d<t dd| D|d<t dd| D|d<t dd| D|d<t dd| D|d<d|d||d<tt}|D]}dD]z}t| d|t| d|} } | d usb| d urqb| d!krd"n| d#krd$nd%} || | | qbqZd&D],} || rdt|| |d'| <qzttjdd} WntyBi} Yn0i} |D]X}|dd(|dd(|dd(|d }|| vs|d)| |d)krL|| |<qLg}| D]\}}| |}|sΐqt|d*t|d+}}t|d,t|d-}}d ||||fvrqd.|d/kd.|d0k}}|||d1||d1||d1||d1qt|d2krt|}t|pd3}|tt|}|jt||||||d4||d4|d5|S)6aLive read of the pinnapi-vs-API-Football comparison. Three separate questions, deliberately reported apart because they answered differently as of 2026-08-09 and conflating them caused real confusion: ACCURACY paired Brier per fixture. The only clean test -- paired, uses every fixture, no entry reconstruction. Needed ~580 fixtures. It FLIPPED SIGN (+0.0096 -> -0.0078) when n went 26 -> 77, which is why nothing below target should be believed. DIVERGENCE how often the two feeds produce a different bet. This is NOT evidence either is better -- only that the choice matters. BIAS API-Football rounds odds to 2dp (1.50 where pinnapi says 1.5917). Measured as a monotonic gradient: it underprices favourites and overprices long shots. This is a MECHANISM and was stable across sample sizes. r r Z snapshotscSsttdS)Nsf)setrrrrRr z!pinnapi_summary..ZlchomeZawayZ game_dateZ xlow_stalerZ xlow_freshrcSs&i|]\}}|ds|dr||qSrr)rkvrrr Yr z#pinnapi_summary.. sig_fixturescss*|]"}|dr|d|dkrdVqdSrrr#Nrrr!rrr \r z"pinnapi_summary..samecss"|]}|dr|dsdVqdSr%rr&rrrr']r af_onlycss"|]}|dr|dsdVqdS)rrr#Nrr&rrrr'^r pin_onlycss2|]*}|dr|dr|d|dkrdVqdSr%rr&rrrr'_s differgY@ agree_pct)haZ fair_fresh_Z fair_stale_Ng?favg?middogr/r0r1bias_|ZtsZ fair_fresh_hZ fair_fresh_aZ fair_stale_hZ fair_stale_ag?HA g& .>g\(\?)acc_nacc_meanacc_seacc_zacc_loacc_hi)CAPTUREexistslistrrrlen collections defaultdictgetadditemssumvaluesrappend statisticsZmeanjsonloads ACC_CACHE read_textrZpstdevmathZsqrtupdate)outrZfxrr ZsignZsgtfrstbcacheZlastZdiffsZftrZfhZfaZshZsaZyhZyamZsdZserrrpinnapi_summary9sz      *    *     rZcs4sdSdvrddkr dnddkr0dnd}dd krDd nd d dd }d ddddddddddddddddd||d}nd}dfdddD}d |d!d"dd#d$dd%d&dd'd(d)dd*d+dd,d-dd.|d/S)0Nz=

pinnapi vs API-Football

No capture data found.

r9r=rzpinnapi more accurater>z!API-Football more accuratezno measurable differenceiDz — still z1 short of the 580 target, so treat as directionalz!
  • Accuracy (paired Brier, n=z): r:z+.5fz (SE r;z.5fz, z r<+.2fz ), 95% CI [z, z ] → z
    • z=
  • Accuracy: too few settled fixtures to report.
    • z / c3s@|]8}d|vr|dd|tdddVqdS)r3 nanr\cN)rEr)rrWprrr'sz!build_pin_html..r2z)

    pinnapi vs API-Football

      z
    • Signal divergence (r$z7 fixtures that produced a signal): identical r(z (r,z.0fz%), API-Football only r)z, pinnapi only r*z%, both fired but DIFFERENT bet r+z. Divergence shows the feed choice matters; it is not evidence either is better.
    • Rounding bias (API-Football minus pinnapi): a. A monotonic gradient = API-Football's 2dp rounding underpricing favourites and overpricing long shots. This was stable across sample sizes — the one solid finding of the exercise.

    Traps. The accuracy figure flipped sign (+0.0096 → −0.0078) when n went 26→77, so anything below target is noise. And a positive result justifies recalibrating the xlow band against the fresh line and re-testing — NOT switching the live feed, which stays on API-Football so the frozen forward tests keep running. Given the rounding gradient, a band tuned to a biased anchor will not transfer to an unbiased one.

    Re-run: python research\pin_accuracy_test.py and python research\pin_fresh_vs_stale.py

    )joinrE)raZverdicttargetZaccZbiasrr`rbuild_pin_htmls\         rdcCst}|dpd}tt}d|ddd|d|dd d |d d d |d d d|dd d|dd d|dd d|d}td|}t|rdnd|rdSdS)NZxhigh_n0z

    Reminder: review xhigh (and revisit Poly-US)

    Frozen 2026-08-09. Set as a five-week checkpoint.

    Where xhigh stands (from the forward-test log, run Zrun_date?z)

    • n = z< of the ~480 needed to resolve the effect
    • WR Zxhigh_wr-z %   ROI Z xhigh_roiz%   z Zxhigh_zz#
    • xlow for comparison: n=Zxlow_nz ROI Zxlow_roiz % z Zxlow_zaw

    Read it with these in mind

    • xhigh's discovery numbers (+15.1%) came from data that was already the xlow forward-test window. That was selection, not validation — expect the real effect to be smaller.
    • The reason it was frozen was measurability, not ROI: a price floor cut the required sample from ~201,000 bets to ~480. 17 candidates were swept and none reached z>1.96.
    • xhigh overlaps xlow. Both fire on Kalshi CLV in the same band. Two green numbers here are largely the same bets.
    • If n is still well short of 480, say so and wait. Do not read a partial sample as a verdict.

    The Poly-US question that was deferred

    Swept 2026-08-09, not frozen: the price floor reproduced (+16.1% at minP≥0.20, needing only ~370 bets) but train and test disagreed in sign and persistence hurt every variant. The cheaper next step is to split xhigh's own realised bets by venue rather than freeze a second rule.

    Poly-US also pays a maker rebate where Kalshi charges a taker fee. If the floor effect survives, that is an execution reason to prefer it — a separate question from whether the signal is better.

    Re-run: python research\rule_sweep.py and python research\forward_test_unique_book.py

     z.[REMINDER] Review xhigh + revisit Poly-US ruleZsentz send failedrr#)rrErdrZrprint)rrSZpin_htmlbodyZokrrrmains2        )+ rk__main__)__doc__rCrrLrPrKsyspathlibrpathinsertZ daily_signalsrrr?rNrrrZrdrk__name__exitrrrrs&  V,4