Grading Record — September 2026 Releases
Predictions frozen in: research/predictions-2026-09.md (committed 2026-08-27; rows never edited after release). Graded: 2026-09-19. Rows #20–24 by hand against the AMS Colorado Direct Hay Report dated Thu Sep 10, 2026 (live PDF, sha256 be950a2b…, read by pdftotext and visually; data/2026-09-19/ams_2905_2026-09-10.pdf); rows #25–27 against the CPC ENSO Diagnostic Discussion of 10 Sep 2026 and wksst9120.for. Every hand grade was then verified by grade_predictions.py --check (17 mechanically checkable rows match the frozen protocol). Rows #17–19 (NASS August prices) grade by script on Sep 29 (NASS calendar; not Sep 30 as the predictions file assumed).
AMS Colorado Direct Hay Report, Sep 10, 2026
| # | Predicted | Observed | Source line (verbatim) | Outcome |
|---|---|---|---|---|
| 20 | SE CO FOB-farm per-ton alfalfa trade prints again (YES) | NO | Southeast section contains exactly two rows: Alfalfa - (Ask/Per Bale) Small Square Bundled 11.50 F.O.B.-Stables Covered 1st Cut and Prairie/Meadow Grass - Premium (Trade/Per Bale) Small Square 1400 11.75 11.75 F.O.B.-Retail Covered 1st Cut. No per-ton alfalfa trade, ask, or FOB-farm line of any kind. | MISS |
| 21 | SE FOB-farm wtd avg $315 / $250–380 (conditional on #20) | — | condition not met | NO-PRINT |
| 22 | NE dairy per-ton alfalfa wtd avg $330 / $265–395 (if prints) | none | Only NE per-ton alfalfa trade: Alfalfa - Good (Trade/Per Ton) Large Square 3x4 2000 260.00 260.00 F.O.B.-Farm/Ranch New Crop. The 3000 … 315.00 … F.O.B.-Dairy row is in the San Luis Valley, Hay (Organic) section, not Northeast. | NO-PRINT |
| 23 | Volume 2,500 t / 800–4,200 | 8,500 t | Tons: Hay 8500 1338 1488 (This Week / Last Reported (8/24/2026) / Last Year). Composition: 3,000 t SLV organic alfalfa contract + 1,000 t SLV organic grass + 500 t SLV organic alf/grass + 2,000 t NE alfalfa + 2,000 t Mtn/NW grass. Row sums check to 8,500. | MISS (high; NEAR cutoff 4,410; 3.4× point) |
| 24 | Demand at least "Good" | "Good to Very Good" | Demand Good to Very Good all markets of new crop hay. | HIT |
Single-target note (audit commitment): the grading target — a completed SE Colorado FOB-farm commercial alfalfa trade ≥100 t — did not print. September is UNOBSERVABLE on the primary target unless the Sep 24 report prints one. The nearest large commercial print in Colorado is the NE-region 2,000 t of Good 3x4 at $260 FOB-farm — 23× the tonnage of the 88-t $325 lot the August 27 report was built around, and $65 under it.
CPC ENSO, Sep 10, 2026
| # | Predicted | Observed | Source line | Outcome |
|---|---|---|---|---|
| 25 | Advisory continues | El Niño Advisory | ENSO Alert System Status: El Niño Advisory (discussion line 8; PDF p.1 read visually) | HIT |
| 26 | Weekly traditional Niño-3.4 at discussion time: +2.7 / +1.4 to +4.0 | +2.7 | wksst9120.for line 02SEP2026 … 29.4 2.7 — the newest line on Thu Sep 10 (weeks are Sun–Sat centered on Wednesday; the 09SEP line, +2.9, posted Mon Sep 14 and is also inside the range). The discussion's "+1.8°C in Niño-3.4" is the monthly relative index (Fig. 2 caption) and the Sep 14 deck's "2.0°C" is the weekly relative value — neither is the frozen metric. | HIT (point exact) |
| 27 | Historic-event probability 69 / 59–79 | 75 | During the October-December 2026 season, there is a 75% chance of a historic event that would exceed the strength of previous El Niño events dating back to 1950 (+2.5°C or more for a 3-month RONI value). (discussion; repeated on deck slide 25, Sep 14) | HIT (+6 vs point) |
Summary (from scripts/scorecard.py, 2026-09-19 — never hand-typed)
- Interval coverage: 6 of 16 two-sided numeric ranges contained the outcome (+2 NEAR). Standing record across both graded cycles.
- Cluster A (economically central): 5 graded — 1 HIT (#14), 4 MISS (#P1, #1, #9, #20); 1 NO-PRINT (#21); #17 pending. The economically central rows are still 1-for-5. Both September Cluster-A misses were the same failure: we predicted the target would be observable and it wasn't.
- Cluster B: 1 NEAR (#16), 1 MISS (#2), 3 NO-PRINT; #19 pending.
- Cluster C (regime checks): 17 graded — 10 HIT, 1 NEAR, 6 MISS, 1 NO-PRINT; #18 pending. The ENSO rows went 3-for-3 this cycle with the metric pinned in advance — the erratum machinery (L8) worked.
- Mean absolute point error: price rows $24.80 (n=5, unchanged — no price row resolved this cycle); volume rows 3,847 t (n=2: −1,694 then +6,000 — the AMS weekly volume is not forecastable at this width and should be dropped from future sheets or made categorical); probability rows 9.7 pts (n=3).
- Judgment layer vs anchor (first readable cycle): #26 anchor +2.6 → judgment +2.7 → observed +2.7 (judgment helped, +0.1); #27 anchor 69 → judgment 69 → observed 75 (no adjustment made; miss +6 but inside range); #23 anchor 1,900 → judgment 2,500 → observed 8,500 (both wrong; judgment direction right, magnitude hopeless). Too few resolved numeric rows to say anything about whether judgment beats the anchor. #17–19 (anchor 210/207/125, judgment 235/230/155) resolve Sep 29 and are the real test.
- Calibration floors (L7): price ±$65, volume ±1,694 t (unchanged); anomaly ±1.3°C; probability ±10 pts. The ratchet rule — floors only fall once two-sided coverage ≥70% on ≥10 rows across ≥2 cycles — is now enforced in
calibration.py(previously caveat text only). Coverage is 38%, so nothing releases.