Grading Record â Rows #17â19 and #28
Graded: 2026-10-02. Rows #17â19 frozen 2026-08-27 in research/predictions-2026-09.md; row #28 frozen 2026-09-19 in research/predictions-2026-10.md. Rows never edited after release.
NASS Agricultural Prices, released Tue Sep 29, 2026 (AUGUST prices) â graded by script
Source: https://esmis.nal.usda.gov/sites/default/release-files/796084/agpr0926.txt (saved: data/2026-10-02/agpr0926.txt and data/2026-10-02/macro/agpr0926.txt). Table title verified: "Prices Received for Hay by Type - States and United States: August 2026 with Comparisons" â month-pinned, not July. Graded by grade_predictions.py --apply, then the table was read by hand a second time (L3). July 2026 was not revised for any of the seven states or the US.
| # | Predicted | Anchor â judgment | Observed | Source line (verbatim; cols = All hay Aug25/Jul26/Aug26, Alfalfa Aug25/Jul26/Aug26, Other) | Outcome |
|---|---|---|---|---|---|
| 17 | CO alfalfa Aug $235 (170â300) | 210 â 235 | $235 | Colorado .............: 154.00 222.00 231.00 155.00 225.00 235.00 140.00 175.00 180.00 | HIT (exact) |
| 18 | CO all-hay Aug $230 (165â295) | 207 â 230 | $231 | same line | HIT (+1) |
| 19 | KS alfalfa Aug $155 (90â220) | 125 â 155 | $127 | Kansas ...............: 114.00 123.00 123.00 121.00 126.00 127.00 102.00 105.00 105.00 | HIT (inside a wide range; point â$28) |
Judgment layer vs anchor (the test research/17 set up). These rows were frozen Aug 27. The persistence baseline available then was June (210/207/125), which is what the rows record as anchors. Against it, judgment won on Colorado (#17: 0 vs 25; #18: 1 vs 24) and lost on Kansas (#19: 28 vs 2); MAE judgment $9.7 vs anchor $17.0. July prices (225/222/126) were published Aug 31, after the freeze; persistence from them would have scored $6.7, but that is a later-information benchmark, not a fair comparison. Three rows is not a track record. The Kansas bet (that the state average would follow cash up) failed again â the average moved $1 for the fourth straight month â so Kansas row #42 is pure persistence. (A first draft and the pre-publication review called July persistence the "fair baseline" and said judgment lost; a second review caught that July printed after the freeze. Corrected.)
AMS Colorado Direct Hay Report, Thu Sep 24, 2026 â graded by hand
Source: data/2026-10-02/ams/ams_2905_2026-09-24.pdf (sha256 797319fdâŠ; retrieved two ways: ams_2905.pdf and AMS_2905.pdf, identical bytes), read with pdftotext -layout and by eye.
| # | Predicted | Observed | Source line (verbatim) | Outcome |
|---|---|---|---|---|
| 28 | No SE CO per-ton FOB-farm alfalfa trade of any size (NO) | YES | Southeast section: Alfalfa - Fair (Trade/Per Ton) / Large Square 3x4 25 250.00 250.00 F.O.B.-Farm/Ranch Weedy 3rd Cut | MISS |
The print is not a grading-target trade (Fair grade, 25 t, under the 100-t bar). It is still informative: a weedy Fair 3x4 lot sold for $250 at the farm in the seller's region. Partial windows (not graded yet): #29 â no SE â¥100-t Good-or-better large-square trade on Sep 24; #30 â no Colorado â¥500-t FOB-farm alfalfa trade on Sep 24 (the 4,320 t is Premium millet at $200 FOB-feedlot); #31 â no NE CO per-ton alfalfa row and no "Dairy" anywhere in the report; #32 â first half: Trade activity Moderate. Demand Good to Very Good all markets of new crop hay.
Summary (from scripts/scorecard.py, 2026-10-02 â never hand-typed)
- Interval coverage: 9 of 19 two-sided numeric ranges contained the outcome (+2 NEAR). Up from 6 of 16; all three new numeric rows hit.
- Cluster A (economically central): 7 graded â 2 HIT (#14, #17), 5 MISS (#P1, #1, #9, #20, #28); 1 NO-PRINT (#21). Both observability calls missed, in opposite directions: #20 said a SE trade would print and none did; #28 said none would print and a small one did. The local tape is too thin to call either way, which is what the 2.5% base rate already said about the â¥100-t version.
- Cluster B: 1 HIT (#19), 1 NEAR (#16), 1 MISS (#2), 3 NO-PRINT.
- Cluster C: 18 graded â 11 HIT, 1 NEAR, 6 MISS, 1 NO-PRINT.
- Mean absolute point error, price rows: $19.12 (n=8), down from $24.80 (n=5).
- Calibration: coverage 47% (9/19) is below the 70% release threshold, so the ±$65 local price floor and ±$30 state-row floor stay.