Southeast Colorado alfalfa price forecasts, graded in public.

Two Reports Compared — September 19, 2026

What is being compared. (A) SeptemberHayReport.md as corrected the same day after the independent review (research/17 + research/18). (B) IndependentHayReport-2026-09-19.md, written by a second model from the same saved data after its review. Both were read in full; every factual claim in (B) was checked against the saved files and holds.

Verdict

For the seller's decision, (B) is the better document. It refuses to average heterogeneous trades into a single McClave price, it puts procedure (test, weigh, three written bids, convert to net at the stack) ahead of any number, it makes the cost of waiting explicit with break-even arithmetic, and it treats the $180 offer as a stale note to be re-bid rather than something to accept or reject on a "floor". Those are the right disciplines for a private sale decision, and (A) reached them only after being corrected.

For the public product, (A) is the better document, because (B) does not attempt it. (A) carries the graded scorecard with cited lines, the frozen predictions with source specs, the two-target grids, the calibration floors, the adjustment ledger, the watch calendar, and the verification appendix. (B) explicitly declines to publish a grid or replace frozen predictions. A public accountability product needs the machinery; a seller needs the discipline.

The final report is therefore a merge, already applied: (A)'s seller section now follows (B)'s procedure-first structure, break-even table, "no automatic winter premium" budgeting rule, and $180 handling; the reference ranges from completed trades are kept but labeled as comparables, not a valuation of the stack. The grid, scorecard, predictions and calendar remain as (A)'s public layer, with the winter lift labeled discretionary.

Point by point

Question(A) corrected(B) independentBetter
Is a local price stated?Yes: $270 G/P, $265 Good, ±$65, from tonnage-weighted comps incl. the $170 lotNo: "$260 NE CO Good as a negotiating reference"; refuses to average(B) for the seller; (A) for the graded public target, which must be a number
Handling of the $170 OK lotIncluded after correction; drives the nowcast down $10Kept visible as a counterexample; treated as something to investigate, not average(B): investigation beats inclusion in an average
NetbackWithdrawn; no netback computedExplains why an origin price cannot be a netback; freight shown as replacement-cost illustrations with a sensitivity column(B): the sensitivity column (12–20 ¢/t-mi) is the honest way to show freight
Winter timing"Consistent with Nov–Dec, not established"; grid still carries a discretionary +6%"No automatic winter premium"; scenario table with implications, no dollar path(B) for budgeting; (A)'s +6% is defensible only as a labeled judgment
Cost of waitingAbsent until mergedBreak-even: ($250+$10)/0.97 ≈ $268; a $25 rise earns ~$7(B), clearly; this is the single most useful thing either report says to a seller
$180 offerFirst "reject", then "judge against bids""Neither accept nor reject on an old note; seek better bids; if bids stay near $180, find out why"(B)
Water/droughtFull detail (canal dry, 25% Jun–Sep, JM 5.6%, Bent County D1)Same facts, correctly limited to what they imply (replacement supply, not stack quality); uses the like-for-like 26% Jun–Aug figure(B) on inference discipline; (A) on completeness
El Niño / relief riskDetailed outlook; drought-outlook used to shape the gridNamed as "meaningful relief risk" with the CPC caveat that conditions can worsen firsttie
Grading / accountabilityFull scorecard, ledger, predictions #28–39, calibration, adjustments ledgerCites the project's 6/16 and 1/5 honestly; publishes nothing gradeable of its own(A)
Verification§11 appendix with two methods per claim, withdrawn claims listedEvery claim linked to a saved file; no appendix table(A) on form; both adequate
Exit rulesKept, seniorKept, explicitly "risk limits, not peak-price signals"tie; (B)'s phrasing adopted
Length / usability for the sellerLong; the seller must find their sectionTwo pages, decision-first(B)

What (B) gets right that (A) should keep permanently

  1. Do not average heterogeneous lots into a precise local price for a specific stack. The public target needs a number; the seller does not. Separate them.
  2. Procedure before price. Test, weigh, three identical-spec bids, net at the stack.
  3. Break-even before "hold for the peak". Carry cost and shrink make a $25 price rise worth ~$7.
  4. Freight as a sensitivity, not a point. 12–20 ¢/ton-mile is the honest range; the parity script's single rate should print that band.
  5. Counterexamples are investigated, not smoothed. The $170 lot is a phone call, not a data point to weight.

What (A) keeps that (B) does not provide

  1. The frozen, source-specified predictions and the machine-checked ledger, which are the only way this project ever learns whether its judgment beats persistence.
  2. The two-target separation (state benchmark vs local), with floors that now fail closed.
  3. The adjustment ledger that names every discretionary move (winter +6%, drought-outlook roll-off, milk on #31).
  4. The calendar of gradeable checkpoints and the site that publishes misses.

Where (B) is thin

Rule added

L11 is amended again: the seller section of any report follows the procedure-first, break-even, no-automatic-premium structure; reference ranges from comparables are labeled as such and never presented as the stack's value; an outside review of the inference layer (not just the data) runs before publication, because the September 19 first version passed all 39 tests and every data check and was still wrong in three load-bearing places.