NEWS

---

grainPlan 0.2.1

• Documentation names kernR's mechanism-consistency test directly.

grainPlan 0.2.0

Bug fixes

• grain_decision and grain_plan now store "ungrounded" for anything not grounded; the bracketed token of decideR 0.2.0 and earlier is still accepted.
• plan_nitrogen_rate(), plan_nitrogen_rate_from_manifest(), plan_variety(), plan_variety_from_manifest(), and plan_grade_target() raised a correctly-typed decline inside decideR (an abstained decideR::decision, itself stamped decideR_abstention), but returned an unstamped outer grain_decision whenever the wrapped decision abstained -- whether from ordinary evidence-driven abstention, a manifest producer's own typed abstention, or a manifest declaring an inferential_target the verb does not price. The outer object's class stayed c("grainPlan::grain_decision", "S7_object") regardless, invisible to the ORCHESTRA fleet's producer-agnostic decline predicate (is_orchestra_decline()). Fixed: the outer grain_decision's class vector is now prepended with grainPlan_abstention and orchestra_refusal whenever @abstained is TRUE, matching decideR's own stamping convention. Prepending, not replacing, so @ access and the registered print.grain_decision method are unaffected. What is refused is unchanged -- only how the decline is signalled to a fleet-level gate (GP-03 closeout).

Documentation

• The nitrogen-rate vignette chunk drew an independent posterior column per candidate rate; decideR's decisiveness gate is a paired comparison, so independent columns inflated the between-rate contrast variance and could mask a real, decisive profit difference. The chunk now shares one residual draw across rate columns, with the coupling requirement documented in-line (GP-01).
• The vignette's headline nitrogen example priced lasrosas.corn yield (quintals/ha) at $0.25/unit, roughly 100x below a realistic grain price, so the "profit-optimal" recommendation degenerated to the status-quo rate with no comment. Repriced to $22/quintal; the example now recommends a non-zero rate and a new captioned figure shows the profit-vs-rate curve behind it (GP-15, GP-16).
• The variety-ranking vignette chunk used each genotype's plot-level residual spread as the posterior standard deviation of the genotype mean, which overstates that uncertainty by roughly the square root of the replicate count; corrected to the standard error of the mean (sd / sqrt(n)). The prose claiming "the recommendation is the leading genotype" is now consistent with the printed decision, which previously abstained (GP-07).
• The variety short-list ledger printed integer candidate indices with no genotype label; the vignette now joins the variety names back onto the ledger before printing (GP-08).
• Printing a decideR::grade_band_value schedule dumped its closure body and a run-specific memory address into the rendered vignette; it is now rendered as its breaks/values/labels attributes (GP-14).
• The grade-target example's cost was high enough relative to the quality shift that the profit-optimal action degenerated to the status quo, at which point the printed rationale ("raising the posterior probability of clearing a higher quality band") did not describe what a zero-input action does; the example cost was lowered so the demonstration reaches a genuine, non-degenerate recommendation, and the full candidate ledger is now shown.
• Minor grammar and naming fixes: "nitrogen response trial" hyphenated to "nitrogen-response trial"; the lasrosas.corn trial location corrected to "Las Rosas" (agridat's own spelling), from the vignette's prior "Lasrosas farm".
• README.Rmd now renders without pandoc smart-punctuation substitution (md_extensions: -smart), so README.md no longer carries Unicode en dashes in prose; regenerated.
• The vignette was restructured to the orchestra's vignette quality bar: a stated grower question up front, a fixed Why/What/Do/Read/Limits/What-to- read-next/Reproduce shape, two ggplot2 figures (a profit-vs-rate curve and the wheat protein-band schedule) replacing the base-graphics plot and the raw attribute dump, every table through knitr::kable(), and a new season-composition example showing a plan that inherits an upstream manifest producer's own typed abstention alongside a grounded decision, read against the plan's grain_plan_table() and grain_is_grounded() output. The besag.met variety example is now labelled crop = "corn" (the trial is 64 corn hybrids across six North Carolina counties, not wheat); citations added for both agridat sources (Anselin, Bongiovanni & Lowenberg-DeBoer, 2004; Besag & Higdon, 1999).

Bug fixes

• plan_nitrogen_rate() now errors on a constraint argument, which decideR::decide_input_rate() silently dropped (it honours only min_ess/ess out of ...) -- a caller capping the candidate rate set saw the cap vanish with no warning (GP-02).
• plan_nitrogen_rate_from_manifest() and plan_variety_from_manifest() now check two manifest-contract obligations before pricing draws: a producer's own typed abstention (summary$abstained, contract v1.1) forces the status-quo action instead of a confident recommendation (GP-03), and a declared inferential_target other than one the verb is built to price does the same (GP-04). Both previously rode straight through to a priced answer.

Development

• Restored the R-CMD-check GitHub Actions workflow (dropped at c6069e7 for lack of a cross-repo secret to read the private decideR dependency); green once GH_PAT is added to the repo.

grainPlan 0.1.1

A safety and consistency release. No public function is removed or renamed, but the genomic manifest tail changes behaviour in one deliberate way (below).

Behaviour change

• plan_variety_from_manifest() no longer fabricates genomic uncertainty. When a manifest carries point GEBVs, the per-genotype reconstruction standard deviation is now resolved from the manifest itself (a prediction-error variance pev, a standard error gebv_se, or a reliability together with a genetic_var) or from an explicit caller gebv_sd. The previous silent default of gebv_sd = 1 is gone: when no uncertainty is available the function now errors rather than let an arbitrary, scale-blind constant decide whether a variety switch fires or abstains. The resolved provenance is recorded in the decision context as gebv_sd_source. Callers who already pass gebv_sd explicitly are unaffected.

Improvements

• plan_nitrogen_rate_from_manifest() gains a grounding override argument, matching plan_variety_from_manifest(), so a caller can force a manifest's grounding token either way.
• Abstention rationales are now exhaustive across every decision verb: all four abstention reasons decideR can emit (ungrounded input, insufficient evidence, no feasible action, too few effective draws) get plain-language, grower-facing phrasing from a single shared helper.
• Runnable manifest examples and a manifest walkthrough section were added to the vignette, and the README firewall example now executes on render.
• Added an explicit R (>= 3.5.0) floor (the S7 requirement) and an API-stability statement.

Internal

• Removed an unused null-coalescing helper.
• Added targeted tests: the reconstruction-scale dependence, the low-effective- sample and no-feasible-action abstention branches, a functional grade-target cost, a nitrogen price-scaling invariant, and an independent verification that the variety decision computes the correct expected merit.

grainPlan 0.1.0

First release. grainPlan is the grain-specific decision orchestrator of the ORCHESTRA agricultural analytics stack -- the practitioner last-mile that turns an upstream member's inference result into an actionable grain-production plan. It is a thin grain-semantics layer on top of decideR, reusing that package's decision engine, manifest-native pipeline tail, grade-band loss library, and grounding firewall rather than re-implementing decision theory.

New features

• Grain decision verbs that wrap decideR's engine with grain meaning: plan_nitrogen_rate() and plan_nitrogen_rate_from_manifest() price a nitrogen rate by expected profit; plan_variety() and plan_variety_from_manifest() rank a variety from a multi-environment trial or a breeding_values manifest under a downside-aware merit loss; plan_grade_target() chooses a quality-driving input under a grade-band payoff.
• A reusable grain value / loss library that consolidates the economics the orchestra runs kept hand-building: wheat_protein_bands(), barley_malting_premium() (both decideR::grade_band_value() schedules built from published grade conventions), and n_rate_economics().
• A season orchestrator: plan_season() composes several grain decisions into one grain_plan, combining their grounding worst-case so a plan resting on any unverified decision abstains as a whole; grain_plan_table() flattens a plan for a report.
• S7 result types grain_decision (one recommended grain action wrapping a decideR::decision) and grain_plan (an ordered season of decisions), with a grain_is_grounded() predicate and print() methods.
• The Independent Oracle Principle firewall is inherited and enforced at every verb: an ungrounded input forces abstention to the status quo, never a confident wrong recommendation.