Effect Methodology: From Verified Operation to Attributable Outcome

Amendment D. Joins verification, seasonal analysis, farm history, and reconciliation into an intervention-to-outcome methodology with three claim strengths, per-crop baselines, per-hazard attribution windows, and an explicit no-claim state.

Verification proves an operation happened. Nothing here proves it worked. The Verification Model caps its claim at the operation is logged and cross-checked; an agency accepting a bundle under that regime is buying evidence of activity, not evidence of impact. Semai’s honest answer to did it help is separate from its answer to did it happen, and the two answers travel under different disclosures.

Amendment D defines that separate track. It joins four pieces that already exist and are unjoined: Seasonal Analysis correlates index against yield with R² and a residual map, the Risk Model farm_history layer learns from what happened on a farm, Estate Group reconciles forecast against actual harvest by field, and the Aggregation Model defines forecast, confidence degradation, and k-anonymity. The methodology adds no new aggregation category, invents no new causal machinery, and refuses more claims than it makes.

Amendment D routes public money. A confident wrong answer here costs more than an honest not yet. The largest section of this page is When no claim can be made; read it before reading the positive case.

Three claim strengths

The methodology carries three claim tiers, in increasing strength and decreasing availability. Every effect measure, baseline, and disclosure below anchors to which tier it can support.

Tier	Meaning	What it lets the platform say	Availability
Observed change	A measured difference in the field’s own state, without an attribution.	”This field returned to the healthy band 9 days after the mitigation was applied.” No claim about cause.	Fast. Available as soon as the attribution window elapses.
Attributable change	An observed change paired with an intervention log, an attribution window, a disposed set of confounders, and a primary baseline that formed.	”On this field, the mitigation is associated with a return to band inside its rulecard window. The paired prior cycle showed no such recovery in the same phenology window.”	Medium. Requires the primary baseline to form and confounders to be controlled.
Counterfactual impact	An attributable change generalised across a cohort with a defined comparison arm, respecting k-anonymity, aggregated per Aggregation Model discipline.	”Across the 47-field cohort, the treated arm recovered inside its rulecard window at a rate N percentage points higher than the untreated arm; the reconciled harvest gap is M% area-weighted, at min-confidence C.”	Slow. Scheme-level only. Refused if k or comparability fail.

Every downstream artifact discloses which tier it supports. A bundle that claims counterfactual impact but only demonstrates observed change fails the disclosure rule and cannot ship. This is the same discipline the Aggregation Model reproducibility rule applies to rollups.

Vocabulary

These terms are strict; do not use them interchangeably. In particular the three claim-strength words above (observed change, attributable change, counterfactual impact) carry distinct disclosure rules and are not synonyms for “effect”.

Term	Definition
Intervention	A logged operation intended to change an outcome. See What counts as an intervention for the measurable and non-measurable split.
Effect measure	A quantity a claim rests on. Four are defined below and ranked by claim strength.
Attribution window	The interval after an intervention during which an effect claim is admissible on this platform. Tied to the rulecard’s days_to_mitigate, capped per crop. See The attribution window.
Baseline	The comparison the effect measure is taken against. Three candidates; per-crop primary and fallback. See The baseline.
Cohort	A set of comparable fields for counterfactual-impact claims. Comparability keys are Amendment A’s band-resolver keys (crop, ecosystem, stage, region) plus pre-intervention deviation from band at the attribution window’s opening. Never combined across crops. See Cohort selection bias and the fifth key.
Comparison arm	A subset of a cohort defined by intervention status: treated (received the intervention) or untreated (did not, in the same window). Not a randomised assignment; a documented split.
Rulecard yield-impact curve	The mapping from severity band and stage to projected yield loss, per Risk Model → yield-impact layers. Used as the counterfactual for the avoided-loss effect measure.
No-claim state	The declared outcome when the methodology cannot honestly compute a claim. Not silent, not defaulted; a positive assertion that the answer is not yet with a stated reason.

Effect is not attribution. A field returning to band is an observed change. Saying the mitigation caused it is an attributable change and requires a baseline. Saying the scheme would have lost N% without disbursement is a counterfactual-impact claim and requires a cohort. The platform will show observed change without attribution; it will not show attribution without a baseline.

What counts as an intervention

An intervention is a logged operation intended to change an outcome. The list is deliberately narrow: an operation that leaves no on-farm signal cannot support an effect claim at field granularity regardless of how much money is attached to it.

Measurable at field granularity

Intervention	Log source	Attribution basis
Applied application plan	VRA Maps map.applied event, per Verification Model source events.	The map ties a rate to a zone; the paired satellite pass corroborates uptake. Both are in the record.
Completed field check that changed a decision	scout.completed event with a paired state transition on a finding (per Finding Provenance) — promoted, demoted, contradicted, or resolved.	The check’s outcome is bound to a specific rulecard-firing or human finding. A scout.completed with no state change is not an intervention; it is corroboration.
Irrigation action logged as a source event	map.applied on an irrigation plan, or an explicit irrigation.completed event if the estate integrates a controller.	Same cross-check discipline as any application plan.
Harvest declared	harvest.declared per Verification Model.	Not an intervention on the current cycle, but the terminal signal against which yield-delta claims are computed.

Not measurable at field granularity

Item	Reason	Where it can land
Disbursed subsidy	The subsidy is disbursed to an entity, not applied to a field. There is no on-farm signal that the money existed. What is measurable is what the recipient did with the money, if the follow-on operation was itself logged as one of the measurable interventions above.	Scheme-level counterfactual-impact claims only, at cohort level, with a documented comparison arm. Field-level subsidy attribution is refused. See Guardrails → Refusals and What the agency sees.
Advisory content delivered	An advisor artifact per Semai Advisor overview is a Task, Explanation, or Proposal. Delivery is logged; adoption is not, unless the user then executes a measurable intervention.	Advisor influence surfaces as an antecedent to a measurable intervention, not as an intervention itself.
Insurance payout	Same shape as subsidy: an entity-level transfer with no on-farm signal.	Same scheme-level cohort disposition.
Training or extension visit	No scout.completed semantics apply; no rulecard state changes.	Cannot support any effect tier at field level.

Decision: subsidy attribution. Field-granularity attribution of a subsidy disbursement to a field-level outcome is refused. Scheme-level cohort attribution is the honest answer and is the only path by which a subsidy claim reaches counterfactual impact. This is written as decision rather than option because the alternative (letting agencies claim field-level subsidy attribution on any signal correlated in time) is the failure mode the Verification Model already refuses for did-it-happen claims; extending the refusal to did-it-help is consistent.

Effect measures, ranked by claim strength

Four measures. They are not a menu; they are a hierarchy. The methodology maps each intervention type to a primary measure and lists the fallbacks in order.

Rank	Measure	What it says	Claim tier ceiling	Horizon
1	Yield delta	Actual harvest against the forecast for the same field, per Estate Group reconciliation.	Attributable change (field-level) or counterfactual impact (scheme cohort).	End of cycle. Slowest.
2	Avoided-loss against yield-impact curve	The delta between the rulecard’s projected yield loss at intervention time and the resolver’s projection now, using the rulecard yield-impact curve as counterfactual. Reported as physical response (recovery to band, days to recovery) at any resolver layer; reported as a monetary (ringgit) figure only when the resolver lands at farm_calibrated. See Avoided-loss and the resolver-layer gate.	Attributable change (physical); attributable change with monetary figure only at farm_calibrated.	Intervention close.
3	Recovery back inside band	A state claim: the field returned to the healthy band per Amendment A within the attribution window.	Observed change primary; attributable change only when paired with a baseline.	Days after mitigation.
4	Days to recovery	Duration between mitigation and return to band. Useful for comparing interventions on the same hazard; not for attribution against a baseline.	Observed change only.	Days after mitigation.

Avoided-loss and the resolver-layer gate

The Risk Model resolves yield-impact through four layers, in order: farm_calibrated → by_region → by_ecosystem → literature_v0. The resolved layer is the trust score for the avoided-loss measure: it is graded, already computed, and already inspectable on every rulecard-firing.

The gate.

Resolver layer	Physical response (recovery to band, days to recovery)	Monetary avoided-loss (ringgit figure)
farm_calibrated — curve tuned on this farm’s own history	✓	✓
by_region	✓	Refused
by_ecosystem	✓	Refused
literature_v0 — literature default	✓	Refused

A monetary avoided-loss figure may be reported only where the rulecard’s yield-impact resolved at farm_calibrated. At any layer above that, the platform reports the physical response and declines the ringgit figure. This is stricter than untrusted curve → refuse: a literature-default curve is not untrusted for physical response; it is untrusted for money. Publishing a ringgit figure derived from literature_v0 against an agency disbursement is the single most expensive error available on this page.

The chain to farm_calibrated. The farm_calibrated layer is built from farm_history per Risk Model → yield-impact layers. farm_history calibration was gated on corroboration in Finding Provenance → Feedback into farm_history. So the chain that unlocks a ringgit figure on a given farm is: corroborated resolved finding → feeds farm_history → enables farm_calibrated resolution on that hazard for that farm → unlocks monetary avoided-loss. A farm with no corroboration history sees physical response only, regardless of how many interventions it has logged.

Primary measure per intervention type.

Intervention	Primary measure	Fallbacks (in order)
Applied application plan (fertiliser, plant protection)	Avoided-loss against yield-impact curve	Recovery back inside band, then yield delta at end of cycle
Completed field check that changed a decision	Recovery back inside band	Days to recovery for comparative reporting
Irrigation action	Recovery back inside band	Yield delta at end of cycle if the water stress rulecard has a trusted yield-impact curve
Subsidy disbursement (scheme-level only)	Yield delta across the cohort	No field-level fallback exists; the tier does not degrade to attributable change at field level

The yield-impact curve resolver layer is the trust score for the avoided-loss measure. See Avoided-loss and the resolver-layer gate below; the ringgit figure is admissible only when the Risk Model resolver lands at farm_calibrated.

Saturation is a structural precondition, not a degradable confounder

Above roughly NDVI 0.80, the index stops discriminating. A closed oil palm canopy at prime and a tapped rubber block at full canopy sit above that ceiling permanently, not intermittently. Two of the four crops cannot demonstrate effect on their default index at their most valuable stage: not sometimes, not under cloud, structurally and always.

The failure mode this closes is the worst one this page can produce. A scheme report that silently ran NDVI on prime oil palm or on tapped rubber measured nothing at all and will return a null effect that looks like an honest null. An agency reading it cannot tell the difference between the intervention did nothing and we did not measure. That is worse than a refusal because it is indistinguishable from one.

Rule. Effect measurement on oil palm at prime phenology and on rubber under tapping must run on NDRE per Indices, not NDVI. If NDRE is unavailable for the attribution window (drone imagery does not provide the red-edge band, or the corroborating pass is missing NDRE), the claim is refused, not degraded. The refusal cites saturation and states which pass or index was missing.

This is a precondition on the effect artifact, enforced in Guardrails → Preconditions. It is not one confounder among many. The confounder table below carries saturation only for the intermittent case (a paddy field driving into saturation late in the season, a young pineapple canopy approaching closure); the structural case for oil palm at prime and rubber under tapping is resolved here, above the confounder table, and never degrades to observed change.

The baseline

Three candidates, each with a failure mode named in the brief. Amendment D’s decision is a per-crop primary and per-crop fallback, because the crop cycle models differ enough that a single global primary would be dishonest.

The candidates

Candidate	What it compares	Failure mode
Before and after (same field)	The field’s state in the attribution window against the same field before the intervention.	Confounded by phenology: Amendment A’s band moves anyway as the crop advances through stages. A recovery observed after mitigation may be the same recovery the band model predicts without any intervention.
Prior comparable cycle (same field)	The field’s state in the intervention window against the same field in a prior cycle at the same phenology, filtered by Amendment B comparability.	Limited by how many cycles exist for the field, and by season-to-season change in variety, weather, and ecosystem. New plantings and switched varieties fail comparability.
Cohort (comparable untreated fields)	The treated field(s) against a documented arm of comparable fields that did not receive the intervention in the same window. Comparability keys: crop, ecosystem, stage, region per Amendment A, plus pre-intervention deviation from band at the window’s opening — the fifth key that keeps the untreated arm honest against selection bias. See Cohort selection bias and the fifth key.	Requires enough untreated fields at comparable deviation to satisfy k-anonymity at k=5 per Aggregation Model, and a documented arm (not a residual left over from a treatment rollout). A cohort drawn without deviation-matching cannot support counterfactual impact and degrades to attributable change.

Per-crop decision

Crop	Primary baseline	Fallback	Argument
Rice (paddy)	Cohort	Prior comparable cycle then before-and-after	Paddy runs Main and Off Seasons per Crop Cycle Models; a field has few prior comparable cycles and season-to-season variety and weather change hurt them. Enough fields exist per scheme to form cohorts.
Oil palm	Prior comparable cycle	Cohort then before-and-after	Oil palm’s Production Year and tree-age cohorts give a field many prior comparable windows. Cohort formation is possible but competes with the field’s own strong history.
Rubber	Prior comparable cycle	Cohort then before-and-after	Rubber’s tapping panel model is continuous with strong panel-year history; comparison against the same panel in the prior tapping year controls for panel biology better than a cross-block cohort would. The annual wintering dip is excluded from the attribution window in any baseline.
Pineapple	Cohort	Before-and-after (prior comparable cycle rejected)	Pineapple’s Planting Batches mean multiple batches per year; batch-to-batch comparison is confounded by forcing timing and batch age. Cohort is primary; prior comparable cycle is not a defensible fallback and is skipped.

Before-and-after is never primary. In every crop it is the last-resort fallback, used only when neither prior-cycle nor cohort can form. When it is used, the claim-strength ceiling is observed change; the methodology refuses to promote a before-and-after comparison to attributable change.

Amendment A’s band resolver keys on crop, ecosystem, and stage. Reusing those keys as cohort comparability keys is what makes cohort baselines computable for what a healthy band looks like. It is not enough for what an untreated arm looks like: see Cohort selection bias and the fifth key immediately below.

Cohort selection bias and the fifth key

Fields that received an intervention were usually the worse ones. That is why they got it. A cohort matched only on crop, ecosystem, stage, and region therefore compares a distressed treated arm against an average untreated arm and systematically understates the effect. The bias is not noise; it has a direction, and it is always the same direction.

This matters more than it would elsewhere because cohort is the primary baseline for paddy and pineapple per Per-crop decision. Two of the four crops rest their strongest available claim on an arm that is systematically not comparable unless the deviation gate is enforced.

Rule. Cohort comparability requires a fifth key beyond Amendment A’s four: pre-intervention deviation from band at the attribution window’s opening. Amendment A already computes deviation on every field to place it inside or outside the healthy band; the key is available without new signal work. Untreated arm members must fall inside the same deviation bracket as the treated arm at the window’s opening, at bracket widths defined per crop (draft: 0.5σ of the crop’s band width; the exact widths are a Risk Model calibration and are noted in Open Questions).

Failure mode. A cohort drawn without deviation-matching produces a counterfactual-impact claim that reads as an honest number and understates the intervention. An agency using that number to decide the next round of disbursement under-funds a working programme. The refusal is loud rather than a downgrade: no counterfactual-impact artifact ships from a cohort that has not passed the deviation gate.

What degradation looks like. If deviation-matching drops the untreated arm below k=5, the artifact degrades to attributable change (per-field), not to a counterfactual-impact claim on the un-matched cohort. Attributable change on the treated field with its own baseline (prior-cycle where possible, before-and-after as ceiling-at-observed-change fallback) is honest; a bias-inflated counterfactual-impact is not.

The attribution window

The window during which an effect claim is admissible. Tied to the rulecard’s days_to_mitigate per Risk Model, capped by per-crop cycle geometry.

Rulecard-anchored, crop-capped

Rule	Meaning
Lower bound	The attribution window opens at days_to_mitigate * 0.5 after intervention. Before that the intervention has not had biological time to work; any observed change is background.
Nominal window	days_to_mitigate to 2 * days_to_mitigate. This is the admissible interval for an effect claim on this hazard.
Upper bound	Beyond 2 * days_to_mitigate, confounders dominate (weather, phenology drift, subsequent operations). Effect claims after this point are refused; observed change may still be logged without attribution.
Crop ceiling	Whichever of the crop caps below arrives first, the window closes.

Per-crop ceilings

Crop	Ceiling	Why
Rice (paddy)	End of the current season’s harvest per Crop Cycle Models. Interventions in previous seasons attribute to those seasons.	Paddy seasons are 3 to 4 months; attribution across seasons is a variety-and-weather comparison, not an intervention comparison.
Oil palm	12 months from intervention, capped at the current Production Year.	One production year is the longest defensible horizon before multi-year weather and management effects dominate.
Rubber	The current tapping year, with the annual wintering dip explicitly excluded.	Wintering surfaces as an annual NDVI decline and will masquerade as intervention failure or success depending on timing.
Pineapple	End of the current Planting Batch at harvest.	Batched forcing means attributing across batches confounds intervention effect with batch-age and forcing timing.

Rulecard-anchored, not calendar-anchored. An intervention against a fast-cycle disease (rulecard days_to_mitigate = 14 days) has a nominal window of 14 to 28 days after intervention. An intervention against a slow-cycle nutrient deficiency (rulecard days_to_mitigate = 45 days) has a nominal window of 45 to 90 days. The methodology does not apply a single “60 days after any intervention” rule; the rulecard already knows the biological horizon and the effect methodology inherits it.

Confounders

The claim is only as strong as the disposition of these confounders. Each is either controlled for (the methodology has a rule that neutralises it) or refused if uncontrollable (the claim degrades to observed change or to no-claim).

Confounder	Source	Disposition
Weather during the attribution window	Weather layer, per Risk Model.	Controlled for in cohort baselines by requiring the comparison arm to share the same weather cell; controlled for in prior-cycle baselines by rejecting the cycle if a documented extreme (drought, flood) does not repeat. Refused if weather is missing.
Phenology drift (the band moves anyway)	Amendment A band resolver.	Controlled for by anchoring the attribution window to stage_at_intervention and requiring the baseline to compare at the same stage. Before-and-after cannot control this; that is why before-and-after ceilings at observed change.
Ecosystem	Amendment A.	Controlled for in cohort baselines by requiring ecosystem match. In prior-cycle baselines the ecosystem is by definition constant.
Variety	Field metadata.	Controlled for by requiring variety match on the field or on cohort members. New plantings and switched varieties fail comparability under prior-cycle; cohort is the only path.
Region	Amendment A.	Controlled for in cohort baselines by requiring region match. Cross-region cohorts are refused.
Saturation (NDVI ceiling in dense canopy, intermittent case only)	Amendment A distortions; Indices.	Structural cases (oil palm at prime, rubber under tapping) are resolved as a precondition above the table; see Saturation is a structural precondition. This row covers the intermittent case only (paddy driving into saturation late in the season, a young pineapple canopy approaching closure): switch to NDRE or SAVI when available; refuse otherwise.
Soil background (young canopy, sparse cover)	Amendment A distortions.	Refused if uncontrollable when the crop stage is below closed-canopy; switch to a soil-tolerant index or refuse the claim.
Wintering (rubber)	Rubber wintering dip.	Controlled for by excluding the wintering window from the attribution window. If the attribution window and the wintering window overlap and the rulecard requires the full window, the claim is refused.
CAM lag (pineapple)	Pineapple CAM physiology.	Controlled for by lengthening the lower bound of the attribution window to account for lag; the methodology uses the pineapple-specific days_to_mitigate from the rulecard rather than a generic index-response horizon.
Subsequent operations in the same window	Operation log.	Controlled for by exclusion: if a second measurable intervention on the same hazard occurred inside the attribution window, the attribution splits (both interventions share the credit at the observed-change tier; neither reaches attributable change alone).
Cloud cover blocking the corroborating pass	Verification Model weak-grade.	Controlled for by the same weak-grade disclosure verification uses. An effect claim built on a weak-graded pass carries the weak flag through to the agency artifact.
Cohort selection bias (fields that received intervention were the worse ones)	This page; Cohort selection bias and the fifth key.	Refused if uncontrollable. Cohorts must match on pre-intervention deviation from band at the attribution window’s opening; a cohort that cannot be matched degrades to attributable change and does not support a counterfactual-impact claim. Directional bias (always understates effect), not noise; cannot be resolved by adding samples.

An effect claim that ignores any applicable confounder is refused, not downgraded. The refusal is loud; see Guardrails.

When no claim can be made

The largest section, per the brief and per the product’s habit. Every condition below produces a no-claim state: a positive assertion that the answer is not yet, with a stated reason and a stated condition that would allow a claim to become possible. Silence is not permitted; the platform’s honest posture is to say we cannot tell yet, here is why, here is what would need to be true.

Condition	What is shown	What would need to be true
Primary baseline cannot form and no fallback is defensible for this crop	Observed change only, with disclosure: “No comparable baseline available; showing state change only.”	Enough prior cycles (oil palm, rubber), or enough comparable untreated fields (paddy, pineapple), to form the primary.
Attribution window has not elapsed	Intervention log with countdown: “Attribution window opens in N days; recovery to band is not yet a claim.”	The window’s lower bound (days_to_mitigate * 0.5) elapses without a subsequent intervention on the same hazard.
Attribution window has closed (past 2 * days_to_mitigate or crop ceiling)	Observed change only if state is still consistent; no attributable claim: “Attribution window closed; any recovery observed now is not attributed to this intervention.”	Not recoverable for this intervention. A subsequent intervention resets the clock for itself, not for the prior one.
A confounder listed as refused if uncontrollable applies	No-claim state with the specific confounder named: “Not enough soil-corrected data at this stage; NDVI is soil-contaminated below closed canopy.”	The confounder resolves (canopy closes; index switches to soil-tolerant; wintering window ends) inside the attribution window.
Cohort forms below k at k=5 per Aggregation Model	Cohort claim refused: “Cohort has fewer than 5 comparable fields; scheme-level claim withheld.”	Additional comparable fields enter the cohort (through the scheme’s own growth, not through relaxing comparability keys).
Cohort cannot match on pre-intervention deviation (selection bias uncontrollable)	Counterfactual-impact refused; per-field attributable-change reported where each field’s own baseline forms: “Untreated arm cannot be matched on pre-intervention deviation; scheme-level claim would understate the effect. Reporting per-field attributable change instead.”	Enough untreated fields at comparable deviation enter the cohort. Relaxing the deviation gate is not offered as a path; the gate is the point. See Cohort selection bias and the fifth key.
Cohort breakdown would fall below k (the k-anonymity-trap case)	The breakdown is refused, not silenced; the aggregate remains: “Breakdown by ecosystem withheld at k=5; total scheme claim remains available.”	Cohort growth.
Forecast-vs-actual reconciliation is not available for the yield-delta measure	Yield delta refused; fallback to next measure in the hierarchy: “Reconciled harvest not yet available; showing avoided-loss estimate against rulecard curve.”	Harvest is declared and reconciled per Estate Group.
Monetary avoided-loss requested but resolver did not land at farm_calibrated	Ringgit figure refused; physical response reported instead: “Yield-impact curve for this hazard resolved at by_region; showing recovery to band and days to recovery, monetary figure withheld.”	The farm accumulates enough corroborated resolved findings on this hazard for farm_history to calibrate the curve; see Avoided-loss and the resolver-layer gate.
Two interventions on the same hazard inside the window	Attribution split; each intervention ceilings at observed change alone: “Two interventions in this window; credit is not attributable to either alone.”	A future intervention on the same hazard with a clear window (no second intervention inside 2 * days_to_mitigate).
Cross-crop rollup requested for an effect claim	Refused per Aggregation Model — never average across crops: “Effect measured on paddy at ripening and oil palm at prime are not the same quantity. No cross-crop average will be shown.”	Not resolvable. The refusal is structural, not conditional.
Weak-graded satellite pass (cloud) is the only corroboration	Claim carries the weak flag; the tier ceilings at attributable change with weak disclosure. Counterfactual impact is refused on weak-only cohorts.	A subsequent unblocked pass corroborates.

The no-claim state is a first-class artifact. It bundles into the Verification evidence envelope, is logged per the Audit Envelope, and is displayed on every surface an effect claim would be displayed on. An agency that receives a no-claim artifact has the same tamper-evident guarantee as a positive claim: they can prove the platform refused, when, and why.

How it aggregates

Field-level effect measures roll to scheme-level answers under the existing Aggregation Model discipline. Amendment D adds no new category. It adds one composition rule and one k-anonymity extension.

The existing discipline (unchanged)

• Never average an average. Field-level effect claims do not average into a scheme-level claim; they compose per the aggregation category applicable to the measure. Recovery back inside band is a count with an area weight per field; days to recovery is a per-field duration and is aggregated as a distribution (median, IQR), not as a mean.
• Area-weight intensive measures. Yield delta expressed as a percentage is intensive; when rolled to scheme level it is area-weighted by the treated area on each field. Absolute yield delta in kg or tonnes is extensive and sums directly, subject to the reconciliation source.
• k-anonymity at k=5. Any breakdown of the scheme claim that would land below k=5 fields is withheld; the aggregate remains available.
• Confidence as minimum of inputs. The scheme-level effect claim’s confidence is the minimum of the field-level confidences composing it. A single weak-graded field lowers the scheme confidence; it does not average with the strong-graded ones.

The Amendment D additions

Cohort-breakdown k-floor. Once a cohort is defined at k, any breakdown of that cohort (by crop, ecosystem, region, intervention timing) inherits the k floor and refuses breakdowns that would violate it. The failure mode this closes: a scheme covers 12 fields, 7 treated, 5 untreated. The cohort has k=5 exactly at the untreated arm. Any further breakdown (by ecosystem, by region) would drop below k on one side. The methodology withholds the breakdown; the aggregate remains available. This is a natural extension of the existing k rule and is stated explicitly because the shape (subset of a cohort must also respect k) is not obvious from the general rule.

Attribution-window composition at scheme level. When rolling field-level attributable-change claims to a scheme-level counterfactual-impact claim, the scheme-level attribution window is the intersection of the field-level windows, not the union. A scheme with 100 treated fields whose windows range from Day 14 to Day 90 supports a counterfactual-impact claim only over the interval where every field’s window is open. The scheme claim outside that interval degrades to observed change. This preserves the minimum-of-inputs discipline the aggregation model already applies to confidence.

Cross-crop refusal is structural. As with the band model, effect measures do not average across crops. A national dashboard showing “scheme-level effect: N%” across paddy, oil palm, rubber, and pineapple is refused; the dashboard shows four separate claims or none.

What the agency sees

Not a screen design. The claims the platform is willing to make, in tiers, and the evidence attached to each.

The claims

Tier	Claim wording (illustrative)	Attached evidence
Observed change	”On this field, recovery to healthy band observed 9 days after the mitigation was applied on 12 Jun.”	Before-and-after band states, indexed pass timestamps, Verification bundle for the map.applied event.
Attributable change	”On this field, the mitigation applied on 12 Jun is associated with a recovery to band inside the rulecard window. Paired prior-cycle (Main Season 1 · 2024) at the same phenology showed no such recovery.”	All observed-change evidence, plus: baseline identifier (prior-cycle citation or cohort id), the confounders’ dispositions, the rulecard window used, the attribution window’s open and close dates.
Counterfactual impact	”Across the 47-field cohort (paddy, MD2 ecosystem, Perlis, ripening stage), the treated arm recovered inside the rulecard window at a rate 22 percentage points higher than the untreated arm. Reconciled harvest gap is 8.4% area-weighted, at min-confidence C. k=23 treated, k=24 untreated.”	All attributable-change evidence for each field composing the cohort, plus: cohort definition (comparability keys), comparison-arm construction, k on each arm, aggregation rules applied, reconciliation source.
No claim	”Attribution window opens in 4 days; recovery to band is not yet a claim. No effect will be shown before 16 Jun.” Or: “Cohort has 3 comparable fields; scheme-level claim withheld at k=5.”	The specific reason from the When no claim can be made table, and the condition that would allow the claim to become possible.

The disclosure envelope

Every tier’s artifact carries:

• The tier itself, named. No artifact conflates tiers.
• The intervention log reference (event id, actor, timestamp, verification bundle).
• The baseline reference (which of the three candidates was used, why the primary was chosen or the fallback was required).
• The attribution window’s open and close dates.
• The confounders’ dispositions (which were controlled, which if any were noted as weak).
• The verification grade of the corroborating pass; weak grade carries through.
• The reconciliation source if the claim rests on yield delta.
• For counterfactual-impact tier: cohort id, k on each arm, aggregation rules applied.

An artifact missing any of these fields fails the disclosure rule and is refused by the Guardrails audit line.

The subsidy case, stated bluntly

An agency disbursing subsidy under Amendment D can receive:

• Not a field-level attributable-change claim tying the subsidy to a specific field’s outcome. Refused; see What counts as an intervention.
• Not a counterfactual-impact claim for the scheme unless a documented untreated cohort exists and satisfies k. Refused otherwise.
• Yes an operation-log summary of measurable interventions taken by subsidy recipients (application plans applied, field checks completed, irrigation actions logged), each carrying its own effect artifact at whatever tier that intervention’s evidence supports.
• Yes a no-claim artifact for the scheme where a claim cannot be made, with the reason stated.

The agency’s disbursement round can be directed by the aggregate of the recipient interventions’ effect artifacts. It cannot be directed by a fabricated field-level subsidy attribution.

Guardrails

Amendment D follows the shared guardrails template. No new category is introduced.

Category	Rule
Input validation	Every effect artifact requires: intervention_ref (a verifiable source event per Verification Model), hazard_ref (a rulecard id or literature ref), field_ref and crop_cycle_ref, attribution_window (open and close ISO 8601), effect_measure (one of the four measures), baseline (one of the three candidates and which was primary vs fallback), confounder_dispositions (an entry for each applicable confounder), verification_grade, and claim_tier. Counterfactual-impact artifacts additionally require cohort_id, treated_arm_k, untreated_arm_k, and aggregation_rules.
Preconditions	The intervention must be one of the measurable types in What counts as an intervention. The rulecard must exist and its days_to_mitigate must be set; effect artifacts against a rulecard without days_to_mitigate are refused. The field’s crop_cycle must be the one the intervention was logged against. Cohorts must form on Amendment A comparability keys plus pre-intervention deviation from band at the attribution window’s opening, and satisfy k on each arm at the matched deviation bracket per Cohort selection bias and the fifth key. Index precondition: effect measurement on oil palm at prime and rubber under tapping must run on NDRE; a claim on either at those phenologies computed on NDVI, or on any index without a red-edge band, is refused per Saturation is a structural precondition.
Refusals	Cross-crop effect averaging (structural refusal, see How it aggregates). Field-level subsidy attribution. Attribution claim before days_to_mitigate * 0.5 has elapsed. Attribution claim after 2 * days_to_mitigate or the crop ceiling. Before-and-after used as anything above observed change. Cohort claim below k=5 on either arm. Cohort breakdown that would drop below k. Cohort claim built without pre-intervention deviation matching per Cohort selection bias and the fifth key. Effect claim built on a weak-graded pass presented as counterfactual impact. Monetary avoided-loss claim on a rulecard whose yield-impact resolved above farm_calibrated (see Avoided-loss and the resolver-layer gate). NDVI-based effect claim on oil palm at prime or rubber under tapping. Any artifact missing a disclosure field per What the agency sees → disclosure envelope.
Confirmations	Publishing an attributable-change artifact for the first time on a field (safety-critical for downstream farm_history calibration per Finding Provenance → Feedback into farm_history). Publishing a counterfactual-impact artifact to an external agency bundle (safety-critical). Overriding the primary baseline with a fallback when the primary could form but the author judges it weak (requires stated reason; enters audit).
Soft warnings	Attribution window closes in the next 72 hours (“close the artifact now or attribution will be refused”). Cohort at k=5 exactly (“any breakdown will drop below k”). Weather confounder marginal (“controlled but the comparison arm’s weather cell disagrees on one week”). Rulecard days_to_mitigate was updated during the attribution window (“window bounds recomputed; earlier draft claims may be stale”).
Rate and scope limits	Cohort formation is capped per scheme per reporting period; the aggregation cost of cohorts scales with cohort size and repeated cohort formation on the same scheme is rate-limited to prevent p-hacking through re-drawing the untreated arm. No hard cap on read paths.
Audit	Every artifact write logs actor, source (system | ui | api | integration), all input-validation fields above, the reason chosen from the no-claim table if applicable, and hashes of the source-event bundles referenced. State transitions on effect artifacts (draft → published → superseded → withdrawn) each write one row per the Audit Envelope.
Escalation	An artifact contradicting a previously published artifact on the same field and cycle (raise_contradiction per Risk Model) escalates to agronomist and freezes both artifacts pending review. A refused counterfactual-impact artifact for a scheme where the agency has a pending disbursement decision escalates to estate_manager for out-of-band communication. Three refused artifacts on the same scheme in 24 hours escalate to org_admin (possible scheme-level data quality problem).

Seed

Fixtures target the observed, attributable, and counterfactual tiers, the no-claim state, and the subsidy case that lands correctly at scheme level only. Fixture ids follow the human-readable convention per Field Data Model; no UUIDs. The eff_ prefix is decided here.

Fixture id	Field(s)	Crop	Tier exercised	Purpose
eff_seed_paddy_bb_recovery_01	field_muda_b1	Rice	Observed change	An map.applied fungicide plan against rice bacterial leaf blight. Attribution window opens 7 days after, recovery to band observed at day 12. No baseline forms (single field, no comparable prior cycle). Ceilings at observed change per Amendment D policy.
eff_seed_palm_ganoderma_priorcycle_02	Oil palm block, blk_semenyih_e3	Oil palm	Attributable change	Sanitation and mounding intervention against ganoderma; prior-cycle baseline forms from the same block’s Production Year 2024 at the same tree-age cohort. Confounders (weather, ecosystem, variety) all match. Reaches attributable change.
eff_seed_rubber_wintering_refused_03	Rubber field, field_ldg_perak_r3	Rubber	No claim	Intervention scheduled in the wintering window; attribution window would overlap with the wintering dip. Rulecard requires the full window. Claim refused with the wintering reason; no-claim artifact published.
eff_seed_pineapple_forcing_cohort_04	Pineapple cohort across batch_johor_p1, batch_johor_p2, batch_negeri_p1 (treated: 6 fields) and deviation-matched untreated cohort of 8 fields	Pineapple	Counterfactual impact	Forcing regime change tested against a matched-batch-age cohort matched on pre-intervention deviation from band at the window’s opening per Cohort selection bias and the fifth key. k=6 treated, k=8 untreated after deviation gate. Cohort reaches counterfactual impact; ecosystem breakdown withheld (would drop below k on one arm).
eff_seed_paddy_avoidedloss_by_region_05	Paddy field, field_kedah_a2	Rice	Attributable change (physical only)	Intervention against a rulecard whose yield-impact resolves at by_region (no farm_history on this hazard yet). Physical response reported (recovery to band, days to recovery); monetary ringgit figure refused per Avoided-loss and the resolver-layer gate. Exercises the resolver-layer gate.
eff_seed_paddy_two_interventions_split_06	Paddy field, field_muda_c4	Rice	Observed change (split)	Two interventions on the same hazard 8 days apart, both inside 2 * days_to_mitigate. Attribution split; each intervention ceilings at observed change alone.
eff_seed_subsidy_scheme_cohort_07	47-field paddy scheme, sch_perlis_2026a	Rice	Counterfactual impact (scheme-only)	Subsidy scheme with 23 treated fields (received input vouchers redeemed for logged application plans) and 24 untreated comparable fields. Field-level subsidy attribution refused per policy. Scheme-level counterfactual-impact claim published with all disclosure fields.
eff_seed_subsidy_scheme_no_untreated_08	30-field paddy scheme, sch_kedah_2026a	Rice	No claim (scheme)	Same shape as fixture 07 but with no comparable untreated cohort available (the scheme covered every eligible field in the district). Counterfactual-impact refused; no-claim artifact published with the k-not-met reason.
eff_seed_palm_yielddelta_reconciled_09	Oil palm estate, est_negeri_op2	Oil palm	Attributable change (yield delta)	End-of-Production-Year reconciliation per Estate Group; yield delta computed against prior Production Year at the same tree-age cohort. Exercises the slowest, strongest field-level effect measure.
eff_seed_weak_cloud_ceiling_10	Paddy field, field_kelantan_b7	Rice	Attributable change (weak)	Corroborating pass was cloud-blocked; verification grade is weak. Claim reaches attributable change but carries the weak flag through the disclosure envelope. Counterfactual-impact promotion refused if fixture 10 is added to a cohort.
eff_seed_paddy_deviation_gate_refused_11	20-field paddy scheme, sch_selangor_2026a	Rice	No claim (cohort → per-field degrade)	Scheme with 12 treated fields at high pre-intervention deviation and 8 untreated fields at low deviation; deviation-matching leaves 3 untreated at the treated arm’s deviation bracket. Counterfactual impact refused per Cohort selection bias and the fifth key; artifact degrades to per-field attributable change on the treated 12.

Coverage properties.

• Every effect measure appears at least once.
• Every no-claim reason appears at least once (attribution window, cohort below k, cohort deviation-gate refused, wintering, resolver above farm_calibrated, two interventions, no untreated arm).
• Every per-crop primary baseline is exercised on its intended crop.
• Weak-grade cloud disclosure travels end-to-end.
• Subsidy is exercised in both the possible-cohort case and the no-comparable-untreated case; field-level subsidy attribution never appears.
• Cross-crop rollup is not seeded; the refusal is structural and is not exercised through fixtures.

Docs Delta

Amendment D is silent, ambiguous, or contradictory against the current docs in the places below. Each item is enumerated, not resolved silently.

Id	Where	Gap
AD-01	Verification Model states its claim as the operation is logged and cross-checked.	Amendment D reads that as the ceiling on did-it-happen claims and defines a separate track for did-it-help. The verification-model page should reference the effect-methodology page as the sibling module, and should state explicitly that verification does not claim outcome.
AD-02	Aggregation Model defines five categories but does not name a cohort-breakdown k-floor rule.	Amendment D states one: once a cohort is defined at k, breakdowns inherit the k floor. The aggregation-model page should incorporate this as a clarification of the k-anonymity rule, not a new category.
AD-03	Risk Model farm_history layer learns from what happened on a farm.	Amendment D’s attributable-change claims are the input to farm_history learning; the risk-model page should state that only attributable-change and counterfactual-impact tiers feed farm_history, and observed change alone does not. This aligns with the Finding Provenance → Feedback into farm_history rule requiring corroboration before calibration.
AD-04	Risk Model resolver layer is the trust score for avoided-loss and the resolver already exposes it (farm_calibrated → by_region → by_ecosystem → literature_v0).	The risk-model page should state explicitly that the resolved layer is downstream-consumable as a trust score, and that the effect methodology gates monetary avoided-loss on farm_calibrated. This is a cross-reference, not a new field.
AD-05	Seasonal Analysis presents index-vs-yield correlation with R² and a residual map at field level.	This is a descriptive, not causal, analysis. The seasonal-analysis page should state its relationship to effect methodology: it is a correlation exercise, not an attribution one, and its outputs feed Amendment D as inputs (evidence for observed change) rather than as effect claims themselves.
AD-06	Estate Group reconciles forecast against actual harvest at field level.	The reconciliation is the raw material for yield-delta claims. The estate-group page should expose the reconciliation output as a formal source for Amendment D’s yield-delta effect measure, including the reconciliation timestamp and confidence.
AD-07	Crop Cycle Models enumerates paddy Main and Off seasons, oil palm Production Year, rubber tapping panels, and pineapple Planting Batches, but does not enumerate attribution windows.	The per-crop ceilings on attribution windows in The attribution window belong on the crop-cycle-models page as a companion table, so any consumer building an effect artifact resolves the cap without cross-reading Amendment D.
AD-08	Finding Provenance defines corroboration for farm_history feedback (Feedback into farm_history section).	Amendment D and Amendment C both bear on farm_history calibration. The finding-provenance page’s corroboration rule and Amendment D’s attributable-change tier are compatible but the interaction (a corroborated finding is an input to an attributable-change claim; the effect artifact then feeds farm_history) should be stated on both pages so the calibration pipeline is not implicit.
AD-09	Semai Advisor overview does not describe advisor behaviour when a user asks for an effect claim that would land in a no-claim state.	The advisor should decline to fabricate an effect and instead return the no-claim artifact, per Failure Modes → Level 2 (Downgrade artifact). The advisor overview page should reflect this explicitly.
AD-10	Audit Envelope does not list the effect-artifact fields (claim_tier, cohort_id, treated_arm_k, untreated_arm_k, attribution_window, no_claim_reason).	The envelope is extensible per its per-module extension pattern; Amendment D names the extension fields.
AD-11	Notification Preferences does not describe delivery for effect artifacts.	Attributable-change and counterfactual-impact artifacts are delivered to different roles than raw verification bundles; the preferences page should account for this. In particular, no-claim artifacts should be delivered to the agency contact who requested the claim, not suppressed.
AD-12	Role Model does not name authoring authority for effect artifacts.	Amendment D’s confirmations (publishing attributable-change; publishing counterfactual-impact to an external agency) imply role gating that the role-model page should surface. Draft resolution: agronomist publishes attributable-change; estate_manager or regional (agency officer) publishes counterfactual-impact bundles externally.
AD-13	Indices enumerates NDVI, NDRE, NDWI and others but does not state the structural saturation ceiling for NDVI on prime oil palm and tapped rubber.	The saturation precondition in Saturation is a structural precondition rests on this. The indices page should state that NDVI is not an admissible effect-measurement index for oil palm at prime or rubber under tapping, and point at NDRE as the required alternative.
AD-14	Risk Model computes deviation from band on every field but does not expose it as a downstream key for cohort formation.	The fifth cohort comparability key in Cohort selection bias and the fifth key depends on the deviation value being consumable at the same call site as the four Amendment A keys. The risk-model page should state that deviation is a public, downstream-consumable field of the resolver, not just an internal computation.

Open Questions

These are decisions the literature did not resolve. Amendment D enumerates them; it does not answer them.

1. by_region admissibility for monetary avoided-loss with a widened error band. Avoided-loss and the resolver-layer gate admits monetary figures only at farm_calibrated and refuses them at by_region, by_ecosystem, and literature_v0. Whether by_region should be admissible with a stated widened error band (for schemes in regions where enough estates have calibrated but this specific farm has not) is undecided. The narrower question replaces the earlier open question about the trust-score shape, which is closed: the risk-model resolver layer is the trust score, and it is already graded.
2. Cohort formation cadence. Cohorts can be formed per-artifact (each effect claim draws its own untreated arm at publication time), per-reporting-period (the scheme’s untreated arm is fixed at the start of the reporting round), or maintained continuously (the platform holds a scheme’s cohorts as a first-class object). Each has integrity and cost trade-offs; the current draft is silent on which is default.
3. NADMA disaster relief as a fifth agency use case. Subsidy bodies and takaful insurers fit the intervention-effect model cleanly. NADMA relief is a different shape: the intervention is disbursement after damage, and the counterfactual (would the recipient have recovered without relief) is much harder to construct against a cohort because comparable untreated cases in a disaster zone are politically and ethically constrained. Amendment D does not resolve the disaster-relief attribution model; a companion amendment may be needed.
4. Interaction with human-raised findings per Finding Provenance. A scout.completed from a human finding may be an intervention (it changed a decision) or it may be corroboration (it confirmed a machine finding without changing the state). The current draft distinguishes by whether a rulecard-firing state changed; whether a human finding on its own (no machine pairing) can support an attributable-change claim without a paired rulecard is undecided.
5. Yield-delta reconciliation timing. Estate Group reconciliation lands at end of cycle. For long-horizon crops (oil palm production year, rubber tapping year) the yield-delta effect artifact is a year late. Whether an interim reconciliation checkpoint (mid-cycle FFB yield for oil palm; mid-tapping-year latex output for rubber) is admissible as a partial yield-delta claim is undecided.
6. Cross-scheme aggregation for national dashboards. A national dashboard aggregating counterfactual-impact claims across many schemes must respect k on each contributing scheme and preserve the minimum-of-inputs confidence. Whether a national-level cohort (drawing untreated arms across schemes) is admissible or whether national claims are strictly summations of scheme-level claims is undecided; the current draft implies the latter but does not state it.
7. Retroactive attribution. An intervention logged and cross-checked months after the fact (a late-log per Verification Model) may still fall within an admissible attribution window on paper. Whether late-log interventions can support attributable-change and counterfactual-impact claims, or only observed change, is undecided.
8. Deviation bracket widths per crop. Cohort selection bias and the fifth key requires deviation-matching at brackets defined per crop; the current draft uses 0.5σ of the crop’s band width as the working figure but the exact widths are a Risk Model calibration that should be established against real cohort formation attempts across paddy and pineapple schemes.

Related

• Verification Model — the did-it-happen track this page’s did-it-help track sits alongside.
• Aggregation Model — the rollup discipline effect claims obey; the source of the k-anonymity and never-average-an-average rules.
• Risk Model — the source of the rulecard days_to_mitigate, the yield-impact curve, and the farm_history layer effect claims feed.
• Crop Cycle Models — the per-crop cycle geometry that caps attribution windows.
• Finding Provenance — the source of the corroboration rule for farm_history feedback, which composes with Amendment D’s attributable-change tier.
• Seasonal Analysis — the descriptive index-vs-yield correlation whose outputs feed effect claims as evidence.
• Estate Group — the forecast-vs-actual reconciliation that yield-delta claims rest on.
• Semai Advisor overview — the safety floor effect artifacts obey, and the advisor behaviour when a claim would land in a no-claim state.
• Guardrails Template — the eight categories this page fills.
• Role Model — the authoring authority effect artifacts require.
• Audit Envelope — the envelope every effect artifact extends.