Proposed Vietnam validation protocol
Planning document · 7 September 2026 · no partner commitment or validated field result.
Scope and start gate
Target 200–300 unique hectares across approximately 150–250 plots and one or two cooperatives in An Giang or Dong Thap. Target 300 ha means 600 hectare-seasons over two crops. Parcel size and farmer count are not interchangeable; confirm both at enrolment. Begin the nine-month schedule only after a cooperative, research collaborator, crop calendar, safe water-management plan and data agreements are in place. A missed planting window can extend the schedule.
LuaZero is an evidence tool. A local agronomist and farmers make irrigation decisions. Participation must not require unsafe drying, surrender of land rights or a promise of carbon income.
Work packages and acceptance evidence
| Window | Deliverable | Acceptance evidence |
|---|---|---|
| M1–2 | Partner setup, lawful data plan, boundaries, local reporting, 20 loggers | Signed agreements; consent records; field inventory; calibrated equipment; approved sampling protocol |
| M3–5 | First-season paired observations and emissions references | Acquisition inventory; raw logger exports; timed farmer records; laboratory QA logs; calibration-only split |
| M6–8 | Frozen-model evaluation in season two | Version freeze; held-out fields/times; confusion matrix; coverage and error reports; staff-time records |
| M9 | Monitoring package, external pre-assessment, go/no-go | Written reviewer feedback; unresolved-issue register; measured cost model; buyer renewal decision |
Satellite and water-level validation
- Confirm actual Sentinel-1 coverage and usable acquisition dates. Do not promise a pass over every field every six days. A first check over An Giang on relative orbit 18 descending returned 30 acquisitions in seven months with a median gap of 6 days but 7 gaps of 12 days, which is the pattern the monitoring design has to survive; see `satellite-observed-revisit.md`. Document resolution, effective valid pixels and edge exclusions; small irregular parcels can be unsuitable.
- Predefine farmer-report, satellite and logger time-alignment windows. Keep acquisition time separate from reporting time. Do not equate one instantaneous observation with a week-long water regime.
- Collect local water-table depth and canopy/crop stage. Use agronomist-defined labels; satellite backscatter alone does not establish safe subsurface water depth.
- Stratify by crop stage, irrigation pattern, residue practice and relevant field conditions. Place 20 loggers across representative fields; revise allocation if stratification is inadequate. Twenty devices are a budget assumption, not a statistical power calculation.
- Separate training/calibration from evaluation at field and season level. Freeze preprocessing, model, threshold and handling of missingness before the independent test. Prevent adjacent observations from the same field leaking into both splits.
- Report balanced accuracy (target ≥85%), class-specific precision/recall, false-drainage acceptance, confusion matrix and uncertainty. Bootstrap by field where appropriate; correlated observations are not independent samples.
- Report usable coverage against all required windows (proposed target ≥70%). Missing imagery, low confidence, contradictory logs and tiny unusable parcels remain in the exception register.
- Investigate conflicts with a field officer; preserve original records and dated corrections. “All records aligned” triggers human review, never automatic credit issuance.
Emissions and agronomic validation
A qualified research collaborator designs reference-field measurements and laboratory QA, including the relevant baseline/project replication per stratum and seasonal sampling intensity. The initial allowance assumes a narrow, scientifically defensible deployment; additional strata require a revised budget. Seek at least three baseline and three project reference fields per stratum where the selected measurement approach requires it. This is not a substitute for a complete sampling and power assessment.
Compare modelled and reference seasonal reductions at the appropriate aggregation level. Target ±20% relative difference only where the measured reduction is materially above zero; otherwise use absolute error and interval overlap. An estimate falling within a ministry-reported range is not validation. A written verifier pre-assessment is not certification.
Track yield, residues, nitrogen inputs and SOC-conserving management. A local agronomist specifies safe thresholds and stop/intervention rules for stress, salinity, irrigation failure or adverse yield signals. The research design must support interpretation of variability. The 2026 SOC rule update must be reviewed in the monitoring design. Do not count SOC gains as rice-methane credits.
Adoption, cost and equity
Target ≥70% of plots logging in at least 10 of 12 weeks and <20% participant drop-off between seasons. Measure officer-assisted entries separately. Record farmer minutes, officer minutes and follow-up trips; do not assume every household has reliable connectivity or owns a suitable device.
Obtain understandable Vietnamese consent and document withdrawal, access, correction, retention and benefit-sharing arrangements. Do not collect land certificates or identity documents unless necessary and lawfully handled. The public demo must remain synthetic. Do not promise payment before a buyer/programme agreement.
Separate research, establishment and recurring delivery costs. The full USD 55,000 pilot over 600 hectare-seasons equals USD 91.67/ha-season, before imputed founder time. A scale target below USD 5 must be demonstrated through measured recurring costs and sensitivity to field visits; it is not the pilot average. Seek a paid-renewal decision from the prospective buyer, with its basis documented.
Proposed budget
| Cost item | USD |
|---|---|
| Reference-field measurements / laboratory | 12,000 |
| Verifier pre-assessment | 8,000 |
| Field coordination, two people, nine months | 9,000 |
| Travel and in-country work | 6,000 |
| Legal / data compliance / contracting | 5,000 |
| Farmer onboarding and training | 4,000 |
| Twenty installed water-level loggers | 3,000 |
| Cloud and processing | 2,500 |
| Contingency | 5,500 |
| Total | 55,000 |
Base USD 49,500; contingency 11.1% of base, or 10% of total. These are planning allowances, not quotations. Founder development is in-kind; log its replacement cost. Device allowance is USD 150 installed per unit and field coordination averages USD 500/person/month; confirm realistic local quotes, maintenance, taxes, travel and replacement requirements before committing the scope.
Failure response
No partner or permission: postpone deployment. Weak satellite performance: retain the field-logging/evidence service and narrow satellite use to triage, or stop that module. Low observation coverage: use ground evidence and report the limitation. Excessive field costs: renegotiate scope/pricing instead of claiming software-only costs. No credible accounting pathway: do not market credit issuance. Unresolved agronomic harm: stop the intervention and investigate under local professional supervision.