What the record shows
A governed learning loop can be inspected in operation.
AI proposals, control checks, firmware authority, telemetry, scorecards, lessons, and failures remain connected in a public operating record.
Greenhouse learning-loop proof
A real AI learning loop, exposed to physical consequences.
Verdify Lab is a live greenhouse where AI can reason over operating context and propose narrow tactics, but deterministic controls and firmware retain physical authority. Telemetry, resource use, failures, and outcomes decide what the loop learns next.
The transferable point: the greenhouse is not the client solution. It is a public pattern for keeping knowledge, authority, evidence, and learning connected inside one governed system.
Read the claim correctly
Proof, boundary, and next decision belong together.
What it does not show
The comparison is not isolated causal proof.
The windows were not randomized or weather-normalized. This record does not prove full autonomy, yield optimization, profit optimization, or agronomic causality.
Why it matters
Evidence can decide what the system earns next.
The useful outcome is not that AI is always right. It is that influence is constrained, checked, scored, corrected, and reviewed before scope expands.
Authority separation
Reasoning and physical authority are separate by design.
The lab is useful because the roles are explicit. AI proposes. Controls constrain. People judge. Telemetry verifies. Outcomes become learning signal.
AI proposes. Controls constrain. People judge. Telemetry verifies. Outcomes become learning signal. Every stage has a named owner, a constrained interface, and evidence that can confirm—or contradict—the plan.
-
AI planning
Reads operating context and proposes narrow climate tactics, hypotheses, or no-change intent through approved fields.
-
Control layer
Checks parameters, ranges, required identifiers, ownership, freshness, and unsupported writes before intent moves forward.
-
Firmware authority
Owns relay behavior, applies local safety rules, and controls physical equipment from trusted state. AI must not bypass it.
-
Evidence layer
Records telemetry, delivery logs, readbacks, scorecards, summaries, and lessons so the result stays traceable.
-
Human operations
Sets intent, reviews exceptions, and decides whether the system should tune, hold, stop, or earn more scope.
Full operating comparison
Every metric travels with the window and caveat.
The public comparison shows a different operating record across climate, stress, water, modeled energy, modeled cost, and planner score. It is useful precisely because the tradeoffs remain visible.
Fixed windows
Comparison design
Planner-offline and AI-planning-online periods are published with dates and caveats.Public
Operating context
The lab publishes its planning, climate, safety, evidence, and known-limit surfaces.Dated
Snapshot discipline
This commercial summary records the source and review date instead of presenting live values as static facts.Versioned
Commercial record
The prior May 19 commercial snapshot remains in the repository for comparison and audit history.| Metric | Planner offline | AI planning online |
|---|---|---|
| Average AI plans/day | 0.0 | 3.0 |
| Both-axis compliance | 20.6% | 54.7% |
| Temperature compliance | 44.9% | 70.9% |
| VPD compliance | 30.3% | 74.5% |
| Graded compliance | 63.4% | 67.3% |
| Combined temperature + VPD stress (axis-hours/day; 48 possible) | 29.9h | 13.1h |
| Water/day | 429.6 gal | 240.1 gal |
| Runtime-modeled electric energy/day | 19.1 kWh | 29.1 kWh |
| Modeled operating cost/day (source currency) | 7.89 | 8.31 |
| Planner score | 25.9 | 52.7 |
Comparison window: Planner offline: April 22-25, 2026. AI planning online: April 26-May 2, 2026.
Claim limit: Dated operational comparison with weather, solar, equipment, crop, and instrumentation confounders; not a controlled A/B test or isolated causal proof.
Snapshot: July 10, 2026. Registry: 2026-07-10.commercial-greenhouse-proof.v3.
Inspect the source comparisonWhat transfers
The pattern works wherever a workflow has stakes, authority, evidence, and review.
A greenhouse is physical, but the architecture is familiar: narrow tools, deterministic gates, readbacks, private evals, an exception lane, and feedback that stays with the organization.
Enterprise diligence
AI prepares source-linked evidence; security and legal owners approve outbound claims; reviewer edits improve the answer base.
Product change review
AI assembles traceability; required owners sign off; missing links become visible knowledge work rather than hidden uncertainty.
Operations exceptions
AI prepares deviation evidence and routes the case; authorized reviewers decide; the decision improves future preparation.
Known limits
The evidence is useful because the limits stay visible.
- The planner-offline and planner-online windows are real production periods, but they were not randomized or weather-normalized.
- Shade, outdoor humidity, solar gain, vent capacity, nozzle geometry, sensor placement, and equipment limits still constrain outcomes.
- Light measurement, intake sensing, leaf wetness, root wetness, and some water attribution need improvement before stronger agronomic claims are defensible.
- Timeouts, missed cycles, stale plans, and no-change acknowledgements remain in the operating record because reliability is part of proof.
- The current evidence covers climate, stress, resource use, planning outcomes, and lessons. Yield, quality, and profit require their own evidence layer.
Inspect the record
Follow the evidence to its source.
This page is a dated commercial interpretation. The Live Lab is the technical source of truth for current plans, scorecards, lessons, failures, and known limits.
Live evidence map
Operations, planning quality, resource use, archives, and public-safe exports.
Safety design
Why the planning system cannot control relays or bypass firmware authority.
Planning loop
How context, tool calls, dispatch, scorecards, and lessons fit together.
Baseline comparison
The fixed-window operational comparison with caveats.
Operating caveats
Weather, sensor, equipment, and physical-world constraints behind the numbers.
Lessons
Validated lessons that constrain how the planner should operate next.
From proof to your workflow
Bring one workflow that should become an owned learning loop.
Start with what AI may do, what it must not do, who retains authority, which feedback should be captured, and what evidence would justify expansion.