Defensible claim
Verdify can show an auditable AI planning loop in a real operating environment: the agent proposes, control layers inspect, firmware enforces, and telemetry verifies.
Greenhouse case study
A real example of Verified AI in operation.
Verdify Lab is a live greenhouse where an AI agent can reason over operating context, but does not receive direct authority over physical equipment. It is the public proof environment for the Verdify Method: map the workflow, define controls, build the narrow path, verify outcomes, and operate from evidence.
Executive thesis
The point is not that AI runs a greenhouse. The point is that AI can be given a verified operating role.
The greenhouse is intentionally physical: weather changes, sensors drift, equipment wears, humidity moves, water and energy matter, and bad actions have consequences. That makes it a useful test of the business question Verdify is built around: can an AI agent improve a workflow without being handed unsafe authority?
Claim limit
This does not prove isolated causal lift, full autonomy, yield optimization, profit optimization, or agronomic causality. The operating record includes weather, crop, hardware, operator, and instrumentation caveats.
Executive takeaway
The result that matters is not that the model is always right. It is that model influence is constrained, logged, checked, scored, corrected, and reviewed over time.
Verdify Method in operation
The greenhouse is the method, made inspectable.
The lab turns a commercial method into a visible operating system. Each step has an equivalent in the greenhouse and in a client workflow.
| Method step | In the greenhouse | Why an executive should care |
|---|---|---|
| Map | Map the physical workflow: crop needs, weather pressure, equipment state, sensor quality, resource use, operating goals, and failure modes. | The work starts by understanding the real system, not by dropping an agent into a vague process. |
| Define Controls | Define what the AI agent may write, which controls inspect the write path, and which actions remain outside model authority. | AI authority is intentionally narrow before implementation expands. |
| Build | Build a controlled planning path that can write tactical intent or acknowledge no change, with required IDs, accepted fields, and readbacks. | The implementation is an operating system with checks, not a demo script. |
| Verify | Verify with telemetry, delivery status, compliance, stress hours, water, energy, scorecards, and visible caveats. | The question is not whether AI sounds plausible; it is whether the record supports the next operating decision. |
| Operate | Operate through plans, no-change acknowledgements, lessons, known limits, dashboards, and periodic review. | The loop improves only when evidence supports tuning, holding, stopping, or expanding. |
The AI agent plans. Control layers constrain writes. Firmware controls. Telemetry verifies. Scorecards and lessons close the loop.
Authority separation
Reasoning, controls, physical authority, evidence, and human operations are separate by design.
The lab story is strongest when the operating roles are explicit. The AI agent can propose. The control layer constrains. Firmware controls. Telemetry and scorecards verify. Humans inspect the record and handle exceptions.
AI planning role
The AI agent reads operating context and proposes controlled climate tactics, hypotheses, and setpoint intent through approved fields.
Control layer
The write path checks parameters, ranges, required IDs, ownership, freshness, and unsupported writes before a plan can become operational intent.
Firmware authority
ESP32 firmware owns relay behavior, applies local safety rules, and controls physical equipment from trusted state.
Evidence layer
Telemetry, delivery logs, readbacks, scorecards, daily summaries, and lessons record whether the plan worked.
Human operating surface
Dashboards, alerts, briefs, and public pages surface state and exceptions without becoming a hidden actuation layer.
Evidence reviewed
The case study is based on an operating record, not a slideware claim.
The evidence window reviewed for this public case study uses live database state through May 19, 2026, plus the public lab pages and implementation records. These numbers are evidence that the loop is inspectable; they are not a claim of optimal control.
274,281
Climate rows
Sensor and operating records from August 5, 2025 through May 19, 2026.
57
Daily summaries
Operating summaries from March 24 through May 19, 2026.
196
Planning journal rows
AI agent planning journal records, with 190 validated at review time.
48
Planner deliveries
Plan writes and no-change acknowledgements since May 12, 2026.
67
Planner policy controls
Registry parameters available to the planning path; 37 are required in full routine plans.
50
Active lessons
Planner lessons, including 30 medium- or high-confidence lessons.
Operational data
The fixed-window comparison is useful because the caveat travels with the metric.
The planner-offline and AI-planning-online windows show a materially different operating record. Verdify keeps the weather and operating caveats next to the table because executive readers should see both the result and the claim limit.
| Metric | Planner offline | AI planning online |
|---|---|---|
| Average AI plans/day | 0.0 | 3.0 |
| Both-axis compliance | 20.1% | 54.7% |
| Temperature compliance | 45.3% | 70.5% |
| VPD compliance | 30.1% | 74.4% |
| Stress-axis hours/day | 29.8h | 12.9h |
| Water/day | 427.5 gal | 222.7 gal |
| Estimated electric energy/day | 2.6 kWh | 1.2 kWh |
| Operating cost index | Baseline | Lower observed energy-cost load |
| Planner score | 28.0 | 56.9 |
Source: Verdify Lab case study and public evidence. Planner offline: April 22-25, 2026. AI planning online: April 26-May 2, 2026.
Registry: 2026-05-19.commercial-greenhouse-proof.v1. Lab reference elevation: 5,090 ft. Normal planning frequency: Up to 3 operating plans/day.
Caveat: Operational comparison evidence with weather and solar confounders; not a controlled A/B test or isolated causal proof.
Inspect the ComparisonCaveats and known limits
The evidence is useful because the limits stay visible.
A high-trust AI story should not hide the conditions that weaken the claim. In the lab, reliability debt, sensor gaps, weather, and physical constraints stay part of the public record.
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 part of the operating record because reliability is part of proof.
The current public proof is about climate, stress, resource use, planning outcomes, and lessons. Yield, quality, and profit need their own evidence layer.
Business workflow translation
The greenhouse pattern transfers when the workflow has stakes, authority, telemetry, and review.
The greenhouse is a physical case, but the operating pattern is the same one Verdify applies to software, operations, procurement, quality, compliance, support, finance, and field-service workflows.
| Transfer principle | Greenhouse example | Business workflow equivalent |
|---|---|---|
| Typed actions, not broad access | The agent writes through approved planning fields instead of receiving raw relay authority. | Give an agent narrow tools, parameter limits, required IDs, and clear ownership instead of broad system access. |
| Controls before authority | The control layer checks the write path before any plan becomes operational intent. | Check schema, policy, approvals, conflicts, freshness, and role authority before an output affects a system of record. |
| Readbacks over assumptions | The loop records plan delivery, accepted fields, no-change acknowledgements, readbacks, and failures. | Confirm what downstream systems accepted. Treat missing confirmation as an operational event. |
| Scorecards decide expansion | Compliance, stress hours, resource use, delivery status, and lessons determine whether the loop should tune, hold, stop, or expand. | Score AI-assisted work against service levels, cycle time, error rate, compliance exceptions, customer impact, or other workflow outcomes. |
| Humans keep the exception lane | Dashboards and alerts surface state and exceptions; they do not hand physical authority to the model. | Use AI to summarize, route, and prepare exceptions while accountable people approve high-risk or ambiguous decisions. |
| Evidence beats trust | The Live Lab publishes plans, scorecards, caveats, lessons, known limits, and public-safe operating records. | Create a proof surface stakeholders can inspect: logs, decisions, source traces, metrics, caveats, and known limits. |
Enterprise diligence: AI prepares evidence; security/legal approves outbound claims.
Medtech design change: AI assembles traceability; QA/RA signs off.
Recall readiness: AI reconciles lot records; recall owners decide action.
Cleantech pilots: AI packages proof; technical and commercial owners approve claims.
Aerospace NCR/MRB: AI prepares deviation evidence; authorized board decides disposition.
Operating lesson: separate reasoning from authority, then measure the result.
Inspect the source of truth
The web page matters because the proof keeps changing.
A static PDF would go stale as soon as the next plan is written or acknowledged. The Live Lab keeps plans, evidence, caveats, lessons, and known limits inspectable.
Live evidence map
Operations, planning quality, resource use, archives, and public-safe exports.
Safety design
Why the AI agent does not 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 client workflow
Bring Verdify one workflow where AI needs a verified role.
A useful first conversation starts with the workflow, what AI might do, what it must not do, which systems or people stay authoritative, and what evidence would justify expansion.