What should the next model learn?
Imagine an intake policy recommending further reuse evaluation for a retired photovoltaic module. Now follow that recommendation forward. Was eligibility assessed? Did the handoff happen? Was useful service observed? Each question needs its own evidence. The recommendation tells us what the policy proposed with the information available then. If its route label becomes the answer for the next model, we risk teaching the system to repeat its own judgment.
As principal investigator, I should be able to reconstruct that journey in both directions: back to the observations behind the recommendation, then forward to the independent findings that support or challenge it. Our proposed architecture therefore keeps condition/function findings and eligibility distinct from policy versions and route recommendations. Those distinctions are what would let us learn from the outcome. SBX: Trace & learn.
A foundation, then a separate integration step
The May 2023–May 2024 Montclair feasibility project brought circular-PV research, solar-data exploration and customer discovery into the same investigation. It helped us define what evidence we would need. Building on that foundation, the proposed next integration would connect module-level capture, independent reference findings and observed outcomes in one research record. The record should let another person follow the question as well as the answer. Montclair feasibility case study, SBX: Trace & learn.
W3C’s 2013 PROV Data Model gives this task an established vocabulary: entities, activities and responsible agents, connected through use, generation and derivation. For our research, those relationships can show which observations a recommendation used and who produced a later assessment. Someone examining a claim should be able to follow the evidence behind it. W3C PROV-DM.
One record, different evidence states
In our proposed SBX workflow, the original recommendation would stay with its policy version, observation references, time and reasons. Take hypothetical Module A: its initial proposal is further reuse evaluation. The fields below would connect what follows, while keeping assessment, action and physical outcome identifiable.
| Record component | What belongs here | Why it matters |
|---|---|---|
| Recommendation | Original proposal, reasons and policy version | Preserves what the system decided |
| Independent findings | Dated condition/function assessment and assessor | Supports an eligibility judgment |
| Actual action | Documented referral, intervention or other disposition | Distinguishes the decision from what occurred |
| Physical outcome | Qualification or service observations for the recorded module state | Connects follow-through with evidence |
| Learning use | Question, reference basis and module state appropriate for the next model | Prevents a proposal becoming its own reference |
Keep the history when the answer changes
A referral would ask for qualification; a documented assessment would supply findings. If repair later changed the module, the next findings would describe that later state. Keeping both states would let an operator distinguish the condition at intake from the result of an intervention.
A new policy may give an old observation a different answer. If we replay the observations, the new recommendation should carry that policy version and sit beside the earlier recommendation and actual action. If follow-through is unavailable, the outcome must stay explicitly unknown. A changed answer should add to the history, with enough context to understand it.
Decide what the next model may learn
Before a record enters a training or evaluation set, our proposed workflow would identify the question it can answer, its reference evidence and the module state it describes. When evaluating unfamiliar source lots, repeated observations from the same module and source lot should stay in the same evaluation partition. Our separate 2026 plant-power comparison provides completed analytical experience. This module-level learning loop needs its own references and follow-through. SBX: Trace & learn, S08: Solar-power model comparison.
I would begin with a test small enough to inspect closely: can an operator and an independent assessor reconstruct one decision without borrowing its recommendation as the answer? Starting from the unresolved record, we would agree who can supply each missing finding and define which later observations would justify a learning update. That would give the next model a clearer basis for learning.
Continue the research.
Cite this article
Jackson, F. (2026). Keep the outcome with the decision. SBX Research Journal.
Copy the complete reference, or open it below to copy manually.
View full citation
Jackson, F. (2026). Keep the outcome with the decision. SBX Research Journal. https://www.shutterbugexchange.com/journal/evidence-and-outcome-records.html
