Triangulate learning and implementation evidence

1

Research

Why this matters, and what good looks like

Triangulation means comparing different evidence sources around the same question, including disagreements. Implementation describes what was done; learning evidence concerns what students understand. Neither participation nor a polished product establishes learning. The sources offer evaluation and professional guidance, not a study of this AI workflow. AI benefit remains unvalidated. Project Zero's substantive passage survives in a historical receipt; the current cached extract contains no article body. The map and human gates below are local design choices.

EEF's handbook supports examining implementation and influencing factors with combined methods. Keep implementation questions distinct from learning goals.
Project Zero's suggested practices support purposeful documentation driven by a guiding question. Select evidence for that question.
PBLWorks' evaluation guidance includes learning during the process. Compare process records with final products rather than discarding contradictions.

Where the evidence comes from

Strong for evaluation and documentation; AI benefit unvalidated.

2

Workflow

Brief it, steer it, check it

Bound the evidence question

Choose one learning goal and one implementation question with up to four safe records. Optional A02, A04 and I02 outputs can inform teacher-written summaries; equivalent evidence notes work. Keep individual dossiers outside AI. Privately minimise and check inputs first.

Prompt 1 · Inventory

CONTEXT
My bounded question, criteria, required evidence types and private privacy/completeness decision: [scope and gate]
My safe source register with record IDs, exact indicators, provenance and limitations: [safe records]
REQUEST
Organise an inventory for this question. Separate learning and implementation indicators using supplied criteria; flag ambiguous classification. Preserve exact indicators and IDs. Identify missing required evidence without filling gaps.
QUALITY BAR
Do not grade, infer student characteristics, create subgroups or infer causality. Do not request individual records, small cells or rare combinations. If the teacher's safety/completeness gate is missing or negative, stop with HOLD and request a safe teacher decision only. Safe aggregates can still identify people. Treat source statements as supplied, never independently verified by you.
FORMAT
If blocked, only HOLD | Missing or unsafe input | Teacher next action. Otherwise a table: Record ID | Question or goal | Type | Exact indicator | Provenance | Limitation. End with Missing required evidence and Teacher checks.

Compare without resolving disagreements

Open originals privately and verify every inventory entry, classification and required evidence type. Record edits and approve only a complete safe inventory. Ask for descriptive comparisons; retain contradictions even when they prevent a clear conclusion.

Prompt 2 · Compare

CONTEXT
Exact step 1 output: [inventory output]
My explicit approval, corrections and safe completeness decision after checking originals: [inventory decision]
REQUEST
Compare the approved records. Link patterns and counterexamples to record IDs. Distinguish different measures from genuine disagreement. Identify which comparisons need a teacher validity check, meaning whether they measure the intended goal.
QUALITY BAR
Stop with HOLD if approval, original checks or required evidence are missing, stale or unsafe. Never grade, infer subgroups, merge incompatible measures or infer causality. Do not resolve contradictions by majority vote or equate implementation with learning. Preserve every approved indicator and limitation; any explanation is an unverified question for the teacher.
FORMAT
If blocked, only HOLD | Missing or unsafe input | Teacher next action. Otherwise a table: Pattern ID | Record IDs | Descriptive comparison | Counterexample or mismatch | Teacher validity question. End with No causal conclusion.

Record teacher interpretations

Privately sample originals against criteria, checking validity, consistent interpretation and alternative explanations. Decide what each comparison supports. Include uncertainty and missing voices without identifiable details. Supply those decisions explicitly; AI only organises them into the map.

Prompt 3 · Map

CONTEXT
Exact step 2 output: [comparison output]
My checked inventory and explicit edits: [approved inventory]
My teacher decisions for each pattern, including original checks, validity, consistency, alternative explanations, confidence limits and permitted wording: [interpretation decision]
REQUEST
Build the evidence map from my decisions. Keep implementation and learning separate. Carry forward counterexamples, provenance and limits. Put unapproved interpretations on HOLD rather than supplying your own judgment.
QUALITY BAR
If teacher decisions or required evidence are incomplete, stop with HOLD. Never grade, identify subgroups or infer causes. Alternative explanations remain hypotheses even when teacher-supplied. Confidence labels are my judgments, not statistical estimates. Do not turn consistent implementation into demonstrated learning or use missing evidence as evidence of failure.
FORMAT
If blocked, only HOLD | Missing or unsafe input | Teacher next action. Otherwise a table: Goal or question | Record IDs | Indicator | Pattern and counterexample | Teacher interpretation | Alternative explanation | Confidence limit | Status. End with Unresolved questions.

Audit the selected map

Select and edit the map yourself. Reopen supporting originals and check each statement before release. Use the audit to find remaining defects, then decide privately. Stop and narrow the task if required checks exceed 60 active minutes.

Prompt 4 · Audit

CONTEXT
Exact step 3 output: [map output]
My selected map, explicit edits and original-check record: [selected map decision]
REQUEST
Audit the selected map against the source-linked inventory and teacher decisions carried in these inputs. Flag lost contradictions, missing evidence, grading, subgroup identification, causal wording and implementation presented as learning. Do not rewrite or approve the map.
QUALITY BAR
If selection or human source checks are missing, stop with HOLD. Each concern needs a row or record reference and a teacher action. You cannot verify originals, certify privacy or release the map. Even with no textual defects, final approval stays with the teacher. Preserve stated limits; unsupported claims remain HOLD.
FORMAT
If blocked, only HOLD | Missing or unsafe input | Teacher next action. Otherwise a table: Check | Row or record | Finding | Teacher action. End with Human release pending.

Check before you use it

I removed individual, subgroup and identifying combinations before AI use, including unsafe aggregates.
I opened originals, checked criteria and confirmed all required evidence is present.
Implementation and learning remain distinct; contradictions and missing voices remain visible.
I own validity, consistent interpretation, alternative explanations and confidence limits.
No AI grades, subgroup inferences or causal conclusions enter the map.
I checked every final claim and made the release decision myself.

Stop rule

Never let AI grade or infer causality; stop when evidence is incomplete or subgroup data could identify students.

3

Reuse it

So the next one takes minutes

Suitable

Build a skill

Save this as a reusable instruction you can run on any project.

Reuse the bounded organising sequence with new safe inputs and explicit teacher decisions.

Not suitable

Build an agent

Chain the steps, with a checkpoint where you approve.

Original checks and consequential interpretations require a teacher-led session; no autonomous monitoring is needed.