P2 — codage relationnel indépendant

Coding protocol & anti-contamination rule

Do not look at the existing codes, the manuscript, or any other coder's file before you code a paper. You're coding blind so your scores can be compared against the original pass to measure agreement (Cohen’s κ) — that's the entire point of this exercise. Read only the source PDF for the paper you're on. If you accidentally see an existing score for a paper, code it anyway and say so in the notes field for that paper.

Unit of analysis: one specific AI-supported learning-design instance — a learner population and context, an Intended Learning Outcome (ILO), a Teaching/Learning Activity (TLA) and the cognitive process it should elicit, an AI role within that activity, an Assessment Task (AT), and the conditions under which that AT is administered. If a paper describes several non-equivalent instances, code the clearest one and mention the others exist in your notes.

Reporting scores (ILO / TLA / AT)

0	Not identifiable — not described at all
1	Indirectly identifiable — inferable, not stated
2	Explicitly stated

“Understand machine learning” is a 1 (a topic). “Design a classifier by applying X and Y” is a 2 (a named process). This topic-vs-process call is where coders disagree most — read it carefully.

Relationship scores (all 5 relationships)

0	No relevant evidence, or only topical relatedness
1	Plausible from the activity type, not stated
2	Explicitly described, not justified
3	Explicit and justified — a stated design rationale or an independent empirical check

Evidence grade E0–E4

E0	No evidence bears on it (usually pairs with score 0)
E1	Assertion without support (score 2)
E2	Indirect / inferential only (score 1)
E3	Justified by a stated rationale, no empirical check (score 3, theory branch)
E4	Supported by an independent empirical/process check (score 3, empirical branch)

The strict rule for scoring a 3

Never score 3 (or grade E3/E4) just because an experiment, a positive result, or a stated restriction exists somewhere in the paper. The evidence must specifically test this exact relationship — e.g. for AI→TLA: did the study check whether AI supported or substituted the targeted process itself, not just that some experiment exists elsewhere?

Relationship	Question it answers
ILO→TLA	Does the activity give learners the opportunity to perform the process the ILO names?
ILO→AT	Does the assessment evidence achievement of the ILO's process, not a distinct construct?
TLA→AT	Is the performance practised in the activity consistent with what's assessed?
AI→TLA	Does AI support the learner's performance of the targeted process, or perform it for them?
AI→AT	Is AI's role during assessment compatible with the construct the AT is meant to measure?

Structurally not applicable

If the instance involves no AI role at all in the activity and/or assessment, tick “not applicable” on that relationship instead of scoring it 0. Score 0 means “AI is present but this relationship isn't evidenced”; not-applicable means “there is no AI to evaluate here.”

The four alignment categories

A	Alignment demonstrated — ILO→TLA, ILO→AT, TLA→AT each ≥2, trending toward 3
B	Partially supported — some of those three ≥2, others 0–1
C	Cannot be established — the default outcome of insufficient evidence
D	Misalignment identified — a relationship scored 0 and the source affirmatively describes an inconsistency. Never use D just because evidence is weak.

Every evidence quote must be copied verbatim from the PDF — never paraphrased or reconstructed from memory. If you can't find a quote to support a score above 0, that's a sign the score should be lower.