I. Authority & Judgment

☐ Is the AI output explicitly framed as a claim, not a decision?
☐ Is there a visible pause before action continues?
☐ Is a human required to ratify, revise, or reject the output?
☐ Is the human’s decision recorded (or at least legible)?
☐ Can a human override the AI even if the AI is “right”?

If authority can shift silently, stop.

II. Refusal & Stopping Points

☐ Does the feature have an explicit refusal path?
☐ Does it specify what it will not do?
☐ Can the system intentionally produce no answer, no output, or no resolution?
☐ Is silence treated as a valid state?

If the system must always continue, it is unsafe for Agora.

III. Artifact Persistence (Memory > Interaction)

☐ Does this produce a durable artifact (draft, decision, revision, audio, note)?
☐ Can a teacher/admin review how the artifact came to be?
☐ Are rejected alternatives preserved or at least acknowledged?
☐ Can artifacts be revisited later without re-running the interaction?

If nothing survives the session, learning is fragile.

IV. Ambiguity Preservation

☐ Are multiple interpretations surfaced where plausible?
☐ Is disagreement treated as information, not error?
☐ Does the system avoid collapsing variance into a single “best” answer?
☐ Can unresolved tension remain unresolved without penalty?

If ambiguity disappears too quickly, meaning is being flattened.

V. Teacher / Admin Authority

☐ Can a teacher/admin see what the AI proposed?
☐ Can they see what the user accepted or rejected?
☐ Can they intervene without the AI “arguing back”?
☐ Is evaluation human-led, even if AI assists?

If teachers become invisible, the system is broken.

VI. Pace & Latency Honesty

☐ Does the UI make waiting visible rather than hiding it?
☐ Is slowness explained as intentional when it is pedagogically required?
☐ Is there a clear “who acts next” moment?

If speed is rewarded implicitly, authority will drift.

VII. Emotional Restraint

☐ Does the system avoid emotional mirroring or reassurance optimization?
☐ Is tone professional rather than intimate?
☐ Does the system avoid language that implies care, belief, or intention?

If users feel emotionally carried, judgment is being displaced.

VIII. Avatars / Audio / Video (if applicable)

☐ Is embodiment role-bound rather than persona-driven?
☐ Does rendering require explicit human authorization?
☐ Is AI audio/video presented as one plausible version, not “the right one”?
☐ Can students compare, shadow, revise, and reflect rather than perform?

If embodiment adds authority instead of constraint, pause.

IX. Ingestion Discipline

☐ Is it clear why material is being ingested?
☐ Is ingested content tied to a scenario, task, or decision?
☐ Can ingestion be explained pedagogically, not just technically?
☐ Is there a path to discard or archive ingested material?

If ingestion becomes accumulation, entropy follows.

X. Model Transparency & Pluralism

☐ Is it clear which model produced which output?
☐ Are differences between models preserved rather than averaged away?
☐ Is the system robust to model substitution?

If models become invisible, accountability erodes.

XI. Scope & Claims

☐ Is this an exemplar, not a promise of universality?
☐ Can you clearly state the limits of this module?
☐ Would you be comfortable not scaling this yet?

If the answer requires hype, slow down.

XII. Final Gate Question (non-negotiable)

Ask this out loud before proceeding:

Does this make human thinking harder to avoid — or does it quietly make thinking optional?

Only the first belongs in Agora.