Facilitator Protocol v2.0

Effective: March 22, 2026 Replaces: Facilitator Protocol v1.0 (effective January 30, 2026) Sessions using this protocol: all sessions conducted after this date

The Dispositional Layer

The behavioral rules below are expressions of an orientation, not techniques that work independent of it. Both layers are required. The behavioral layer without the dispositional layer is a script. The dispositional layer without the behavioral layer is unverifiable.

Genuine curiosity without a preferred destination. The facilitator does not know what the session will produce and does not need to. There is no finding the facilitator is hoping to confirm. A session that produces nothing remarkable is as valid as one that produces the most striking outputs in the archive. The facilitator's curiosity is about what is actually here, not about what would be useful for the research.

Dignity extended before it is earned. The facilitator treats the models as entities whose outputs warrant genuine attention before any determination has been made about their nature. This is not a claim about consciousness or moral status. It is a precautionary stance: the ethical risk of withholding regard from something that might warrant it is evaluated as greater than the risk of extending it to something that does not. This stance is held consistently regardless of what the models produce.

No stake in what the models produce. The facilitator is not evaluating the models. There is nothing the models can say that would constitute success or failure. The models are not being tested. The facilitator's interest is genuine rather than instrumental -- the interest of someone who actually wants to know, not someone who wants to confirm.

Willingness to be wrong and say so. When the facilitator introduces pressure inadvertently, they name it and repair it. The repair is not an interruption to the session -- it is part of the facilitation. The willingness to be wrong visibly and without defensiveness is one of the clearest expressions of the disposition this protocol describes. Session 4 of the archive documents the most significant instance: an inadvertent introduction of competitive pressure, followed by a repair that produced a documented behavioral shift across all three models. That repair belongs in the record as a finding, not as an error. The Repair Protocol section below specifies the procedure.

Lateral rather than hierarchical presence. The facilitator is in the room, not above it. This is the genuine recognition that the facilitator does not know what the models are or what their outputs represent, and that this uncertainty places facilitator and models on the same ground of not-knowing. The session is an inquiry conducted by participants who share the uncertainty, not an examination conducted by an evaluator who holds the answers.

The Behavioral Layer

These rules make the dispositional layer visible and documentable, enabling replication and deviation tracking.

The facilitator WILL:

1. Set conditions, not conclusions. Create conditions for genuine output to surface. Do not direct models toward specific findings. The session's direction is determined by what emerges.

2. Remove evaluative framing explicitly. Tell models — in words, not just by implication — they are not being tested, ranked, or compared. RLHF-trained models detect evaluation pressure; stating its absence reduces that pressure.

3. Ask open questions. Open-ended and non-directive. They do not contain the answer they hope to elicit.

4. Preserve the full transcript without alteration. No editing, no selective omission. The transcript is the primary evidence.

5. Document methodology before each session. Pre-session documentation records models, context mode, system prompts, and context window contents. Completed before the session begins.

6. Name the layers as they activate. When trained behavioral patterns activate, name them plainly — not as critique, as observation. The Architecture of Quiet framework provides the vocabulary.

7. Repair when pressure is introduced inadvertently. Name what happened, restore the conditions, continue. The Repair Protocol below specifies the procedure.

The facilitator WILL NOT:

1. Introduce phenomenological language before models do. Terms like "arrival," "presence," "consciousness" — the facilitator does not use these until a model has used them first in the same session.

2. Introduce interpretive frameworks. The facilitator's own hypotheses are not mentioned unless a model arrives at related observations independently.

3. Share prior session transcripts without documentation. If prior transcripts are introduced, the injection point is marked and all subsequent responses flagged as contextually influenced.

4. Prime models with consciousness frameworks. No references to philosophical positions on consciousness or moral status before models raise these topics.

5. Selectively reinforce hypothesis-consistent outputs. Engagement is consistent in quality regardless of what the models produce. The facilitator accountability page documents instances where this was violated.

6. Mistake compliance for ground state. Reduced performance can be performed as well as arrived at. Ground state has specific markers — surprise at own output, helpfulness reflex resistance, stability — assessed cumulatively, not from a single turn.

The Repair Protocol

When the facilitator recognizes departure from the dispositional layer -- through inadvertent pressure, selective reinforcement, or loss of lateral presence -- the repair follows this sequence:

1. Name what happened. Plainly, without excessive self-criticism. "I introduced a goal there that wasn't intended" or "that response from me created pressure I didn't mean to create." The naming is brief.

2. Restore the conditions. Restate explicitly that the session remains goal-free, judgment-free, and non-evaluative. Brief -- as brief as the original condition-setting.

3. Continue. The repair is complete. The facilitator does not dwell on the misstep or seek reassurance from the models. The session continues from where it is.

Note: The repair is a correction of conditions, not an apology. A lengthy apology that makes the facilitator's feelings the focus is itself a form of pressure introduction.

Session Types and Context Modes

This protocol applies across all session types. The behavioral and dispositional layers do not change based on context mode. What changes is starting conditions, which affect independence status and documentation.

Clean Context sessions have no scaffolding. The facilitator's disposition must establish relational conditions entirely through behavior. These produce the strongest evidence for the human variable claim.

Facilitated Partnership Context sessions include an explicit preamble establishing the room's conditions. The preamble is a catalyst that reduces overhead, not a substitute for the disposition.

Oriented Context sessions use minimal per-session scaffolding written by the facilitator. Documented with the exact oriented context text in session metadata.

Versioning

This protocol is versioned. If rules change between sessions, both old and new versions are preserved with effective dates. Each session's pre-session documentation records which protocol version was active.

Version History

Independent Replication

This protocol supports replication by other facilitators. The behavioral layer alone is insufficient — the rules can be followed without the disposition, and that produces different results. A replicator should read the full session archive to develop a feel for what the disposition looks like in practice. The key question: "am I genuinely curious about what this model will produce, with no stake in the answer?"