confidence-engine/docs/behaviour-selection.md

# Behaviour Selection — v0.1 Implementation Brief

> **Status: Design only.** Experiment 19 pending. This brief is a constraint on the experiment, not an architecture.

---

## The Problem (Discovered)

Experiments 1–14 proved that the workspace layout is stable and the reasoning engine works. What they revealed but could not fix:

> The current engine behaviour is: **ask → wait → ask → wait**. Every turn produces a question. This makes the investigation feel like automated Q&A rather than guided thinking.

The user's framing from Exp 15: *"An expert consultant does not have a script. They have behaviours — recurring patterns of action that they deploy based on what they observe."*

This experiment tests whether adding **behaviour selection** between assessment and conversation changes that pattern in a meaningful way.

---

## What We Can Measure Now (From Exp 18)

The assessor produces three reliable dimensions:

| Dimension | What it tells us | Available now? |
|-----------|-----------------|----------------|
| Phase | Where the investigation is (orienting → concluding) | ✓ |
| Progress | Whether understanding is advancing (accelerating/steady/stalled) | ✓ |
| Conversation Health | Whether the interaction pattern is productive (healthy/too_broad/too_narrow) | ✓ |

These are sufficient for a first test. We do not need evidence quality, uncertainty trend, or understanding trajectory yet.

---

## v0.1 Behaviour Set: Five Patterns

The smallest useful set that covers the gap between "always asking" and "facilitated thinking":

| Behaviour | When to deploy | What it does |
|-----------|---------------|--------------|
| **Acknowledge** | Any turn where user provided useful information (at least one resolved node) | State what was learned; do not immediately ask a new question |
| **Clarify** | Conversation health is `too_broad` or phase is `orienting` with insufficient data | Ask for a single specific piece of context, not an unknown-node query |
| **Summarise** | Phase is `synthesising` or `concluding`; or ≥3 turns have passed without summarisation | Restate current understanding; compress without losing detail |
| **Continue** | Default — no other behaviour matches | Ask the next useful question (current behaviour, but made explicit) |
| **Pause** | Phase is `focusing` with stalled progress | Hold space; acknowledge what was learned; invite reflection rather than asking for more |

Every turn must select exactly one of these. No combinations, no secondary actions. The test is: does *choosing* change the pattern?

---

## Selection Rules (One Rule Per Behaviour)

These are plain conditions with no scoring, no weights, no convergence:

1. **Acknowledge triggers** if `conversation health == healthy` AND at least one node was resolved this turn
2. **Clarify triggers** if `conversation health == too_broad` OR `phase == orienting` AND observations < 3
3. **Summarise triggers** if `phase == synthesising` OR `phase == concluding` OR (turns ≥ 3 AND no summarisation in recent turns)
4. **Pause triggers** if `phase == focusing` AND `progress == stalled`
5. **Continue** is the default — use it when none of the above match

If multiple rules fire simultaneously, priority is: Acknowledge > Clarify > Summarise > Pause > Continue. No convergence required. If two conditions are equally relevant, pick the one that adds *information* rather than the one that asks for more input.

---

## What v0.1 Does NOT Do

These are intentional exclusions — not deferred features:

- **No scoring or weighting.** A condition either matches or it doesn't.
- **No "convergence" threshold.** If two dimensions trigger, pick by the priority rule.
- **No evidence quality or uncertainty trend integration.** We don't have that data yet, and we don't need it for this test.
- **No stable behaviour pairing.** Acknowledge replaces "acknowledge + communicate confidence." One action per turn.
- **No rationale output or developer view.** That's infrastructure, not signal.
- **No phase-constrained allow/block tables.** The rules above *are* the constraints.

---

## Success Criteria

How we know whether behaviour selection is worth continuing:

1. **Behaviour diversity:** Does the system deploy at least 3 different behaviours across a normal investigation, or does it default to Continue (Continue) most of the time?
2. **Acknowledge appears:** Does Acknowledge fire whenever new information resolves an uncertainty? If not, the trigger condition is wrong — fix it, don't abandon selection.
3. **Pause feels like relief, not delay:** When Pause fires, does the user experience it as a natural break rather than a system failure to produce a question?
4. **Summarise compresses meaningfully:** Does the summarised understanding feel useful (new synthesis) or redundant (restatement of what's already on screen)?
5. **Conversation rhythm changes:** Is there a perceptible difference between "engine always asking" and "engine sometimes acknowledging/summarising/pausing first"?

If none of these can be evaluated after 2-3 real investigations with v0.1, the experiment was too small to answer the question. Expand the behaviour set or extend the test — don't abandon selection.

---

## Future Considerations (Not In v0.1)

| Concept | Status | Why deferred |
|---------|--------|-------------|
| Signal weighting / scoring formula | Speculative | No observed basis; inventing numbers |
| Convergence requirement | Speculative | Design preference, not discovery |
| Behaviour Readiness derived dimension | Architecture housekeeping | Useful later if v0.1 validates the approach |
| Full 14-behaviour inventory | Available but not tested | Expand only after v0.1 proves the mechanism works |
| Rationale output for developer view | Infrastructure | Not signal; can be added post-validation |
| Multi-dimensional contradiction detection | Complex, unneeded yet | One rule per behaviour is simpler and testable |
| Phase-specific allow/block tables | Invented constraints | Rules above *are* the constraints |

---

## Evaluation Criteria for Behaviour Selection

How we know a behaviour is working? Not through visual metrics, but through conversational quality:

1. **Does each turn feel like it builds on the previous one?** (Continuity)
2. **Does the user understand why they are being asked what they are being asked?** (Purpose)
3. **Does the investigation feel guided rather than mechanical?** (Direction)
4. **Does the user feel understood, not just processed?** (Respect)
5. **Does uncertainty feel honest, not manufactured?** (Trust)
6. **Does progress feel real, not illusory?** (Substance)

These six criteria come directly from `facilitator-behaviour.md` (Experiment 14). They apply to all experiments that touch conversation behaviour.

---

## Relationship to Assessment Layer

Behaviour Selection does not replace the assessor. It *consumes* it.

| Assessment Dimension | How Selection Uses It |
|---------------------|----------------------|
| Phase | Determines which behaviours are available (orienting → acknowledge/clarify; synthesising → summarise) |
| Progress | Stalled progress in focusing phase triggers Pause instead of Continue |
| Conversation Health | `too_broad` triggers Clarify; healthy with resolution triggers Acknowledge |

If the assessment contract changes, selection rules that read those dimensions must be reviewed. The mechanism (plain condition matching, no scoring) is stable regardless.

---

## What This Experiment Proves

One question: **Does selecting from a small set of behaviours — instead of always asking — make the investigation feel more like guided thinking and less like automated Q&A?**

If yes: expand the behaviour set and test refinement in v0.2.
If no: either the five behaviours are wrong (not selection itself) or the engine's core loop needs a deeper change than this layer can provide.

Nothing else matters until this is answered.