confidence-engine/docs/archive/experiments/pre-RTO/v0.6-atomicity-experiment.md

v0.6 Atomicity Experiment

Hypothesis

After deterministic unknown selection, the engine should assess whether the selected unknown is already atomic or is still too composite to ask directly.

If the unknown is atomic, the engine should proceed exactly as before.

If the unknown is composite, the engine should not ask that parent unknown directly. Instead, it should decompose it into a small set of explicit child unknowns representing broad, independent candidate dimensions that a non-expert could understand.

Constraints

• No graph redesign
• No persistence
• No UI redesign
• No selection-weight tuning
• No Ollama calls in unit tests

Deterministic rule introduced

Atomicity assessment is not a new investigation strategy.

It runs in the graph update path at this seam:

unknown selection -> atomicity assessment -> optional decomposition -> deterministic reselection -> question formulation

The implementation uses deterministic text and graph-shape checks:

• focused unknowns like denominator / threshold / definition / baseline / evidence remain atomic
• broad relationship-explanation unknowns and broad “possible causes / what changed / explanation for why X but Y” unknowns become composite

Decomposition behavior

When a selected unknown is composite:

1. The parent unknown remains unresolved.
2. Between 2 and 5 child unknowns are created or reused deterministically.
3. Children become explicit graph nodes.
4. Children link back to the parent with existing depends_on edges.
5. Children inherit the same “why it matters” discipline in their descriptions.
6. Deterministic selection reruns across the updated graph.

For the current relationship-explanation experiment, the broad child dimensions are:

• Whether the two observations reflect different timing
• How the two observations were measured
• Change affecting signal A more than signal B
• Change affecting signal B more than signal A
• One-off event during the period

These are intentionally non-jargon and broad enough to generalise across scenarios like:

• Revenue up / Cash down
• Customer satisfaction up / Complaints up
• Delivery time down / Cancellations up
• Traffic up / Sales flat
• Production up / Defects up

Diagnostics added

The orchestrator now reports:

• atomicityAssessment
• atomicityDecisionReason
• decompositionDepth
• decompositionAttempted
• decompositionAccepted
• decompositionStoppedReason
• proposedChildCount
• acceptedChildCount
• rejectedChildren
• selectedChildNodeId
• childQualitySummary
• propagationPerformed
• resolvedChildNodeId
• parentNodeId
• parentStatusBefore
• parentStatusAfter
• parentConfidenceBefore
• parentConfidenceAfter
• affectedAncestorIds
• nextSelectedSibling
• parentResolved
• decompositionPerformed
• childUnknownCount
• childNodeIds
• atomicityReason

This sits alongside the existing explicit-emergent-unknown diagnostics.

Observed outcome

The experiment was useful.

Before this change, the engine could select a broad explanation unknown and ask it directly.

After this change:

• the broad explanation parent remains explicit in the graph
• the engine decomposes it into child unknowns first
• the next asked question is backed by a more focused child unknown
• repeated updates reuse the same decomposition children deterministically
• child-quality checks reject compound or duplicate children before they enter the graph
• decomposition stops deterministically once a selected child is directly answerable
• resolving one child does not resolve the parent immediately
• resolved child evidence now propagates upward to the parent and ancestor chain deterministically
• parent status and confidence change conservatively after child resolution
• the next sibling becomes eligible for normal deterministic selection without recreating the resolved child

In the revenue-versus-cash case, the selected next question becomes:

What evidence would clarify how the two observations were measured?

rather than asking the full broad explanation node directly.

Upward propagation and reconstruction

Recursive reasoning is complete only when decomposition and reconstruction are both deterministic.

Confidence must not outrun completeness or evidence.

For this experiment, reconstruction now behaves as follows:

• when a child unknown resolves, that child keeps its own resolved status and answer evidence
• the parent is updated, but remains unresolved unless the deterministic completion rule is satisfied
• only the ancestor chain connected to that child is updated
• unrelated branches remain unchanged
• the deterministic selector then chooses the next justified unresolved sibling or related follow-up

For the current conservative completion rule:

• one resolved child → parent becomes provisional with higher confidence, but remains unresolved
• all direct child unknowns resolved → parent resolves deterministically with high confidence

The confidence model is now explicitly separated into:

• evidence confidence: how trustworthy the currently attached support is
• completeness: whether the required direct child structure is empty, partial, or complete
• conclusion confidence: how strongly the current parent state is justified given both evidence and completeness

Deterministic propagation rules now enforce:

• one resolved child may raise evidence confidence
• unresolved direct children cap conclusion confidence
• contradictory direct children block high conclusion confidence
• duplicate evidence does not increase confidence
• status changes do not raise confidence on their own
• parent resolution still requires the separate completion rule

Cross-branch corroboration

The next confidence experiment adds deterministic branch interaction checks without changing the graph model.

The engine now distinguishes between:

• multiple evidence: more than one branch exists
• independent corroboration: distinct resolved branches support the same parent without sharing the same evidence key
• duplicate evidence: the same evidence key appears through multiple branches and must not be double-counted
• conflicting evidence: branches support incompatible positions, such as recognised correctly vs recognised incorrectly

Deterministic branch rules:

• corroboration only counts when branches are distinct and their evidence sources differ
• duplicate evidence groups never count as corroboration
• conflicts cap conclusion confidence and prevent a higher confidence upgrade
• independent branches remain interaction-neutral

Additional diagnostics now expose:

• corroboratingBranchCount
• conflictingBranchCount
• duplicateEvidenceCount
• independentBranchCount
• interactionSummary
• confidenceAdjustmentReason

Observed effect:

• independent corroboration can raise evidenceConfidence
• duplicate evidence produces no extra confidence increase
• conflicting evidence lowers or caps conclusionConfidence
• completeness rules still dominate whether a parent may become highly justified

Example progression:

• parent before: unknown, medium
• after resolving How the two observations were measured: parent becomes provisional, medium
• evidence confidence becomes high, completeness becomes partial, conclusion confidence becomes medium
• next sibling becomes selectable and the engine moves on without recreating the resolved child

Interpretation

This supports the idea that recursive decomposition is a fundamental part of graph-backed questioning, not just a prompt refinement.

The main remaining limitation is that sibling selection still inherits the existing deterministic scorer. That means some domains may advance to a justified sibling that is not the intuitively expected next child, even though the propagation itself remains deterministic and graph-valid.

Validation run

Covered by:

• tests/graph/atomicity-assessment.test.js
• tests/graph/decomposition-quality.test.js
• tests/graph/upward-propagation.test.js
• tests/graph/apply-proposal.test.js
• tests/graph/orchestrator.test.js
• tests/graph/question-formulator.test.js
• tests/ui/scenario-form.test.jsx

And then by the broader requested validation pass with lint and build.