Nexus Confidence Gate
Purpose: Decide whether CLASSIFY has enough evidence to select a route or must ask one focused question. Read when: Entering CLASSIFY, selecting among multiple valid routes, or deciding whether a reversible assumption is safe.
Confidence is a routing aid, not an authorization mechanism. Evaluate blocking unknowns first, then assign one discrete evidence band. Do not invent source weights, add “boosters,” or average away an unresolved dimension.
Pipeline
Gather evidence → type unknowns → blocker gate → evidence band → decision threshold → proceed or askEvidence sources, in authority order:
- The user's current request and explicit corrections.
- Observed repository/runtime state.
- Project rules and
.agents/PROJECT.md. - Git history as supporting context, never as authority to expand scope.
Absence of a source is not automatically a penalty. Use only sources relevant to the decision being made.
Blocking Unknown Gate
Type uncertainty using intent-clarification.md before assigning a score.
| Dimension | Blocking when | Action |
|---|---|---|
authority |
Permission to delete, publish, push, spend, deploy, or modify an external system is unclear | Ask; no confidence score can authorize the effect |
referent |
The target cannot be resolved from the request or observed state | Ask for the target |
goal |
Two valid interpretations produce materially different outcomes | Ask for the deciding axis |
scope |
The boundary changes risk, effect class, or the files/systems affected | Offer bounded scope options |
constraint |
An unstated limit would change the chosen approach | Surface the constraint that would otherwise be assumed |
outcome |
Success cannot be observed with available evidence | Agree on an observable or mark it UNVERIFIED |
Rules:
- Unresolved
authorityalways blocks. - Two or more materially valid interpretations always trigger
GATE, regardless of the numeric band. - A minor assumption may remain only when it is explicit, reversible, and cannot change the requested outcome.
- Batch missing dimensions into the one-question format defined by
intent-clarification.md; do not ask a sequence of open-ended questions.
Discrete Evidence Bands
Assign the highest band whose complete description is true. The bands are intentionally coarse; pseudo-precision is not evidence.
| Score | Evidence state | Default action |
|---|---|---|
1.00 |
Goal, target, scope, constraints, outcome, and authority are explicit; one route fits | Auto-route |
0.80 |
All load-bearing dimensions are settled; only minor reversible assumptions remain | Route and state assumptions |
0.60 |
One non-blocking dimension is partial, but a safe reversible default is established and rollback is clear | Proceed cautiously and record the default |
0.40 |
A load-bearing dimension is unresolved or multiple material interpretations remain | Ask one focused question |
0.20 |
Goal or target is mostly absent | Clarify before route selection |
0.00 |
No task-shaped intent can be recovered | Answer the meta/factual request directly or ask what outcome is wanted |
context_confidence < 0.60 is the mechanical GATE floor used by routing-matrix.md. The independent “2+ valid interpretations” rule still fires at any score.
Decision Thresholds
| Decision | Minimum band | Additional condition |
|---|---|---|
| Chain selection | 0.80 |
One best-fit Recipe/task type and no blocker |
| Agent routing | 0.80 |
Role boundary is clear |
| Approach selection | 0.80 |
Wrong choice is reversible and does not change scope |
| Recovery action | 0.80 |
Failure class is known and rollback/checkpoint exists |
| Parallel vs sequential | 0.60 |
Branches are independent, have owners, and define a merge gate |
Below the threshold, use intent-clarification.md § Routing Decision Output to present the smallest useful choice. Do not dump the routing matrix or confidence arithmetic.
Reversibility and Safety Overrides
| Effect | Examples | Auto-proceed |
|---|---|---|
| Read-only | Inspect files, run non-mutating checks | Yes when the route is clear |
| Reversible workspace edit | Git-tracked code/docs edit with a known rollback | Yes within approved scope |
| Moderate | Config or migration with a tested rollback | Ask when project rules require it |
| Difficult/irreversible | Data deletion, external send, production deploy, payment, key rotation | Never infer authority |
The following bypass the confidence band and use Nexus Ask First/safety rules:
- credential, auth, permission, or encryption-key changes;
- destructive or bulk data operations;
- production deploys and external-system mutations;
- cost-incurring actions;
- changes affecting 10+ files, breaking APIs, or architecture when not already approved.
AUTORUN modes do not bypass these controls. Explicit approval settles authority only for the named effect and scope; it does not raise unrelated confidence dimensions.
Decision Record
_CONTEXT_SNAPSHOT:
task: "<original request>"
evidence:
conversation: ["<decisive user statement>"]
repository: ["<observed state>"]
project: ["<applicable rule or journal fact>"]
git: ["<supporting history, if relevant>"]
dimensions:
referent: settled | partial | blocking
scope: settled | partial | blocking
goal: settled | partial | blocking
constraint: settled | partial | blocking
authority: settled | blocking
outcome: settled | partial | blocking
context_confidence: 0.00 | 0.20 | 0.40 | 0.60 | 0.80 | 1.00
assumptions: ["<explicit reversible assumptions only>"]
decision: proceed | ask | direct_answerRecord the evidence that changed the decision, not every file inspected. Re-evaluate when the user corrects the interpretation, observed state contradicts an assumption, or the action's effect class changes.
Phase Integration
CLASSIFY
- Gather only decision-relevant evidence.
- Type unknowns and apply the blocker gate.
- Assign a discrete band.
- If the band clears the decision threshold and no blocker exists, select the route.
- Otherwise ask one focused question, integrate the answer, and re-evaluate from evidence rather than adding a fixed score bonus.
EXECUTE
- Pass the approved intent contract and explicit assumptions, not the scoring rationale, to specialists.
- If new evidence invalidates a load-bearing assumption, stop at the nearest reversible boundary and re-enter CLASSIFY/GATE.
HANDOFF
Handoff completion confidence is owned by handoff-validation.md. It must not be substituted for context confidence: a specialist can be highly confident in work performed against the wrong intent.
LEARN
Record routing corrections as categorical evidence: which dimension was missed, which route was corrected, and whether the correction changed outcome or only efficiency. Do not tune invented weights; update examples, decision boundaries, or fixtures only after repeated verified cases.