Confidence as Triage

Parallax treats confidence as a routing decision, not a calibrated probability that an agent is correct. The orchestrator does not try to know when an agent is right — it tries to know when it can't be sure, and routes accordingly.

One line: your swarm knows when to ask for help. Failing safely (surface to a human) beats failing silently (an agent claims done and nobody checks).

Why not calibrated correctness?

The tempting pitch is "Parallax extracts a calibrated probability that each agent's output is correct." Parallax deliberately does not build on this, because it is genuinely hard and partly unsolved:

• Most LLM providers don't expose logprobs, and token probability measures fluency, not correctness — models hallucinate confidently.
• Verbalized confidence ("I'm 90% sure") correlates weakly with correctness. A self-reported CONFIDENCE: 0.9 marker is the weakest version of this.
• Rigorous ML approaches (evidential/Bayesian methods, conformal prediction) need training-time access or a labeled calibration set. Parallax orchestrates closed CLI agents — that door is closed.

So instead of introspection, confidence in Parallax is grounded in verification against reality.

The reframe: attention allocation

An orchestrator does not need a probability. It needs to decide: accept this, or get it checked? A good system spends its scarce resources well:

1. Auto-verify where an oracle is cheap.
2. Spend a second agent where that's worth the cost.
3. Spend a human's attention only where nothing else can resolve it.

Confidence is the signal that decides which branch a result takes. Its job is to distinguish "checkable and fine" from "I can't be sure" — not to be right about correctness.

The tiers

For any task, confidence is resolved by the first tier that applies:

Tier	Signal	Trust	Cost
Cheap oracle	Compiler / typecheck / tests pass, schema & constraint checks	high	~free (already run)
Structural acceptance check	"Done means X" turned into a lightweight per-task checklist — confirms shape, not truth	medium	cheap
Second agent	A reviewer / cross-checker when structural checks aren't enough	medium-high	one agent turn
Human	The universal fallback when no automated oracle exists	highest	expensive

The introspection signals (self-report markers, consistency sampling, linguistic hedging) are optional supplements, never the default.

This is why coding is the beachhead: it has the richest, cheapest oracles of any agent domain (compile, typecheck, tests, diff/scope sanity).

The escalation policy

Org-chart patterns declare a per-role escalation policy that consumes these signals. Each role can specify:

• accept — the result is good enough; move on.
• retryBelow — below this confidence, retry the task (optionally with more guidance).
• escalateBelow — below this confidence, escalate to the role's reportsTo (the agent tier) and, ultimately, to a human.

The policy is fed by verification results — test outcomes, typecheck, acceptance checks — with reportsTo as the agent tier and human-surface as the terminal. An engineer whose tests fail escalates to the architect. That needs no ML magic to be obviously correct.

The confidence library

The @parallaxai/confidence library provides the algebra that carries these values through a pattern:

• Confident<T> — a value paired with its confidence and provenance.
• Combinators — cf, best, gate, uncertain, coalesce — for selecting, thresholding, and combining confident values.
• Aggregation — combining multiple confident results (e.g. across a role's instances) into one.

What this is NOT

• Not a promise to verify open-world truth. Parallax verifies where oracles are cheap, structurally-checks where it can, and escalates the rest.
• Not a claim that the swarm is always right. The claim is that the swarm knows when to ask for help.
• Not a fix for hard tasks. Confidence lets you fail safely; it doesn't make real-world tool access or ambiguous-failure recovery easy.

Learn more

• Confidence Scoring — how scores flow through aggregation
• Consensus — combining multiple agent opinions
• Verification Patterns — building verify steps into org-chart workflows

For the full internal positioning, see docs/CONFIDENCE.md and docs/VERIFY.md in the repository.