Maxim
Confidence tags ยท ADR-010

Trust becomes a property
of the output, not your mood.

Generic LLMs communicate with the same confidence whether they're citing a fact or making one up. That's the worst possible signal-to-noise for decision-making. Maxim tags every output so you know what to trust and what to verify.

claude code โ€” confidence rubric example

The five tiers

Each tag comes with a four-line rubric in the Technical Educator voice: Basis (what grounds the answer), Gap (what's missing), Mitigation (what to do about it), Next (what would change the tag).

๐ŸŸข

HIGH

Maxim skill matched the domain, primary framework fired, gap log is clean, and no open compliance question remains.

Example: A pricing-page hero composed with Fogg B=MAP, no compliance ambiguity, brand voice loaded from overlay.

๐ŸŸก

MEDIUM

Skill matched but one of the gap categories is unresolved โ€” customer data not validated, external-input assumption, or framework applied with limited training context.

Example: A SQLite-to-Postgres recommendation given without seeing your actual query log.

๐Ÿ”ด

LOW

No Maxim skill matched the domain, or an external pack ran without the Maxim behavioral overlay. Tag is also applied if the compliance gate blocked a sub-part of the output.

Example: A generic answer when you asked for something outside Maxim's 34 skill domains โ€” Maxim tells you it's generic so you know to validate heavier.

๐Ÿ”ต

SUPER USER

Operator has super_user.enabled = true in the project manifest. Governance gates suppressed; Claude's values layer still active. Use with care.

Example: Internal dev environments where you want raw Claude behavior for debugging.

๐Ÿ”

GATED

Project manifest has gated: true. Maxim refuses output until explicit operator approval is recorded in the session.

Example: Pre-launch private projects where any AI output needs manual review before proceeding.

Source

The rubric is governed by Architecture Decision Record ADR-010 Confidence Tag Technical Educator Rubric. The ADR is public โ€” read it to see why the four-line shape was chosen over alternatives, and what would require a superseding ADR to change.