Methodology

What the evidence means

Device-attested, content-free, replay-safe

The official CLI extracts usage counters from supported local assistant records, signs the payload with a registered device key, and sends it to Burnbook. The server validates the signature, rejects replays, and checks schemas, timestamps, identities, and contradictory counters. Token volume is never a rejection or fraud signal.

Token accounting

Processed and Fresh answer different questions

Processed includes input, cache reads, cache writes, output, reasoning, and tool input. Fresh excludes cache reads.

A cache read means the model reused an earlier prompt or context prefix instead of processing that prefix from scratch. The model still considered those tokens, so they count as Processed, but they were generally faster and cheaper. A cache write is context processed now and saved for later, so it remains Fresh.

Collector registry

Current compatibility

• Claude Code: supported via transcript on cli. Persisted local Claude Code sessions; ephemeral or deleted sessions are not covered.
• Codex: preview via transcript on cli. Persisted local rollouts only; ephemeral and cloud-only sessions are not covered.
• Gemini CLI: preview via otel on cli. Local token telemetry with prompt logging disabled.
• Cursor: preview via import on ide. User-supplied content-free usage exports only.
• Antigravity: preview via import on ide. User-supplied content-free usage exports only.

Scoring

AI Fluency formula: fluency-v1

AI Fluency is the rounded, equally weighted average of four 0–100 pillars. Spending more affects only the log-scaled volume pillar. Formula changes create a new version rather than silently changing the meaning of an old credential.

• Efficiency (25%): 70% cache-hit discipline and 30% output per input context. Input context includes input, cache-read, and cache-creation tokens. Cache performance is normalized from 50% to 100%; output per context reaches its cap at 2%.
• Consistency (25%): equal parts active days in the trailing 90 days, capped at 63, and consecutive weeks with activity, capped at 13.
• Breadth (25%): 70% distinct models, capped at five, and 30% distinct assistants, capped at two, measured over the trailing 90 days.
• Volume (25%): logarithmic interpolation through 0 tokens = 0, 100 million = 60, 1 billion = 80, and 100 billion = 100.

Composite grades are A at 85+, B at 70+, C at 50+, and D below 50.

Evidence coverage is supported-source tokens divided by all collected tokens. Preview evidence can lower coverage but cannot raise a score or ranked total. Formula version and coverage are available in public credential metadata and API responses and should be read together with the displayed score.

Weekly and quarterly Burn boards show supported-source token volume separately. Preview collectors never contribute to ranked totals.

Leaderboard

Competition requires consent

• Public profiles do not automatically enter the leaderboard.
• Participants must opt in, keep their profile public, and have an Active or Cleared integrity state.
• Eligibility requires supported evidence on at least three days in the trailing 90 days and activity within the trailing 30 days.
• Equal scores share a competition rank. Handle ordering is display-only and does not break a scoring tie.
• The Founding Season remains hidden until 100 eligible people join. Once opened, that season never relocks if the eligible population later falls.
• Accounts under integrity review, banned accounts, and preview-only evidence are excluded.

Weekly Burn uses the current UTC ISO week. Quarterly Burn uses the current UTC calendar quarter.

Account owners can see their status and request a review from the private dashboard. The public review and appeal process explains each state without disclosing controls that would make abuse easier.