Model Token & Cost Accounting (I02 / I04)
Disjoint token components per model, plus observed and estimated cost kept apart. Reasoning stays inside output. Unknown cost is not zero. Task, actor, and model token accounting across invocations.
4-Component Token Distribution by Model: Each stacked bar decomposes token consumption into physical token classes:
Uncached Input (Cyan)
Fresh prompt tokens processed from scratch by the model API. High proportions indicate poor prompt cache reuse.
Cache Read (Emerald)
Prefix tokens served from existing prompt cache. High cache hit rates dramatically reduce billing and turn response times.
Cache Write (Purple)
New tokens written to context checkpoint store to establish prefix caching for future turns.
Generated Output (Amber)
Model output tokens, including tool calls and internal reasoning tokens (per I02 disjoint standards).
Disjoint token accounting equations and efficiency thresholds:
Cache Read Ratio (I02)
cache_read / (input_uncached + cache_read + cache_write)
Disjoint Token Sum
input_uncached + cache_read + cache_write + output
Task-Intake Overhead
Fraction of tokens expended on prompt discovery before task assignment
Physical Tokens Only
All values track verified provider counts; currency is never imputed
4-Component Token Distribution by Model
Fresh uncached input vs cache reads vs cache writes vs generation output
| Actor ID | Model Family | Effort | Attribution Bucket | Input Uncached ▾ | Cache Read | Cache Write | Output | Total Tokens | Cache Read Fraction of Input (I02) |
|---|
Technical Lineage, Coverage & SQL Proof (L3)
SELECT * FROM v_mission_token_burn ORDER BY input_uncached DESC, seat_id;
Connected fact rows
Recorded window: —
Total rows: —