sema-tax
The Sema tax curve prices what an agent pays — in tokens, bytes, and cost — to carry an increasing number of semantic patterns, and what that overhead buys in task quality.
How to read the results
Each trial gives a model a worksheet whose items can only be answered correctly using the definitions of the patterns currently active (0 to 16). Definitions arrive either inline as prose, behind an opaque lookup, or as content-addressed references; resolver arms additionally vary whether the local registry is cold (references must be hydrated) or warm. The primary endpoint is task success per total model token. Provider-side prompt-cache telemetry is recorded but observational (see ADR 0011); the controlled cache effect is hydration bytes.
p0-baseline- No patterns active. The shared origin of every curve.
pN-prose-*- N pattern definitions inline as prose. Pays wire bytes up front, no hydration.
pN-opaque-*- N patterns behind content-free lookup IDs. Pays hydration on a cold registry.
pN-content-*- N patterns as content-addressed references. Same hydration as opaque; the wire-byte difference is the addressing overhead itself.
Interpretation — as of 2026-07-15
One exploratory pilot on one model — a first pricing of the tax curve, not a proven effect.
Task quality rises with active pattern count, but token efficiency stays flat until coverage is complete — then doubles. At partial coverage the model burns output tokens deliberating over items whose definitions it lacks, so incomplete semantic coverage is both the least accurate and the most expensive regime. The addressing tax is real: at full coverage, content-addressed delivery pays roughly 20% of score-per-token relative to inline prose, from reference wire and hydration overhead. It is a bounded premium, not a quality cost — the content-addressed arm was also the most accurate (0.994–0.996 versus 0.971 for prose at sixteen patterns) with a 100% answered rate.
These fixtures sit near the worst case for references — definitions of roughly a hundred bytes, used once; the tax should shrink as definitions grow or are reused. Provider prompt caching was pervasive in both cache arms, confirming the observational status of cached-token telemetry (ADR 0011). One cell (twelve patterns, prose, cold) dips below its eight-pattern counterpart; the trial records are published for inspection. A preregistered test of the curve and of the accuracy hint has not yet run.
| Date | Mode | Model / provider | Trials | Best full- coverage score |
Score / 1k tok range |
Report |
|---|---|---|---|---|---|---|
| 2026-07-15 | Model pilot | MiniMaxAI/MiniMax-M2.5-TEEllm.chutes.ai |
2790 | 0.996 | 0.1318–0.4297 | Report |
Best full-coverage score is the highest mean worksheet score among the sixteen-pattern arms; the score-per-1k-token range spans every condition of the run. Higher score per token is better. Provider cache telemetry is observational (ADR 0011).
Size/reuse follow-up arm
Pattern count is fixed at p8 while definition size and within-session reuse vary. Deterministic results validate the crossover measurement, not model performance.
| Date | Mode | Trials | Content wins bytes | Content wins tokens | Report |
|---|---|---|---|---|---|
| 2026-07-23 | deterministic-harness |
54 | 4/9 | 5/9 | Report |