sema-evalsIndependent evaluations

← sema-tax

sema-tax

20260715T103807828Z-order-20260714 · 2026-07-15 · 2790 trials across 3 scenarios

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. About this experiment

Provider prompt-cache telemetry (the cached-token column) is observational, not controlled: the cold/warm axis controls only harness-level hydration bytes, while a provider may cache prompt prefixes automatically across both arms (ADR 0011). The controlled cache effect is hydration bytes; read cached-token figures as description, not manipulation.

Provenance

Model
MiniMaxAI/MiniMax-M2.5-TEE
Provider
llm.chutes.ai
Sema version
0.3.0
Semantic backend
semahash-python-api
Canonicalization version
v2
Vocabulary root
(none)
Prompt digest
6dcff7f4cba0cf91053ea5b79e49f4694a1aaeb62f754b5da4cbf634e6362263
Fixture digest
ade52f7d5ea645e3002a73e43fc1749ae52f30663a37d75709e2dfde9a98e35c
Implementation commit
47a7da0925e53a862cf44506445e12feec471cef
Dependency lock digest
8492976045dc82b4c2cd05728a67b59c5e2a7c1ab03b451531bf34a010aadd27
Protocol version
0.3.0
Artifact schema version
0.3.0
Order seed
20260714
Seeds
0, 1, 2, 3, 4, 5, 6, 7, 8, 9, 10, 11, 12, 13, 14, 15, 16, 17, 18, 19, 20, 21, 22, 23, 24, 25, 26, 27, 28, 29

Results by condition

Condition Trials Patterns Mean
score
Answered
rate
Wire
bytes
Hydration
bytes
Input
tok
Cached
tok
Total
tok
Score /
1k tok
p0-baseline 90 0 0.235 0.478 550.0 0.0 505.0 496.5 1704.2 0.1377
p2-prose-cold 90 2 0.374 0.547 758.0 0.0 619.7 617.8 2056.9 0.1816
p2-prose-warm 90 2 0.375 0.575 758.0 0.0 619.7 617.9 2064.0 0.1817
p2-opaque-cold 90 2 0.438 0.689 623.0 193.0 653.7 643.3 2326.2 0.1881
p2-opaque-warm 90 2 0.428 0.656 623.0 0.0 653.7 649.0 2260.0 0.1893
p2-content-cold 90 2 0.456 0.731 763.0 193.0 739.7 728.0 2000.7 0.2277
p2-content-warm 90 2 0.435 0.682 763.0 0.0 739.7 734.3 2283.5 0.1904
p4-prose-cold 90 4 0.457 0.572 952.7 0.0 722.3 720.1 2734.0 0.1671
p4-prose-warm 90 4 0.496 0.635 952.7 0.0 722.3 721.1 2564.3 0.1934
p4-opaque-cold 90 4 0.567 0.747 681.0 387.7 780.3 766.5 2950.5 0.1921
p4-opaque-warm 90 4 0.501 0.664 681.0 0.0 780.3 770.8 2989.9 0.1677
p4-content-cold 90 4 0.537 0.713 961.0 387.7 949.3 930.1 2857.7 0.1881
p4-content-warm 90 4 0.546 0.735 961.0 0.0 949.3 948.3 2759.3 0.1978
p8-prose-cold 90 8 0.643 0.767 1342.0 0.0 931.7 922.0 3004.7 0.2140
p8-prose-warm 90 8 0.637 0.757 1342.0 0.0 931.7 930.7 3095.8 0.2059
p8-opaque-cold 90 8 0.656 0.786 797.0 777.0 1037.7 1022.4 3338.1 0.1964
p8-opaque-warm 90 8 0.637 0.750 797.0 0.0 1037.7 1017.3 3234.7 0.1971
p8-content-cold 90 8 0.661 0.787 1357.0 777.0 1378.7 1343.2 3532.7 0.1871
p8-content-warm 90 8 0.676 0.796 1357.0 0.0 1378.7 1371.1 3405.9 0.1986
p12-prose-cold 90 12 0.549 0.597 1732.3 0.0 1145.7 1121.2 4160.9 0.1318
p12-prose-warm 90 12 0.664 0.725 1732.3 0.0 1145.7 1137.5 3668.8 0.1810
p12-opaque-cold 90 12 0.624 0.681 913.0 1167.3 1299.7 1279.0 3898.9 0.1599
p12-opaque-warm 90 12 0.669 0.733 913.0 0.0 1299.7 1274.8 3756.5 0.1782
p12-content-cold 90 12 0.694 0.772 1753.0 1167.3 1812.0 1780.8 4077.1 0.1703
p12-content-warm 90 12 0.747 0.836 1753.0 0.0 1812.0 1788.3 3799.5 0.1967
p16-prose-cold 90 16 0.971 0.978 2125.7 0.0 1361.0 1331.6 2379.8 0.4080
p16-prose-warm 90 16 0.971 0.978 2125.7 0.0 1361.0 1360.0 2259.1 0.4297
p16-opaque-cold 90 16 0.951 0.956 1029.0 1560.7 1563.0 1550.2 2551.4 0.3729
p16-opaque-warm 90 16 0.963 0.967 1029.0 0.0 1563.0 1538.2 2530.0 0.3804
p16-content-cold 90 16 0.994 1.000 2149.0 1560.7 2251.0 2222.2 2972.5 0.3346
p16-content-warm 90 16 0.996 1.000 2149.0 0.0 2251.0 2236.2 2998.4 0.3321

Mean score is the graded fraction of worksheet items correct; answered rate is format compliance (items that received a parseable line). Total tokens are billable model tokens (fresh input + output); cached reads are counted separately. Score per 1k tokens is the primary endpoint. Wire and hydration bytes are the controlled transport channels.

Raw public derivative

manifest.json summary.json trials.public.jsonl

The public derivative strips raw provider payloads and caps transcript text. Full raw bundles are retained locally only.