sema-evalsIndependent evaluations

Sema tax size/reuse arm — 20260723T073225662Z-size-reuse-order-20260714

Validates the size/reuse condition grid, per-message and cumulative byte/token accounting, one-time resolver hydration, the size-tier byte bands, scoring, randomization, and reporting only. Deterministic outcomes and token prices are scripted; the token model attributes each definition ingestion once per wire delivery (see ADR 0013).

Deterministic harness validation only. Scores and token prices are scripted; this run measures the controlled byte and reuse geometry and is not empirical evidence about language models.

Crossover surface

SizeRProse semantic B Content semantic BContent wins B Prose tokContent tok Content wins tok
small 1 1347.5 2131.5 no 461.0 666.0 no
small 3 4042.5 4843.5 no 1383.0 1473.0 no
small 9 12127.5 12979.5 no 4149.0 3894.0 yes
medium 1 8465.5 9249.5 no 2328.5 2533.5 no
medium 3 25396.5 11961.5 yes 6985.5 3340.5 yes
medium 9 76189.5 20097.5 yes 20956.5 5761.5 yes
large 1 29106.5 29890.5 no 7572.5 7777.5 no
large 3 87319.5 32602.5 yes 22717.5 8584.5 yes
large 9 261958.5 40738.5 yes 68152.5 11005.5 yes

Condition grid

SizeRDeliveryTrials Mean scoreWire B Hydration BTotal semantic B Total tokScore / 1k B
small 1 prose 2 0.625 1347.5 0.0 1347.5 461.0 0.4638
small 1 opaque 2 0.625 804.0 775.5 1579.5 527.0 0.3957
small 1 content 2 0.625 1356.0 775.5 2131.5 666.0 0.2932
small 3 prose 2 0.625 4042.5 0.0 4042.5 1383.0 0.1546
small 3 opaque 2 0.625 2412.0 775.5 3187.5 1056.0 0.1961
small 3 content 2 0.625 4068.0 775.5 4843.5 1473.0 0.1290
small 9 prose 2 0.625 12127.5 0.0 12127.5 4149.0 0.0515
small 9 opaque 2 0.625 7236.0 775.5 8011.5 2643.0 0.0780
small 9 content 2 0.625 12204.0 775.5 12979.5 3894.0 0.0482
medium 1 prose 2 0.625 8465.5 0.0 8465.5 2328.5 0.0738
medium 1 opaque 2 0.625 804.0 7893.5 8697.5 2394.5 0.0719
medium 1 content 2 0.625 1356.0 7893.5 9249.5 2533.5 0.0676
medium 3 prose 2 0.625 25396.5 0.0 25396.5 6985.5 0.0246
medium 3 opaque 2 0.625 2412.0 7893.5 10305.5 2923.5 0.0606
medium 3 content 2 0.625 4068.0 7893.5 11961.5 3340.5 0.0523
medium 9 prose 2 0.625 76189.5 0.0 76189.5 20956.5 0.0082
medium 9 opaque 2 0.625 7236.0 7893.5 15129.5 4510.5 0.0413
medium 9 content 2 0.625 12204.0 7893.5 20097.5 5761.5 0.0311
large 1 prose 2 0.625 29106.5 0.0 29106.5 7572.5 0.0215
large 1 opaque 2 0.625 804.0 28534.5 29338.5 7638.5 0.0213
large 1 content 2 0.625 1356.0 28534.5 29890.5 7777.5 0.0209
large 3 prose 2 0.625 87319.5 0.0 87319.5 22717.5 0.0072
large 3 opaque 2 0.625 2412.0 28534.5 30946.5 8167.5 0.0202
large 3 content 2 0.625 4068.0 28534.5 32602.5 8584.5 0.0192
large 9 prose 2 0.625 261958.5 0.0 261958.5 68152.5 0.0024
large 9 opaque 2 0.625 7236.0 28534.5 35770.5 9754.5 0.0175
large 9 content 2 0.625 12204.0 28534.5 40738.5 11005.5 0.0153

Every displayed aggregate is recomputed from trials.public.jsonl. Wire, hydration, and model-token channels are reported separately.