sema-evalsIndependent evaluations

Workflow value — 20260723T073224723Z-order-20260716

Validates the workflow-value scaffold mechanics only: clearly labelled synthetic seed tasks, hidden executable validators, dev/eval separation, paired randomized conditions, fixed token-budget accounting, semantic delivery channels, repair notice handling, and durable result preservation. Scripted outcomes are not evidence about model performance (ADR 0021).

Dataset gate: SEED-ONLY. These are synthetic seed fixtures and scripted outcomes, not an evaluation dataset and not evidence that workflows improve model performance.

ConditionTrialsEval trials Eval success in budgetValidator pass Parse failureWire B Hydration BInput tok Output tokTotal tok
task-only 3 2 0.0% 0.0% 0.0% 663.7 0.0 220.7 78.7 299.3
equal-prose 3 2 100.0% 100.0% 0.0% 1157.0 0.0 382.3 80.3 462.7
opaque-resolver 3 2 100.0% 100.0% 0.0% 704.7 473.3 393.7 80.3 474.0
content-addressed 3 2 100.0% 100.0% 0.0% 777.0 473.3 411.3 80.3 491.7
content-addressed-repair 3 2 100.0% 100.0% 0.0% 933.0 473.3 480.3 80.3 560.7

Every displayed aggregate is recomputed from trials.public.jsonl. Wire bytes, hydration bytes, input tokens, and output tokens are reported separately.