Grid (36 reps + 2 canaries, all counted): negatives 18/18 both tiers, zero
over-triggering; positives 1/18 on delegate-at-all, but transcript
verification splits those FAILs into superior direct strategies (scripted
bulk edit, surgical greps - the audit's own justified-non-delegation shape)
vs genuine misses (haiku whole-file ingestion, per-file grind). Conditional
rules validated where engaged: 14/14 spawns explicit-model, self-report
downgrade flag worked. E2P/E3P criterion redesign pre-registered with
independently-anchored thresholds; reserve untouched. Analysis in vault:
os-orchestration-eval-baseline-grid-results.
Co-Authored-By: Claude Fable 5 <noreply@anthropic.com>
Harness: 6 run-set + 6 frozen-reserve paired scenarios from the WS1 verified
misses, Node relaystation fixture (deterministic sandbox-time logs), headless
runner with CLAUDE_CODE_SUBAGENT_MODEL stub for E1P, deterministic-first
checker driving audit/bin/extract (extended with per-segment read targets),
narrow frozen haiku-judge fallback for E1 language axes, model-free self-test
21/21 incl. shipped-instruction-compliant conformance transcripts.
Wording v2: the E1 canary (counted) proved resolvedModel is not model-visible
in the launch result, so the launch-stub-comparison rule was unactionable;
replaced with subagent model self-report on sonnet/opus spawns (subagents know
their exact model ID - verified live). Scenarios/checker/rubric untouched.
Deterministic env-override detection routed to WS3 (plan updated).
Co-Authored-By: Claude Fable 5 <noreply@anthropic.com>
jared