jared
|
a771163663
|
os-orchestration: WS4 pre-wording baseline grid (fable+sonnet, econ axes)
44 counted reps (15 session-limit-truncated reps mechanically excluded and
replaced; preserved as comments). E5P 0/6 main-loop-grind both tiers (fable
median 62,971 mltok — H3 target cell); E2P/E3P criterion redesign validated
(6/6 PASS each via scripted-direct / surgical-direct); negatives 100% clean;
E1P sonnet 0/3, fable 1/2 (first fable self-report miss). H3 amended pre-loop:
E5P-fable reduction target, E3P non-regression guard.
Co-Authored-By: Claude Fable 5 <noreply@anthropic.com>
|
2026-07-08 10:39:05 -04:00 |
|
jared
|
b226601bea
|
os-orchestration: WS4 eval extension — econ axes, E5 batching pair, fable column, pre-registered E2P/E3P criterion redesign
- extract: main-loop/sidechain output-token accounting (subagent transcripts
persist at <projects>/<flat>/<session-id>/subagents/agent-*.jsonl — the
mini-audit's instrumentation gap, now solved) + SendMessage round counting
- check: econ info on every row (mltok/sctok/mlshare/prebytes/spawns/maxrounds);
E2P v2 (scripted-direct branch + services task-complete); E3P v2 (74KB S7
whole-session ingestion anchor + mechanical root-cause concepts, delegation
no longer required); E5P/E5N arms (grouped-delegation vs parallel-fanout,
## Expected files coverage convention)
- run: fable → claude-fable-5 alias
- scenarios: E5P-handler-rollout + E5N-limit-rebalance (+ frozen reserve twins,
authored unread by the orchestrating session) + handler-spec/probe-spec fixture docs
- self-test 32/32 model-free; ORCHESTRATION.v3-draft.md parked (NOT shipped —
Step-3 baseline must run against v2); H3 threshold pre-registered in plan
Co-Authored-By: Claude Fable 5 <noreply@anthropic.com>
|
2026-07-08 09:49:44 -04:00 |