Season 18 Energy Bees

🧬 Season 18 — Operation Energy-Bees

One sentence: give every Bee (specialist model) a measured energy — how wrong and how costly its answer is, checked by running it — and let a child Bee replace its parent only when that energy goes down by more than noise, with a rollback.

Roadmap of record: ROADMAP.md §3.16 · Proposal: docs/AUTOPOIETIC_ENERGY_BEES_PROPOSAL.md · Epic: #208 · How to contribute: Human and Agent Workflow

Why now

Measured on 2026-10-25/26:

• The BOINC night chain (98 rounds on the T4) lowered held-out loss from 2.2705 to 2.1380 (−5.8 %), then plateaued, and its head was chosen best-of-97 on the same small sets it is reported on.
• Loss is not answer quality: the first real coding answer was fast and wrong; adapters trained on terse gold text interfered with other skills.

What comes from AutoevolveAI, and what does not

AutoevolveAI tested these ideas with gates that could fail (docs/EVOLUTION_LAB.md):

Imported (its gate passed)	Not imported (its gate failed)
Hidden-test energy (Phase 1: the self-graded loop claimed 11/20, hidden tests verified 7)	A learned energy predictor that skips execution (Phase 2: 40.0 % vs 44.3 % random)
Promotion gate with equivalence check, A/A noise margin, rollback (Phase 3: 5/5, toy components)	Full-solution lesson memory and verbose retries on small models (Phase 1 UC2/UC4)
Best-so-far and short error prompts for small models (Phase 4 D1/D3/D6)	New crossover operators (8 tested, all hit the same cliff)
One warm verifier process (one Mathlib import was the whole cost of a Lean check)

The energy, in plain words

For an answer y to prompt x:

• Crash, timeout, empty function, sorry: energy 1 000 000 (rejected).
• Otherwise: 100 × (share of hidden tests failed) + conservation/consistency residual + a cost for time and memory that is always smaller than one failed test. So a correct slow answer always beats a wrong fast one; cost only breaks ties between equally correct answers.

Code: src/hivemind/verification/bee_energy.py (PR #207), tests: tests/test_bee_energy_autopoiesis.py.

Tasks (claim one)

Task	Issue	Unblocks	Labels	Depends on
HM-S18-AE-01 energy function in the sandbox	#209	#191	agent-ready, human-review	—
HM-S18-AE-02 disjoint reported split + plateau rule	#210	BOINC chain	agent-ready, needs-prereg	—
HM-S18-AE-03 energy-gap training pairs	#211	#166	agent-ready, needs-prereg	AE-01; owner approval for data
HM-S18-AE-04 promotion gate + registry with rollback	#212	#165, #170	agent-ready, human-review	AE-01
HM-S18-AE-05 warm verifier → bandit reward	#213	#173	agent-ready	—

Suggested order: AE-01 → AE-04 → AE-03; AE-02 and AE-05 in parallel.

Same-day review of the first draft

#	Finding	Fix
R1	Four tasks duplicated open issues (#191, #166, #165/#170, #173)	S18 builds the shared pieces; each task names the issue it unblocks
R2	Correct 6 s answer scored 60.5, half-wrong 5 ms answer 50.1	cost bounded below one failed test; regression test (now 20.4 vs 51.7)
R3	"Plateau rule fires at round ~32"	replayed on the real ledger: first fires at round 47; labelled descriptive
R4	Fixed noise margin 0.5, McNemar unspecified	margin from ≥ 3 measured A/A runs; exact McNemar primary
R5	Acceptance written by the engineering coordinator	all gates are drafts until the scientific lead pre-registers them
R6	10/10 on 10 hand-picked cases proposed as a pre-ranker	dropped

Run it locally

uv sync --all-extras
uv run pytest tests/test_bee_energy_autopoiesis.py -q
uv run python scripts/gwaya_stub_audit.py

No GPU is needed for AE-01, AE-03 and AE-04. AE-02 runs on the T4 night window and needs the owner's go-ahead.