Science Legion

A six-agent squad's first catch was an error in its own benchmark. The formula survived; the answer key did not. See evals/pilot-fourth-order/.

A DSH skill bundle for tiered, verified science campaigns — math, physics, astronomy, engineering, and biomedical evidence appraisal.

It is not a solver and it does not promise open-problem breakthroughs. It is a command layer for DSH subagents and workflows: one commander, a shared blackboard, independent lines of attack, a separate verifier, a red team, and an explicit token/time budget. The unit of progress is a verified claim, not the number of agents.

Cost planning

Before a run above six agents, use the shipped estimator from the bundle directory:

node scripts/estimate-campaign.mjs --agents 24

It scales the measured six-agent pilot linearly and adds 25% coordination overhead for tiers above six. Six agents reproduces the measured ¥1.4543; 24 is about ¥7.27. That is a planning number, not a provider quote: coordination and verification can grow faster than linearly. Put the cap in BRIEF.md before spawning.

Why this exists

Frontier models are good at producing plausible science. Plausible is not verified. I wanted a tool judged by verified claims per token, not by how many agents it can spawn. The first six-agent run caught a bad answer key — mine. That is the product working.

Built by lengrufeng2008 on DeepSeek Harness. The agent runs the campaigns; the human sets the token budget.

What is in the bundle

Skill	Role
science-legion	Router: frame the claim, load the domain skill, choose a tier
legion-command	Squad/platoon/company/legion doctrine, blackboard, worker contract, budgets
math-research	Definitions, proof obligations, counterexamples, symbolic and numerical checks
physics-research	Units, conservation, limits, approximation control, simulation, uncertainty
astronomy-research	Frames, time systems, photometry, distances, selection effects, archives
engineering-research	Requirements, load cases, units, safety factors, tolerances, failure modes, tests
biomedical-research	PICO, study design, absolute vs relative effects, bias, applicability; not medical advice
science-verify	Independent evidence table, reproduction, falsification, provenance

Wolf-pack and flanking tactics

The command skill names the shapes: wolf pack (independent packs with different methods), flanking (reach the claim through a neighbouring field or limiting case), pincer (two independent methods must agree), feint (bounded dead-end mapping), ambush (reserve a counterexample pack), siege (rounds with a blackboard, not one huge fan-out), and retreat (stop and report a negative result). A 1000-agent legion is 10+ companies under leads, never 1000 direct reports.

Scale is a budget decision

Default is a Squad (6): commander, two scouts, calculator, independent verifier, red team. The doctrine then allows 24, 96, and only after validation 384–1000. A 1000-agent run is a compute campaign; estimate the token bill, write the cap in the brief, stop at 70%, and never use agent count as evidence.

The bundle does not spawn those agents itself. It tells the model how to use DSH's existing workflow and subagent tools; the profile's tools, model, and rate limits still apply. No 1000-agent run has been validated here — treat that tier as a guarded research mode, not a shipped fact.

Install

Use the DSH plugin manager with this absolute directory (or the packed tarball). Once the public repo exists, the GitHub form will be:

dsh plugin --profile web add github:lengrufeng2008/dsh-science-legion

Then ask for a hard science problem and invoke the science-legion skill (or the domain skill directly). The router will pick the domain and legion-command.

Use

Use science-legion on: prove or refute ...

The commander writes a brief under research/<slug>/, starts with a squad, and reports claim, evidence, uncertainty, failures, and cost. Verification is a separate role; the author cannot mark their own result verified.

Honest boundaries

• No claim of solving a math/physics/astronomy open problem. A proof, a certificate, a reproducible computation, or a decisive measurement is required.
• No fabricated citations, data, observations, or numbers. Unknown facts are marked UNKNOWN.
• No patient data or credentials in prompts.
• No majority vote as truth; no LLM-only judge; no silent retries.
• 1000 agents can be expensive and slow. The budget gate is part of the product.

Roadmap

• v0.2 (this bundle) — eight skills, ten known-answer evals, two pilots.
• v0.2 — validated 6- and 24-agent runs, a token cost ledger, artifact manifests, and a hard verifier gate.
• v0.3 — a Host legion planner/dashboard and a 96-agent company run with a budget cap.
• v1.0 — budgeted 384–1000-agent campaigns, proof/certificate/measurement requirements, and a public benchmark ledger. Open problems remain attempts until a proof or decisive measurement exists.

No tier is a promise of a breakthrough. The roadmap is about better evidence per token, not bigger numbers.

Review

Standalone showcase: showcase.html. See REVIEW.md for a 60-second code tour and the boundaries of what the pilot does and does not establish.

Tests

npm test

Covers the manifest, frontmatter names, provider registration, lookup, and the mandatory doctrine markers.

Pilot

Two pilots, both on known problems:

• 4 agents, second order (evals/pilot-pendulum/): the formula was supported; the commander's pure-stdlib check was rerun and a separate AGM-only check reproduced the coefficient.
• 6 agents, fourth order (evals/pilot-fourth-order/): the independent verifier caught an error in the shipped answer key. The old expected value 1.066315375 is the sixth-order truncation T6, not the fourth-order T4 = 1.066080729...; the key was corrected, while the formula survived.

Both used the command protocol prompts directly; automatic skill routing is tested by the provider tests, not by the pilots. They are not evidence for 24/96/1000-agent runs or open problems.

The six-agent pilot's cost is now measured from local session logs: ¥1.4543 for 110 calls (5,790,848 cache-read, 155,611 uncached input, 295,723 output; six agent sessions only; main session excluded). That is ¥1.45 per verified answer for this run, not a linear price for 1000 agents. The 1000-agent tier remains untested.

Evals

evals/problems.json contains known-answer problems in math, physics, astronomy, engineering, and dimensional analysis. They are a benchmark, not a claim of open-problem capability. See evals/README.md for the protocol.

License

MIT.