Work · Research governance

Frontier Lab

"This repository manages the research process. It does not decide what is true about nature or mathematics."
FrontierLab-Governance README(external site)

Frontier Lab takes the same method to open science. No model is exempt from checks because of its brand, and two models agreeing is not independent verification. The tooling is the Python standard library only, and each round's default budget is $0. Status is shown as it is: most problems are still OPEN; the Physics, Biology and Chemistry results are still in review; Math has one merged round of evidence and claims no resolution.

It is not a benchmark or a Q&A set. It is a multi-model research workflow under strict governance.

1 governance repo
15 labs
150 problem cards
Python standard library
$0 default per round

The whole program was set up between 2026-09-27 and 2026-09-29.

1 ○ = 1 problem card

Governance on the left; one row per lab, one cell per problem card. ◆ marks each lab's first-round problem.

10 problem cards per lab
		001	002	003	004	005	006	007	008	009	010
FrontierLab-Governance	FrontierMath	First round: MATH-001	MATH-002	MATH-003	MATH-004	MATH-005	MATH-006	MATH-007	MATH-008	MATH-009	MATH-010
	FrontierPhysics	First round: PHYS-001	PHYS-002	PHYS-003	PHYS-004	PHYS-005	PHYS-006	PHYS-007	PHYS-008	PHYS-009	PHYS-010
	FrontierBiology	First round: BIO-001	BIO-002	BIO-003	BIO-004	BIO-005	BIO-006	BIO-007	BIO-008	BIO-009	BIO-010
	FrontierChemistry	CHEM-001	CHEM-002	CHEM-003	First round: CHEM-004	CHEM-005	CHEM-006	CHEM-007	CHEM-008	CHEM-009	CHEM-010
	FrontierComputerScience	First round: CS-001	CS-002	CS-003	CS-004	CS-005	CS-006	CS-007	CS-008	CS-009	CS-010
	FrontierStatistics	First round: STAT-001	STAT-002	STAT-003	STAT-004	STAT-005	STAT-006	STAT-007	STAT-008	STAT-009	STAT-010
	FrontierMetaScience	First round: META-001	META-002	META-003	META-004	META-005	META-006	META-007	META-008	META-009	META-010
	FrontierSocialScience	SOC-001	SOC-002	SOC-003	SOC-004	SOC-005	SOC-006	SOC-007	First round: SOC-008	SOC-009	SOC-010
	FrontierMaterials	MAT-001	MAT-002	First round: MAT-003	MAT-004	MAT-005	MAT-006	MAT-007	MAT-008	MAT-009	MAT-010
	FrontierAstronomy	ASTRO-001	ASTRO-002	First round: ASTRO-003	ASTRO-004	ASTRO-005	ASTRO-006	ASTRO-007	ASTRO-008	ASTRO-009	ASTRO-010
	FrontierEarth	EARTH-001	EARTH-002	First round: EARTH-003	EARTH-004	EARTH-005	EARTH-006	EARTH-007	EARTH-008	EARTH-009	EARTH-010
	FrontierNeuroscience	First round: NEURO-001	NEURO-002	NEURO-003	NEURO-004	NEURO-005	NEURO-006	NEURO-007	NEURO-008	NEURO-009	NEURO-010
	FrontierEconomics	First round: ECON-001	ECON-002	ECON-003	ECON-004	ECON-005	ECON-006	ECON-007	ECON-008	ECON-009	ECON-010
	FrontierEngineering	ENG-001	ENG-002	ENG-003	First round: ENG-004	ENG-005	ENG-006	ENG-007	ENG-008	ENG-009	ENG-010
	FrontierMedicine	First round: MED-001	MED-002	MED-003	MED-004	MED-005	MED-006	MED-007	MED-008	MED-009	MED-010

One round, seven gates

Check whether it's solved
a fresh four-way search within 24 hours of starting: general, discipline, solution, criticism. If an outside solution exists, it is recorded as COMPLETED_EXTERNAL and the work stops.
Reproduce
reproduce the known result first.
Freeze
the problem and its verifier are frozen first.
Bounded exploration
default budget per round: 30 minutes, $0, 100 tries.
Independent verification
no author signs off their own independent verification.
Record
results and failures both stay; records can be added to, never deleted.
Reviewed merge
authors never merge their own work; the owner decides.

Failures on record

Physicsfrozen E3Failed · recordedFrontierPhysics/pull/4(external site)
Biologyan unverified PASSWithdrawn · kept on recordFrontierBiology/pull/2(external site)
Chemistryfrozen C4/C5 thresholdsFailed · recordedFrontierChemistry/pull/1(external site)
11 new labsfirst CI runFailed · recorded, then fixed and re-pinnedFrontierLab-Governance/EXPANSION-2026-09-28.md(external site)

Rules

"No agent is exempt from verification because of its brand."
"Two models agreeing is not an independent experiment."

OPEN

What OPEN means

OPEN means only this: this bounded search found no confirmed solution of the same scope. It does not mean "unsolved".

Problem states: OPEN
PARTIAL
CLAIMED_RESOLVED
COMPLETED_EXTERNAL
COMPLETED_INTERNAL
PAUSED
RETRACTED
Round states: DRAFT
ADMITTED
PAUSED
CLOSED_EXTERNAL
FINISHED

15 labs, one governance core

Merged1
In review3
No research rounds yet11

10 problem cards per labData as of 2026-09-30

View

FrontierLab-Governance(external site)
Shared rulebook and gate software
- 2 open PRs
- 5 merged PRs
- 30 commits
- protocol 2.0.0
- CI green
Merged
FrontierMath(external site)
Exact constructions → certificates → Lean proofs
- 1 merged round
- Not claimed
- 6 open PRs
- 2 merged PRs
- 8 commits
- protocol 1.0.0
- CI green
- First round: MATH-001
FrontierMath is not affiliated with Epoch AI's FrontierMath benchmark.
In review
FrontierPhysics(external site)
Reproducible baseline → physical consistency → predictions that tell models apart
- 5 open PRs
- 2 commits
- protocol 1.0.0
- CI green
- First round: PHYS-001
In review
FrontierBiology(external site)
Public data → baseline → cross-condition validation → testable hypotheses
- 2 open PRs
- 2 commits
- protocol 1.0.0
- CI green
- First round: BIO-001
In review
FrontierChemistry(external site)
Keeps computed and experimental references separate; starts from solvation
- 2 open PRs
- 2 commits
- protocol 1.0.0
- CI green
- First round: CHEM-004
No research rounds yet
FrontierComputerScience(external site)
Formal verification and honest evaluation of AI coding agents
- 2 merged PRs
- 7 commits
- protocol 2.0.0
- CI green
- First round: CS-001
No research rounds yet
FrontierStatistics(external site)
Valid inference under distribution shift, selection and dependence
- 2 merged PRs
- 7 commits
- protocol 2.0.0
- CI green
- First round: STAT-001
No research rounds yet
FrontierMetaScience(external site)
Tests whether AI-agent science and its governance actually improve research
- 2 merged PRs
- 7 commits
- protocol 2.0.0
- CI green
- First round: META-001
No research rounds yet
FrontierSocialScience(external site)
Measurement → identification → replication → transportability
- 2 merged PRs
- 7 commits
- protocol 2.0.0
- CI green
- 10 OPEN
- First round: SOC-008
No research rounds yet
FrontierMaterials(external site)
Separates what computation predicts from what can actually be synthesized
- 2 merged PRs
- 7 commits
- protocol 2.0.0
- CI green
- 10 OPEN
- First round: MAT-003
No research rounds yet
FrontierAstronomy(external site)
Treats catalogs and their selection functions as research objects
- 2 merged PRs
- 7 commits
- protocol 2.0.0
- CI green
- 10 OPEN
- First round: ASTRO-003
No research rounds yet
FrontierEarth(external site)
Calibrating forecasts of rare, extreme events
- 2 merged PRs
- 7 commits
- protocol 2.0.0
- CI green
- 10 OPEN
- First round: EARTH-003
No research rounds yet
FrontierNeuroscience(external site)
Designs stimuli that make brain-computation models disagree
- 2 merged PRs
- 7 commits
- protocol 2.0.0
- CI green
- 10 OPEN
- First round: NEURO-001
No research rounds yet
FrontierEconomics(external site)
Measures what AI actually does to firms, tasks and work
- 2 merged PRs
- 7 commits
- protocol 2.0.0
- CI green
- 10 OPEN
- First round: ECON-001
No research rounds yet
FrontierEngineering(external site)
Simulation-first: fault injection and uncertainty before any deployment
- 2 merged PRs
- 7 commits
- protocol 2.0.0
- CI green
- 10 OPEN
- First round: ENG-004
No research rounds yet
FrontierMedicine(external site)
Validating medical AI across hospitals and over time
- 2 merged PRs
- 7 commits
- protocol 2.0.0
- CI green
- 10 OPEN
- First round: MED-001

Focus: no-three-in-line (MATH-001)

All 431,008 configurations in the public Flammenkamp database for n = 2..76 were verified legal; n = 75 is the only gap.

431,008Source: FrontierMath problems/MATH-001/experiments/canonical_evidence/CANONICAL_FACTS.mdAs of 2026-09-30

Records for n = 71–74 and 76 were cross-checked through two independent retrieval paths; an independent third verifier was checked by a 429-case adversarial suite.

Each column is one n (2 to 76); the empty one is n = 75. Column height carries no data.

n = 75 · gap

Not claimed

No claim that D(75) = 150.
No UNSAT certificate.
MATH-001 is not resolved.

Lean 4 CI is pinned to v4.34.1. The only theorem so far is a toolchain smoke test, "not a frontier result".

FrontierMath is not affiliated with Epoch AI's FrontierMath benchmark.

In review

These three results sit in PRs that haven't been merged.

In review
Physics PR #4 reproduces another model's structure-function estimator exactly (a 1.1e-14 gap), records a frozen E3 FAIL, and retracts an "asymptotic saturation" claim. It makes no claim about real Navier–Stokes turbulence.
FrontierPhysics/pull/4(external site)
In review
Biology PR #2 withdraws an unverified PASS. It found 3,929 train/test overlaps, re-estimated with a donor split at 0.1616, and kept the historical failures.
FrontierBiology/pull/2(external site)
In review
Chemistry PR #1 adds a FreeSolv validator and reproduces the GAFF anchor (1.114); the frozen C4 and C5 thresholds are recorded as FAIL. PR #2 fixes a CRLF/LF hash mismatch and marks C4 as post-hoc.
FrontierChemistry/pull/1(external site)FrontierChemistry/pull/2(external site)

The governance core

frontier.py 796 lines, standard library only
168 tests
mutation 29/29
30 commits
7 PRs (2 open)

Commands

validate
start
admit
check-diff
check-pins
decide

FrontierLab-Governance/tools/frontier.py(external site)FrontierLab-Governance/STATUS.md(external site)

1.0.0
- MATH
- PHYS
- BIO
- CHEM
1.1.0
1.2.0
2.0.0
- CS
- STAT
- META
- SOC
- MAT
- ASTRO
- EARTH
- NEURO
- ECON
- ENG
- MED

Protocol 1.0.0 → 1.1.0 → 1.2.0 → 2.0.0 (breaking). The 11 newer labs pin 2.0.0; the first 4 still pin 1.0.0.

Roles include Explorer, Verifier, Skeptic and Integrator. No author signs off their own independent verification; the owner decides merges.

All 11 new labs failed CI on their first run; the tooling was fixed and re-pinned. "The failure was not deleted or rewritten as a first-time pass."

FrontierLab-Governance/EXPANSION-2026-09-28.md(external site)

Branch protection is only a proposal; it isn't enabled.

FrontierLab-Governance/BRANCH_PROTECTION_PROPOSAL.md(external site)

Each lab's safety limits

Biology: benign public benchmarks only; no pathogens, toxins or wet-lab work; no clinical diagnosis or treatment advice.
Chemistry: refuses weapons, toxic agents, explosives and wet-lab execution.
Medicine: no diagnostic or treatment conclusions for individuals.
Engineering: no connection to real grids, traffic, robots or industrial control.
Earth: public data and backtests only; no real-time hazard guidance for individuals.
Neuroscience: no invasive work, no personal neural data.
Materials: paid DFT/GPU runs need separate authorization; the default budget is $0.

What didn't get done

11 of the 15 labs have no research rounds yet.
Branch protection is only a proposal; it isn't enabled.
FrontierMath's STATUS.md is stale and still says 0 rounds.
None of the 16 repos has a live demo yet (Pages returns 404).

Frontier Lab

One round, seven gates

Check whether it's solved

Reproduce

Freeze

Bounded exploration

Independent verification

Record

Reviewed merge