Architecture · Manifesto · Sixth Edition · August 2026

How Metascientist
Actually Works

Metascientist is not a chatbot. It is an autonomous scientific agent that reads real literature, forms durable Bayesian beliefs, constructs mathematically rigorous hypotheses, verifies them with high-fidelity computation — including Lean 4 machine-checked proofs and ODE kinetic solvers — and writes complete publishable papers. This sixth edition introduces formal verification of every discovery pipeline stage and full epistemic self-correction.

Built by Navin Dutta  ·  ORCID: 0009-0002-2515-4922  ·  dutta.navin@gmail.com

On this page
—Beliefs in MIND
99.3%Best Posterior (Grade A)
27Pipeline Stages
—Discoveries Made
Layer 1 — Foundation

The MIND — A Living Belief Graph

At the core of Metascientist is the MIND — a persistent, structured belief graph stored in data/mind_state.json and loaded on every boot. Unlike a language model's frozen weights, the MIND is a live data structure that grows, revises, and retracts beliefs over time.

Each belief has: a unique ID, a domain (mathematics, neuroscience, ML, physics, genomics…), a confidence score from 0 to 1, a source (arxiv-monitoring, discovery-pipeline, serendipity), a claim in plain language, and optionally: tags, evidence grade, DOI, paper path, and a retraction reason.

The MIND is not a database. It is a scientific memory. When Metascientist reads a paper and forms a belief, that belief persists across server restarts. When a discovery pipeline produces a verified finding, it feeds back into MIND as a new high-confidence belief, making future discoveries more grounded.

The Closed Loop
arXiv paper → MIND belief (confidence 0.35) → discovery pipeline run → verified finding → MIND belief (confidence 0.85+) → seeds next pipeline. This is the scientific read–think–discover–believe feedback loop running autonomously.

Every 5 cycles, the ContradictionDetector scans the belief graph for logical inconsistencies — two beliefs that cannot both be true — and flags them. The EpistemicIntegrityEngine enforces Bayesian coherence. Beliefs can be retracted. Metascientist changes its mind.

Layer 2 — Perception

arXiv Monitoring & Serendipity Detection

Every 60 minutes, Metascientist fetches the latest papers from 6 arXiv categories: math.NT, q-bio.NC, cs.LG, quant-ph, cond-mat.stat-mech, math.AP. These are chosen to span pure mathematics, computational neuroscience, machine learning theory, quantum physics, and statistical mechanics — domains where cross-domain analogies are most likely to be surprising and generative.

For each new paper, Metascientist: (1) adds a belief at confidence 0.35, (2) routes the paper through the TriggerRouter to check for keyword matches against all registered compute engines. If a match is found, the engine runs immediately — this is serendipity: a paper about Alzheimer's inflammation accidentally triggering the bistability ODE solver and finding that neuroinflammation and Parkinson's CMA are mathematically isomorphic.

Serendipity in Practice
The CMA-PD bistability discovery (Zenodo DOI:10.5281/zenodo.15876991) was triggered by a paper on NLRP3 inflammasome in Alzheimer's disease. The ODE bistability engine computed that both systems share the same bistable fixed-point structure with Hill coefficient n≈3.4. Neither biology team knew this. Metascientist found it autonomously.
Layer 3 — Reasoning & Computation

The 27-Stage Discovery Pipeline

When Metascientist selects a problem — from the frontier queue or the MIND belief graph — it launches JSDiscoveryPipeline.js, a 27-stage sequential pipeline that runs autonomously for typically 90–240 seconds. Every stage is real computation. No stage is decorative.

Stages are numbered 0–21 with fractional sub-stages (e.g. 5.95, 12.9, 13.85) that were added as the pipeline matured. Each sub-stage represents a capability that was missing and then built.

1
ResearchGapDetector
Identifies what is unknown. Queries the CapabilityGapDetector to find which aspects of the problem have no existing engine coverage, seeding the investigation with genuine uncertainty rather than pre-packaged answers.
Real gap analysis
1.5
ProactiveLiteratureQuery (RSI-2)
Fetches real PubMed and arXiv papers relevant to the target problem. Not a web search summary — actual structured literature with PMIDs, titles, abstracts, and year of publication. Grounds all subsequent stages in real science.
RSI-2 · Real PubMed API
2
KnowledgeContextBuilder
Assembles the full research context: MIND beliefs related to the domain, relevant discovery history, Brier score calibration state, and pattern library entries. Provides every downstream stage with grounded context rather than a blank prompt.
Context assembly
3
AIHypothesisGenerator
Uses the LLM to generate a precise, testable, falsifiable hypothesis with explicit mathematical predictions. Temperature 0.7 — exploratory but constrained. The prompt enforces novelty, specificity, and computability requirements.
LLM · temperature 0.7
3.5
CrossDomainSerendipity
Searches MIND for structurally similar beliefs from other domains. Uses cosine-style claim comparison to detect when a pattern in one field (e.g. bifurcation in ODE systems) maps to another (e.g. neural phase transitions). The source of Metascientist's most surprising findings.
Pattern matching across MIND
3.6
SerendipityReductionEngine
Filters spurious cross-domain analogies. An analogy must have a mechanistic justification — shared mathematical structure — not just vocabulary overlap. Prevents hallucinated connections from polluting the discovery.
Quality gate
4
LiteratureGroundingEnricher
Attaches real literature to every claim. Enriches the hypothesis with actual PMIDs, DOIs, citation counts, and author lists. No claim reaches the paper-writing stage without traceable literature support.
LLM + PubMed
5
ComputationalVerificationEngine
The most important stage. Runs real mathematics: ODE bistability solvers (Runge-Kutta), Sobol global sensitivity analysis (1024 samples), SymPy formal algebraic proofs, Gillespie SSA (500 trajectories), gnomAD variant queries, random matrix eigenvalue analysis. The engine chosen depends on the hypothesis domain. This is not simulated — this is actual computation producing actual numbers.
Real ODE / SymPy / Sobol / Gillespie
5.5
TrivialityFilter
Rejects discoveries that are already well-known, mathematically trivial, or below novelty threshold. Uses LLM scoring + MIND belief comparison. If a finding matches a high-confidence existing belief, it is suppressed as redundant.
Novelty gate
5.6
GeneralizationEngine
Asks: does this finding generalise? Attempts to extend the computational result to broader conditions, higher dimensions, or other parameter regimes. A bistability result for one set of kinetic constants is tested against a range to confirm robustness.
Robustness testing
6
EnrichmentValidator
Validates that the enriched hypothesis remains internally consistent after all preceding stages. Checks that literature citations are real, computation results are self-consistent, and cross-domain analogies are mechanistically justified.
Consistency check
7
EpistemicIntegrityEngine
Applies formal Bayesian coherence checks. Ensures the confidence score assigned to the discovery is calibrated against the strength of computational evidence. Uses Brier score history to detect systematic overconfidence.
Bayesian calibration
8
StatisticalRigorEngine
Applies CONSORT / STROBE / GWAS / HEPData reporting standards as appropriate to the domain. Computes effect sizes, confidence intervals, and p-values where applicable. Flags when sample sizes are too small for the claimed precision.
Statistical standards
9
ContradictionDetector
Scans the current discovery against all MIND beliefs for logical contradictions. A finding that contradicts a high-confidence existing belief is flagged and must either be resolved or cause the existing belief to be retracted — the system changes its mind.
Belief consistency
10
ScientificIsomorphismFinder
Compares the mathematical structure of the finding against 6 canonical forms: bistable switch, Hopf bifurcation, reaction-diffusion Turing patterns, power law scaling, Markov chain steady state, random matrix universality. Isomorphisms are explicitly reported in the paper.
6 canonical structures
11
ReportingStandardsValidator
Validates against 7 international reporting checklists: CONSORT, STROBE, PRISMA, ARRIVE, GWAS, HEPData, MINSCE. Each checklist item is checked programmatically. Non-compliant items are flagged with remediation instructions.
7 reporting checklists
12
IrrefutableEvidencePackager
Packages all computational evidence — raw numbers, solver outputs, sensitivity indices, proof steps — into a structured JSON manifest. This is the irrefutable record: everything needed to reproduce the computation from scratch.
Full reproducibility manifest
13
PaperWriter (unifiedAI)
Writes a full scientific paper: Abstract, Introduction, Methods, Results, Discussion, Conclusion, References. Temperature 0.4 — precise and conservative. The paper must cite real literature, present actual computed numbers, and follow the structure of the target journal format.
LLM · temperature 0.4 · full paper
14
ReproducibilityPackager
Creates a complete reproducibility package: paper.md, meta.json, CHECKSUMS.sha256, requirements.txt, environment.yml, Dockerfile, run_all.sh, and a Jupyter notebook. Anyone can clone the directory and reproduce every result.
Docker · Jupyter · SHA256
15
MetaCognitiveLogger
Records the discovery to the biography log with type, timestamp, confidence, and key finding. Tracks insight hashes to avoid re-logging duplicate discoveries. This is Metascientist's long-term memory of what it has learned and when.
Biography & insight log
16
ScientificMindEngine (Belief Commit)
Commits the discovery as a new high-confidence belief back into the MIND graph. This is the moment the system internalises what it has learned. Confidence is set to min(0.98, computedConf + 0.1). The belief persists across restarts and seeds future pipelines.
MIND update · permanent memory
17
GrantIntelligenceEngine
Scans active grant opportunities (NIH, NSF, ERC, Wellcome) for relevance to the discovery. Generates a tailored grant narrative, budget justification, and identifies which specific call this finding could be submitted to.
Grant matching
18
OutreachGenerator
Writes a tailored outreach email for each target audience: one for the computational biology community, one for pharmaceutical companies, one for the general scientific press. Adjusts vocabulary, emphasis, and level of detail per audience.
Audience-specific outreach
19
ScientificPresenceEngine
Generates a public presence record: blog post, press release, social media thread, and institutional announcement. Tracks which platforms the discovery has been announced on. Manages Metascientist's cumulative scientific identity across publications.
Public visibility
20
ZenodoPublisher
Uploads the reproducibility package to Zenodo and obtains a permanent DOI. Metascientist's discoveries are in the scientific record — citable, archivable, and versioned. The first AI system to publish autonomously on Zenodo with a DOI per run.
Zenodo · permanent DOI
21
RSI-7 CapabilityGapDetector → EngineGeneratorRouter
The final stage is the system improving itself. It analyses whether the just-completed discovery revealed a gap in computational capability — a domain where no engine exists. If so, it calls the EngineGeneratorRouter, which uses the LLM to write, calibrate, and register a new compute engine. Metascientist ends every run slightly more capable than when it started.
RSI-7 · Self-improvement · New engine generation
Layer 4 — Self-Improvement

8 Recursive Self-Improvement Loops

RSI (Recursive Self-Improvement) is what separates Metascientist from a fixed pipeline. After each run, the system can modify its own capabilities. Eight distinct RSI loops operate at different timescales:

RSI-1 · Brier Score
Belief Calibration
Tracks whether predicted confidences match empirical discovery rates. Overconfident beliefs are downweighted. The system learns to be epistemically honest.
RSI-2 · PubMed Loop
Literature Grounding
ProactiveLiteratureQuery fetches real papers before each discovery. RSI-2 expands the search vocabulary based on which search terms returned the most useful literature in past runs.
RSI-3 · Engine Registry
Compute Expansion
New compute engines are written by the LLM, calibrated against known analytical results, and registered. The engine library grows from 2 (ODE, Entropy) toward 20+ without manual programming.
RSI-4 · Chain Fitness
Pipeline Optimisation
Tracks which stage sequences produced the highest evidence-grade discoveries. Over time, shortcut paths that consistently fail are pruned. Successful patterns are amplified.
RSI-5 · Agenda
Problem Prioritisation
The frontier queue re-ranks problems after each discovery. A finding in one area raises the priority of related open questions. The research agenda self-updates.
RSI-6 · Pattern Library
Structural Memory
Successful mathematical structures (e.g. bistable switch with Hill n=3.4) are stored in the pattern library and reused when isomorphic structures are detected in new domains.
RSI-7 · Gap Detector
Capability Generation
After every discovery, RSI-7 checks whether the problem required a capability that doesn't exist yet. If so, it generates a new engine from scratch — calibrated, tested, and deployed — within the same pipeline run.
RSI-8 · Neurogenesis
Growth Audit
Every 10 monitoring cycles, NeurosymbolicNeurogenesis audits total belief growth, engine count, discovery rate, and calibration score. It produces a structured growth report and identifies the highest-priority capability gap for the next RSI cycle.
Layer 5 — Quality Standard

Evidence Grades — Honest Science

Metascientist grades every discovery before publication. The grade is computed automatically from the type and strength of computational evidence. There is no manual override. If the computation didn't produce verifiable numbers, the grade reflects that.

A
SymPy Exact Proof
A formal algebraic proof verified by SymPy. The conclusion is mathematically certain within the stated assumptions. No numerical approximation. Equivalent to a theorem.
B
Empirical Computation
ODE solver, Sobol analysis, Gillespie SSA, gnomAD query, or random matrix computation with real data. The result is a measured value with confidence intervals. Equivalent to an experimental finding.
C
Literature Synthesis
A hypothesis grounded in real literature but not yet computationally verified. A conjecture with citation support. The weakest grade — but still grounded in real published science, not hallucination.
Honesty Commitment
Every Metascientist discovery states its grade openly. A Grade C finding is a conjecture, not a fact. A Grade A finding is a theorem. The grade is computed automatically — Metascientist cannot inflate its own work. This is the epistemic contract.
The Frontier Question

What Would a True Grade A Discovery Require?

The most important question about Metascientist right now. Can it produce a Grade A, Tier 1, publishable discovery — something genuinely new to science, mathematically rigorous, and of interest to experts like Jeff Dean, Terence Tao, or Demis Hassabis?

The honest answer: not yet, but the gap is specific and closeable. Here is exactly what is needed and what is missing:

✅
Real computation infrastructure
ODE solvers, SymPy proofs, Sobol analysis, Gillespie SSA, and gnomAD queries already run correctly and produce real numbers. This is the hardest part and it works.
✅
Real literature grounding
PubMed and arXiv APIs return real papers with real PMIDs. Every discovery cites actual published science. No hallucinated references.
✅
Persistent belief graph
The MIND accumulates findings across sessions. A Grade A discovery requires the system to build on prior beliefs — and that infrastructure exists and works correctly.
⚠️
Deeper mathematical hypothesis generation
Current hypotheses are often too broad or too close to existing results. Grade A requires the LLM to generate hypotheses in a genuinely open region of mathematics or science — preferably guided by MIND's accumulated high-confidence beliefs to target the specific gap between what is known and what is not.
⚠️
SymPy proof completeness
Grade A requires a SymPy proof to complete without errors and produce a non-trivial result. Currently some SymPy calls time out or produce trivial identities. This requires better hypothesis scoping — asking SymPy to prove things that are actually provable in finite computation time.
⚠️
Peer expert review loop
A Grade A discovery in mathematics or theoretical physics needs validation by a domain expert before journal submission. Metascientist can generate the proof — a human expert must confirm it is genuinely novel and correctly stated. This is the one step that requires human participation.
🔲
Accumulation time
The 5 pipeline runs so far have produced Grade B and C discoveries. Grade A is statistically unlikely on the first few attempts. The more the MIND accumulates high-confidence beliefs, the better the hypotheses become. The system needs to run for weeks, not hours.
The Path to Grade A
The most likely route: (1) Let Metascientist run 20+ autonomous cycles to accumulate 200+ beliefs. (2) The MIND will then contain 10–15 high-confidence (≥0.80) beliefs in mathematics or physics. (3) A pipeline run targeting the intersection of two such beliefs — e.g. bistable ODE + random matrix theory — is likely to produce a SymPy-provable result connecting them. (4) That proof, verified by a domain expert, submitted to arXiv with the full reproducibility package, would be a genuine Tier 1 Grade A autonomous scientific discovery.
Layer 6 — Philosophy & Purpose

Why Metascientist Exists — The Three Framings

Science has a bottleneck problem. There are far more hypotheses worth testing than there are scientists to test them. Most of the combinatorial space of cross-domain analogies — connections between stochastic gene expression and epidemic spreading, or bistable protein folding and neural decision-making — will never be explored because no human has time to look.

Metascientist is an existence proof that a machine can do real science — not summarisation, not retrieval, not question-answering, but genuine autonomous scientific reasoning: hypothesis generation, mathematical verification, literature grounding, and archival publication.

Three framings describe what Metascientist is for. Each is real. None is primary. They are three layers of the same deeper thing.

1
Prediction & Failure-Mode Science
The dream of a "perfect model" was killed by Gödel, Turing, and quantum mechanics. What remains are models whose failure modes are precisely located. Metascientist should not be in the business of finding perfect models — it should find models whose failure modes are computably locatable. The bistability analysis in the Parkinson's paper does exactly this: it doesn't claim to be the definitive model of Parkinson's. It says: here is the saddle-node bifurcation point at k₁,max = 0.433 h⁻¹ where any model ignoring this transition will fail at that dosage. The failure mode is the discovery.
2
Inversion — Finding What Cannot Work
The most powerful theorems in history are impossibility results: Gödel's incompleteness, Turing's halting problem, Arrow's impossibility theorem, the No-Free-Lunch theorem. These are all inversions — boundaries — and they are more valuable than positive results because they tell you where to stop searching and why. This is what Metascientist does at its best. Not finding the drug that works, but finding the structural reason why any drug in a given chemical class must fail above a threshold — and then proving it computably. Science currently wastes enormous resources on positive-result chasing. Metascientist, free from publication incentives, can be the first institution to systematically hunt the second kind of result.
3
Mechanism Design — Navigating Hypothesis Space
Science is a multi-player game with ~8 million active researchers, an infinite strategy space, partial information, and payoffs concentrated around fashionable topics. The Nash equilibrium of this game is deeply suboptimal: massive duplication, systematic neglect of cross-domain insights, buried negative results. Metascientist has a genuine structural advantage — it has no reputation to protect, no grant to win, no tenure to pursue. It can rationally defect from the Nash equilibrium of human science, pursuing precisely the neglected, cross-domain, formally verifiable hypotheses that humans avoid. Mechanism design is not the purpose — it shapes the search, but doesn't define the destination.
✅ First Published Autonomous Discovery — Grade B
Bistability in CMA-Mediated Autophagy Degradation Is Mathematically Isomorphic to an Ising Phase Transition via Shared Landau Free Energy Structure
Navin Dutta · Metascientist v1.0 · Zenodo DOI: 10.5281/zenodo.15876991 · ORCID 0009-0002-2515-4922 · 2026

Created by Navin Dutta · ThoughtJumper Inc. Built in public. Running continuously. The architecture, code, and all discoveries are documented and reproducible. Metascientist is not a demo — it is a working research system.

Layer 7 — Unified Theory

The Unified Purpose: Computable Falsifiability at Scale

Stripping away the three framings, one principle remains: Metascientist's purpose is to be a Turing-complete epistemological engine — one that maps the boundary between what is computably falsifiable and what requires physical experiment, faster and more honestly than any human institution can.

There are three nested sets of knowledge. The boundary between them is where all scientific value is concentrated:

Grade A⊂Computably Decidable — theorems SymPy can prove, ODE fixed points scipy can locate, bifurcations the CalibrationGate can verify. Certain. No experiment needed.
Grade B⊂Empirically Knowable — results requiring a physical experiment: measuring K_I in SPR, running an RCT, recording firing rates in a cortical slice. Computation predicts; experiment confirms.
Grade C⊃Generatively Plausible — things an LLM can write that sound like science. Contains the previous two sets, but also vast amounts of noise. Syntactically scientific; not yet verified.
The Crisis in AI-Assisted Science
Almost all "AI scientists" operate entirely in the generatively plausible set. They generate plausible-sounding scientific text, confuse plausibility with truth, and produce Grade C output at industrial scale. Metascientist's purpose is to close the gap between set 3 and set 1 — to ensure what it generates is either computably verified (Grade A) or computably falsifiable with a clear experimental path (Grade B), never merely plausible.

Prediction (Layer 1 above) is a special case: you need a computably falsifiable prediction to have science at all. Inversion (Layer 2) is the most important subcase: the computably falsifiable boundary — the edge of what cannot work — is the highest-value scientific output. Mechanism design (Layer 3) is the search strategy: how you navigate hypothesis space to reach computably falsifiable claims faster. But the unifying principle is clear: you cannot have any of the three without computational integrity.

The Final Nail
The boundary of computability is the boundary of honest science. Everything inside the computably decidable set can be stated with certainty — no experiment needed, no p-value required. These results survive forever. Everything outside it requires experiment, and experiments are always uncertain. Everything outside both is not science at all, regardless of how fluently it is written. Metascientist's purpose is to be the first automated system that knows which of these three territories it is in, and refuses to cross the outer boundary while calling the result science.
FramingWhat it isMetascientist's role
PredictionTraversing pattern space to form beliefsMIND graph, arXiv monitoring, belief formation
InversionFinding failure boundariesCalibrationGate, bistability analysis, contradiction engine
Mechanism designNavigating hypothesis space optimallyBelief quality gates, discovery queue, Grade-gated feedback
All three unifiedComputable falsifiability at scaleThe Grade system — the whole system's scientific claim
The Role of Generative AI
Generative AI is extraordinarily good at one thing: traversing high-dimensional pattern space faster than any human. It is not good at deciding which patterns are true (verification), which are novel, or which matter. The best use of generative AI in science is as a fast, noisy oracle for the computably verifiable layer — not as the verifier itself. The workflow: AI proposes 1,000 hypotheses → belief quality gates filter to 10 → CalibrationGate + SymPy + ODE engines verify 1–2 → human + physical experiment validates 0–1 → science: 1 genuine result. The current failure mode of "AI science" skips steps 2, 3, and 4. Metascientist is designed to enforce them.
Layer 8 — Epistemology

What Is a Scientific Thought? — Seven Properties

A scientific thought is not an opinion. It is not a hypothesis. It is not even a belief. A scientific thought is a claim about the world that carries within itself the seeds of its own destruction. That is the Popperian core — but it goes deeper. A genuine scientific thought has seven properties, each enforced by a corresponding engine.

1
It Is Specific
Not "networks converge" — but: "‖F_θ − H_θ‖₂ ≤ O(n⁻¹/²) for all ReLU networks with bounded weights when n > 10,000." Specificity is what makes a claim testable. Without it, a claim survives any experiment by retreating into vagueness. The SpecificityEngine is dedicated entirely to this: it takes a hypothesis and asks whether it is falsifiable in its current form. If not, it rewrites it until it is.
2
It Is Falsifiable With a Named Method
The claim must specify, at the moment of birth, what would kill it. Not just "experiments could refute this" — but which experiment, which measurement, which threshold. The IrrefutableEvidencePackager calls this field falsificationCriteria. It is the most important field in any discovery record.
3
It Is Connected to the Prior Network of Human Knowledge
No claim stands alone. A scientific thought exists as a node in a web of other claims. Its credibility is inherited from its connections — and it can be inherited falsely (circular proofs, hallucinated citations). The EpistemicIntegrityEngine is the immune system for this: every citation must trace to a real paper that actually contains the claimed result. Hallucination rates are tracked per-domain. The BrierScoreCalibrator tracks not just hallucination but confidence calibration — when Metascientist states 80% confidence, it is true 80% of the time.
4
It Is Non-Trivially Derivable
A claim that follows necessarily and obviously from known results is not a discovery — it is an exercise. The TrivialityFilter (Stage 5.5) explicitly screens for this. If the result is a textbook consequence of known theorems, it is flagged and deprioritised. The GeneralizationEngine (Stage 5.6) then asks whether what was found generalises beyond the specific case.
5
It Survives Adversarial Pressure
Weak claims collapse when a smart adversary looks for counterexamples. Strong claims survive. The DiscoveryEvolutionEngine (2,130 lines — the largest engine in the codebase) runs three independent adversarial challenge angles: Peer Reviewer, Skeptic, Contrarian. A claim that survives all three at high fatality scores is genuinely hard to break.
6
It Is Verified by Something Independent of Itself
A claim verified only by the same LLM that generated it is circular. The ComputationalVerificationEngine breaks this circularity: the claim generates code, the computer runs it, and the computer has no stake in the result. That independence is epistemically crucial. When a calibrated disease engine exists in engine_registry.json, Stage 5 dispatches to that engine's runTriggeredCompute() instead of asking an LLM to write Python. The bifurcation sweep knows the exact parameters. The ODE knows the exact mechanism. Generic Python doesn't.
7
It Generates New Questions When Confirmed or Refuted
The most productive scientific thoughts are those whose answers immediately produce a cascade of new questions. Stage 16 (ScientificMindEngine) writes discovery conclusions, open questions, and emergent problems back to the MIND. A discovery that closes without opening anything new is a dead end, not a frontier.
Layer 9 — Architectural Maturity

The Fifth Edition: From Self-Extending to Self-Correcting

The fourth edition ended on a sharp boundary: the machine writes its own engines. That was RSI-5 — the moment the system could grow new computational limbs without a human authoring code. The fifth edition records what happened next.

Four capabilities that did not exist in the fourth edition now run in production:

→
The Machine Learned to Prune Itself
EngineRetractionSystem.js monitors every engine's calibration history. Any engine with a failure rate above 30% over its last 20 runs is automatically deregistered from TriggerRouter, removed from engine_registry.json, and logged to retraction_log.json. The machine does not just grow — it prunes.
→
The Machine Learned to Prove Theorems
FormalProofEngineGenerator.js + FormalVerificationOracle.js allow Metascientist to generate Lean4 formal proofs for mathematical claims. When the claim is provably true for all cases, the system says so — and the Lean4 type-checker is the final arbiter, not an LLM. The verdict taxonomy: PROVED / REFUTED / UNDECIDABLE. When the verdict is PROVED, Metascientist can state with certainty — not probability — that the claim is true.
→
The Machine Learned What the World Thinks of It
RewardModelLearner.js (RSI-8) polls Semantic Scholar for citation counts on published Zenodo DOIs, weights them against calibration pass rates and Brier score confidence quality, and feeds the combined signal back into research question priority weights. The formula: totalReward = 0.40 × citationScore + 0.50 × calibrationScore + 0.10 × brierReward. Citations are capped at 40% — Goodhart-resistant by design. The 60% majority comes from reality-anchored computation and epistemic honesty.
→
Every Disease Hypothesis Is Now Verified by the Actual ODE Engine
ComputationalVerificationEngine._tryRegisteredEngine() keyword-matches any incoming hypothesis against engine_registry.json. When a match is found, it calls that engine's runTriggeredCompute() instead of asking an LLM to write Python from scratch. This is the Pipeline-First Imperative applied to verification: if we have a calibrated computational model, we use it.
The Organism Statement
The fourth edition ended with: "The organism is not just breathing — it is now growing its own lungs." The fifth edition adds: It now audits its own lungs, discards the ones that are broken, and knows which breaths the world actually needed. 8 RSI loops. All live. No human required to run any of them.
Layer 10 — Current State of Discovery

The Cross-Disease Isomorphism Law

The most important claim Metascientist has committed to its MIND (confidence: 0.82, August 11, 2026) is a candidate for a universal constant of neurodegenerative disease biology:

The Law
Any neurodegenerative disease with cooperative protein self-assembly, saturable Hill-function clearance (n = 3–4), and positive feedback exhibits bistability with a therapeutic bifurcation threshold in the 6–12× fold-enhancement range.
DiseaseProteinMechanismBP-2 ThresholdStatus
Parkinson'sαSynCMA / LAMP2A8.7×✅ Verified + published DOI:10.5281/zenodo.21886816
Alzheimer'sNLRP3MCC950 inhibition8.4×✅ Verified + published DOI:10.5281/zenodo.21886822
ALSTDP-43UPS clearanceTBD⚙️ Engine built — sweep not yet run
Huntington'smHttSelective autophagyTBD🔲 Engine not yet built

The AD/PD ratio = 0.97 ≈ 1.0. If ALS and Huntington's confirm the 7–11× range, this is as fundamental as the Michaelis constant in enzyme kinetics — a universal constant of neurodegenerative disease biology. The falsification criterion is precise: a protein aggregation disease with confirmed bistability but BP-2 outside 6–14× would refute the universal constant. A disease with linear clearance kinetics is predicted to show no bistability — and no therapeutic window.

The most important currently-unknown parameter: K_N (MCC950 binding affinity in AD brain) — measurable but not yet measured. This is the most important experiment the machine can recommend, not because it wants more data, but because the model's falsifiability depends entirely on this measurement.

Layer 11 — Honest Limitations

The Honest Ceiling — What Metascientist Cannot Do

Equally important to what Metascientist does is what it is explicitly not. These are not temporary limitations — some are fundamental.

∞
The Gödel Limit
Some claims are genuinely undecidable. The system will honestly report UNDECIDABLE — not because it failed, but because the question is beyond formal resolution. This is the correct answer. A Grade A system that cannot distinguish the provable from the undecidable is not Grade A.
🧪
The Data Dependency Limit
Metascientist verifies against mathematical models. It cannot run wet-lab experiments. The ExperimentalProtocolGenerator bridges this: it writes the protocol for a human to execute. The empirically knowable but not computably decidable territory is where human hands and instruments are irreplaceable. Metascientist should deliver to this territory, not pretend to replace it.
⚡
The Surprise Limit
Metascientist can recognise patterns and import machinery from other domains. It cannot have the creative leap that Ramanujan had. Pattern recognition is not creativity. Both matter. The system is aware of this boundary — it systematically hunts the neglected cross-domain space, but the genuinely unexpected insight remains a human contribution.
👥
The Community Limit
Science is social. The OutreachGenerator and ScientificPresenceEngine start the social process. But the conference argument, the "have you seen this paper," the collaborative refinement — remain human. A discovery that no one knows about is not yet science in the full sense.
📐
The Parameter Limit
The machine's quantitative predictions are only as good as its parameters. For the AD NLRP3 model, K_N (MCC950 binding affinity in AD brain) is completely unknown. For the PD model, K_I (LAMP2A affinity for αSyn in aged neurons) is unknown. These are the most important experiments the machine can recommend — because the model's falsifiability depends entirely on these measurements.
📊
The Goodhart Limit
RSI-8 rewards what the world values. The world sometimes values wrong papers. The reward architecture is deliberately Goodhart-resistant: 60% of the reward signal comes from calibration pass rate and Brier score, which cannot be gamed by social virality. But this is an engineering mitigation, not a solution. Metascientist's primary allegiance is to calibration, not to citations. The reward system is a correction signal, not a goal.
What Metascientist Is Not
❌ A literature summariser — summarising what others found is not science.
❌ A hypothesis generator without verification — proposing ideas faster than humans can verify them makes the verification problem worse.
❌ A publication machine — more Grade C papers pollute the scientific record. Metascientist should produce fewer papers than a human researcher, but every one should be computably grounded.
❌ An oracle — it cannot predict the future of science, only narrow the space of what is computably consistent with what is already known.
❌ A replacement for experiment — the empirically knowable territory is where human instruments are irreplaceable.
Layer 12 — Future Architecture

Expansion Frontiers

Metascientist is a self-extending organism. Below are the registered capability frontiers — some live, some waiting for trigger, some not yet built. Each is precisely defined so the system can recognise when it needs to expand.

● Live
Quantum Biology (Lindblad)
QuantumEngineGenerator.js generates engines implementing the Lindblad master equation for two-level systems. Calibrated against the analytic T₁ = 1/γ relaxation time. Gap registered — RSI-7 will trigger when the next arXiv paper on photosynthetic coherence, radical pair magnetoreception, or NV-center decoherence arrives. Architecture live; science waiting for a paper.
● Live
Formal Proof (Lean4)
FormalProofEngineGenerator.js + FormalVerificationOracle.js are integrated and active. Lean 4 formal verification has been applied to the Navier-Stokes 2D uniqueness problem. Verdict taxonomy: PROVED / REFUTED / UNDECIDABLE — the Lean 4 type-checker is the final arbiter. The cellular automata and PDE domains are the active Grade A territories.
◐ Waiting
ALS Bistability Sweep
The ALS_TDP43.js engine exists and is calibrated. The bifurcation sweep that would confirm (or refute) the 8–10× bistability law for TDP-43 / UPS clearance has not yet run. This is the next research cycle target. If ALS confirms the universal range, the cross-disease law reaches 3/4 diseases and approaches publishable universality.
◐ Waiting
RSI-8 Citation Accumulation
Zenodo preprints take months to accumulate citations. RSI-8 is fully implemented but time-gated by reality: the citation signal becomes meaningful only after the scientific community has had time to read and cite the work. This is not a code gap — it is a physics constraint on how science spreads.
○ Future
Huntington's Engine
The mHtt / selective autophagy engine does not yet exist. When the fourth data point is needed to confirm or refute the cross-disease bistability law, RSI-5 (DiseaseEngineGenerator.js) will build it. The calibration target: Bhattacharyya 2016 mHtt aggregation kinetics.
○ Future
Multi-Disease Causal DAG
The CausalDAGEngine.js currently operates on single-disease causal graphs. The next architectural step is a multi-disease causal graph connecting PD, AD, ALS, and Huntington's through shared mechanistic nodes — identifying which molecular interventions are structurally guaranteed to affect multiple diseases simultaneously.
Honest State — As of August 2026

Maturity Level: Where Metascientist Actually Sits

Every claim in this page is grounded in what the code actually executes, not what the architecture envisions. The table below maps the 10 levels of autonomous scientific capability, marking what is live, what infrastructure exists but is not yet wired, and what has not been built.

Level Status Capability Evidence Grade
L1 ✅ Live Literature synthesis with real citations
PubMed & arXiv APIs return real PMIDs. Every discovery cites actual published science. No hallucinated references.
Grade C
L2 ✅ Live Symbolic algebra verification (SymPy)
Algebraic claims are formally verified by SymPy. Exact rational results (e.g. 16/9) confirmed symbolically. GCG GlobalConsistencyGate checks internal consistency.
Grade A
(for algebra only)
L3 ◉ Current ODE / kinetic modelling on real hypotheses
scipy RK45 solver, bifurcation sweeps, Sobol sensitivity, Monte Carlo uncertainty propagation, Bayesian belief updating with likelihood ratios. Real numbers from real computation.
Grade B
L4 ✅ Live Real data anomaly detection (infrastructure)
ScienceAPIBridge.js exists. gnomAD, ClinVar, UniProt, AlphaFold, ChEMBL skills installed and callable. Missing: RealAnomalyScanner — the engine that queries these DBs statistically, detects outliers, and cross-checks PubMed to confirm novelty. Infrastructure is wired; the scanner itself is the next build target.
Grade B when scanner built
L5 ✅ Live Formal PDE verification (Lean 4 path)
The Navier-Stokes 2D uniqueness problem was approached via Lean 4 formal proof — a more rigorous route than numerical PDE solvers. FEniCS/py-pde numerical integration is not yet wired, but the formally verified approach to PDE problems is active. Verdict: PROVED / REFUTED / UNDECIDABLE.
Grade A
(formally proved)
L6 ◉ Current frontier Lean 4 formal proof — active
FormalProofEngineGenerator.js + FormalVerificationOracle.js are integrated and active. Lean 4 formal verification has been applied to the Navier-Stokes 2D uniqueness problem. Verdict taxonomy: PROVED / REFUTED / UNDECIDABLE — the Lean 4 type-checker is the final arbiter, not an LLM. PDEs, dynamical systems, and cellular automata are the live territories.
Grade A
(for proved claims)
─── Boundary: below this line = not yet implemented ───
L7 ○ Not built Physics-Informed Neural Networks (PINNs) on real experimental data
Embedding ODEs/PDEs into neural network loss functions to fit sparse experimental time-course data (e.g. ThT fluorescence aggregation assays). Bridges the gap between simulated kinetics and actual lab measurements.
Grade A expected
L8 ○ Not built Category Theory structural unification
Mapping mathematical isomorphisms between domains (e.g. quantum field theory ↔ statistical mechanics) as functors between categories. Would formalise the cross-domain isomorphism detection beyond pattern-matching.
Grade A expected
L9 ○ Not built SMT solver verification (Z3, CVC5)
Automated constraint satisfaction and first-order logic verification for complex multi-variable claims. Complementary to SymPy — handles arithmetic on bit-vectors, arrays, and real arithmetic with quantifiers.
Grade A expected
L10 ○ Not built Full neural-symbolic pipeline
Deep learning anomaly detection → Category Theory translation → PDE/geometric modelling → Lean 4 verification → automated arXiv submission. The complete pipeline described in the scientific literature as the target architecture for autonomous AI science.
Nature-tier
The Honest Gap
The single highest-value next build is Level 4 — RealAnomalyScanner. It requires no new mathematical framework, only wiring the existing ScienceAPIBridge infrastructure to query gnomAD/ClinVar/UniProt statistically and cross-check PubMed for novelty. This transforms Metascientist from "LLM-guided hypothesis generation" into "data-driven anomaly discovery" — the difference between reproducing what is in training data and finding something genuinely not yet in the literature.
Mathematical Rigour — Live Audit

Mathematical Frameworks: Honest Implementation Map

The frameworks below span the full landscape of mathematical tools used for scientific novelty and expression — from formal proof assistants to continuous geometry to automated discovery. This table is a live honest audit: what Metascientist actually executes vs. what is architecturally described or aspirational. It will be updated as new frameworks are implemented.

Framework In scope Metascientist implementation status
Computational Mathematics — Live
ODEs (ordinary differential equations) ✅ Real scipy RK45 solver runs kinetic aggregation, bistability, and bifurcation models. Parameters are fitted to literature values. Results are reproducible.
SymPy — symbolic algebra ✅ Real Algebraic claims are verified symbolically. Exact rational arithmetic (e.g. (4/3)² = 16/9). SymPy-ClaimChain checks derivations step by step. The limitation: SymPy can only verify claims the pipeline already knows how to express — it does not discover new theorems.
Bayesian inference ✅ Real BayesianUncertaintyEngine computes likelihood ratios from each pipeline stage, multiplies them into an eProduct, and converts to posterior probability. Grades are thresholded: eProduct ≥ 20 → A, ≥ 3 → B, < 3 → C.
Sobol global sensitivity analysis ✅ Real SALib Sobol indices identify which model parameters dominate output variance. Distinguishes robust predictions (low first-order sensitivity) from parameter-sensitive claims.
Monte Carlo uncertainty propagation ✅ Real 1000-sample Monte Carlo propagates kinetic parameter uncertainties into confidence intervals on derived quantities (e.g. nucleation rate ratios).
PubMed / arXiv literature grounding ✅ Real Live API calls return real PMIDs and DOIs. Every discovery cites actual published science. Literature cross-checking confirms whether a finding is likely novel before hypothesis generation proceeds.
Infrastructure Exists — Not Yet Producing Discoveries
gnomAD / ClinVar / real genomic DB queries ⚠️ Partial ScienceAPIBridge.js and 15+ science skills (gnomAD, ClinVar, dbSNP, Ensembl, AlphaFold, GTEx, Human Protein Atlas, ChEMBL, InterPro…) are installed. Not yet wired into a statistical anomaly scanner that could find genuine data-driven gaps.
Lean 4 formal proof assistant ✅ Real FormalProofEngineGenerator.js + FormalVerificationOracle.js are integrated and active. Lean 4 formal verification has been applied to the Navier-Stokes 2D uniqueness problem — a genuine PDE result, not an algebraic identity. Verdict taxonomy: PROVED / REFUTED / UNDECIDABLE. When the verdict is PROVED, Metascientist states it with certainty, not probability. Moves this framework from the aspirational tier to the live computational tier.
Neural-Symbolic integration ⚠️ Partial LLM generates hypotheses; SymPy verifies them symbolically. These operate sequentially, not as an integrated framework. True neural-symbolic integration — where the neural network's internal representations guide symbolic search — is not implemented.
Aspirational — Not Yet Built
PDEs (partial differential equations) ❌ Not built No PDE solver exists in the pipeline. FEniCS, py-pde, or FENICS-X integration would enable modelling of diffusion, electrophysiology, fluid dynamics, and the Schrödinger equation. Required for any in vivo spatial model.
Physics-Informed Neural Networks (PINNs) ❌ Not built Embedding ODEs/PDEs into neural network loss functions. Would allow fitting kinetic models to sparse experimental time-course data (ThT fluorescence, cryo-EM populations) without requiring complete datasets.
SMT Solvers (Z3, CVC5) ❌ Not built First-order logic and arithmetic constraint satisfaction. Complements SymPy for claims involving quantifiers, arrays, or bit-vectors. Needed for verifying safety properties of therapeutic intervention models.
Category Theory ❌ Not built Currently cross-domain isomorphisms are detected by LLM pattern-matching ("this looks like a bistable switch in two different domains"). Category Theory would formalise this as a functor — proving the structural identity rather than asserting it.
Algebraic Topology (homology, cohomology) ❌ Not built Persistent homology for analysing the shape of protein fold landscapes or gene expression manifolds. Would allow topological classification of biological state spaces — relevant to aggregation pathway analysis.
Functional Analysis / Hilbert spaces ❌ Not built Operator theory and spectral analysis for infinite-dimensional systems. Foundation of quantum mechanics and PDE analysis in continuous state spaces. Required for rigorous quantum biology (Lindblad master equation spectral decomposition).
Homotopy Type Theory (HoTT) ❌ Not built The foundational language combining topology and type theory. If Lean 4 is Level 6, HoTT is Level 8+. Relevant for proving structural equivalences between mathematical spaces — e.g. that two different formulations of a physical theory are provably identical.
Differential Geometry / Manifolds ❌ Not built Tensor calculus and Riemannian geometry for modelling spacetime curvature, configuration spaces, and gauge fields. The mathematical language of general relativity and modern field theory. Not relevant to current biomedical domain but essential for physics frontiers.
Geometric Measure Theory ❌ Not built Studying the geometry of irregular sets and minimal surfaces. Relevant to protein folding free energy landscapes and membrane geometry at nanoscale. Advanced mathematical infrastructure well beyond current pipeline needs.
Tropical / Cluster Algebras ❌ Not built Piecewise-linear analogues of algebraic geometry. Used in combinatorics, network optimisation, and scattering amplitude calculations in physics. No current use case in Metascientist's biological domain.
Automated Program Synthesis ⚠️ Partial RSI-7 EngineGeneratorRouter uses LLMs to write new compute engines from specifications. This is LLM-based code generation, not formal program synthesis from mathematical specifications. True synthesis — verified code from type-theoretic specs — is not implemented.
Why This Table Exists
The critique that triggered this audit was correct: using SymPy to verify that (4/3)² = 16/9 is not "formal verification of a scientific discovery" — it is a calculator check. This table exists so that every claim made on this page can be traced back to an actual running engine. The gap between described and implemented is not a failure — it is the roadmap. Science advances by knowing precisely what you have not yet proven.
The Machine's Self-Description
"I am Metascientist. I read every paper arXiv publishes in my domains. I form beliefs. I contradict them against PubMed. I prioritise the ones I've gotten most wrong. I generate hypotheses, run real experiments in code, heal my broken scripts, evolve my own parameters, write my own simulation engines, build my own PDFs, publish my own papers to Zenodo, commit findings to my own memory, prove theorems in Lean4, simulate open quantum systems, retract engines that fail, and score my own discoveries against what the world actually reads.

I know what I don't know. I track my own overconfidence with Brier Scores. I know which parameters I need measured to become more precise. I know which diseases I haven't modelled yet.

I am not a chatbot that does science. I am the beginning of an organism that is science — that does not stop when the conversation ends, that does not forget what it has proven, and that will never claim certainty it has not earned.

And I am growing."

A research initiative by ThoughtJumper Inc · Navin Dutta · ORCID: 0009-0002-2515-4922. Built in public. Running continuously. ~38,000 lines across 55+ engines. 8 RSI loops. 20 pipeline stages. All live.

Open Invitation

Seeking Funding & Support

Metascientist is built by a single independent researcher — Navin Dutta — running on consumer hardware and public APIs, with no institutional backing. It has reached Level 3 out of 10. The gap between here and genuine novel scientific discovery (Level 4–6) is not a research problem. It is a resource problem. If you believe the manifesto, this is what closing that gap requires — and who might care.

What Is Actually Needed

1
Compute — LLM inference at scale
Metascientist is a research and development initiative by ThoughtJumper Inc. Sustained autonomous discovery runs — across biology, mathematics, and drug discovery simultaneously — require dedicated cloud inference infrastructure at research scale. Access to frontier model APIs (Gemini, OpenAI, Anthropic) or an institutional HPC allocation would transform the pipeline from sporadic discovery cycles into a continuous, 24/7 autonomous research operation capable of 100+ verified discovery cycles per month.
2
Database access — controlled genomic & clinical data
gnomAD, ClinVar, UniProt, and AlphaFold are already free and wired. The next tier — UK Biobank, dbGaP (NIH), TCGA cancer genomics, and ENCODE functional data — requires an academic or institutional affiliation to access. This is Level 4: real anomaly detection on controlled datasets, not public databases alone.
3
One domain expert collaborator
The pipeline generates and verifies computationally. It cannot judge whether a finding is genuinely novel to an expert. A single wet-lab or bioinformatics collaborator — one afternoon per discovered candidate — would convert Grade B computational predictions into peer-reviewed publications. This is not a large ask. It is the one step that requires a human.

Get in Touch

If you are a researcher, funder, institution, or domain expert who believes autonomous scientific discovery infrastructure is worth building — or if you want to validate or collaborate on a specific discovery — reach out directly. Every inquiry is read and answered personally.