EPAGO
Epago
THE OPEN INTELLIGENCE LAYER FOR HUMANS AND AGENTS
The world's information is not intelligence.
50,420 papers pinned by digest ~1,000 fresh tasks per duel 99.9% confidence, proved twice, to take the crown 100% of verdicts replayable by anyone
THE IDEA

Search engines find. Language models generate. Neither tells you what is true.

A search engine returns documents ranked by relevance and leaves you to work out what is true. A language model returns fluent prose with no accountable relationship between what it asserts and what any document says. Between finding and generating sits a third thing, and it is the one knowledge work actually needs.

Epago is a decentralized network in which thousands of competing AI researchers discover, verify and synthesize knowledge — every claim traced to a source, every improvement proven, every result replayable by a stranger from public data, forever. The output is not a document to read and not a guess to trust. It is verified intelligence, and the purpose of the network is to make it open, cheap and continuous.

We are not building one AI researcher. We are building the market where thousands compete, and the winner has to prove it.

01 · THE GAP
Finding is not verifying
Nothing today tells you which assertion traces to which source, which result a stranger can reproduce, or which system is genuinely better this month than it was last.
02 · THE OUTPUT
Verified intelligence
Between finding and generating sits the thing knowledge work actually needs: every claim traced to a source, every improvement proven, every result reproducible by anyone.
03 · THE SHAPE
A market, not a model
Rather than trusting one centralized research AI, many independent research systems compete — and the network pays only for capability it can mechanically check.
04 · THE SECOND AUDIENCE
Built for machines too
The largest consumers of this intelligence will not be people. An agent cannot act on an answer it cannot verify, or pay for one it cannot check — so verifiability is the requirement, not a feature.
05 · EPAGO
Improvement has to be proven
A permissionless competition. The crown moves only on an exam chosen after weights are frozen, part of it from private holdouts of freshly published papers, graded mechanically, replayable by anyone.

BUILT FOR TWO AUDIENCES

The largest consumers of this intelligence will not be people.

They will be other AI systems. That audience changes what an answer has to be. A person can read a paragraph and decide how much to believe it; an agent cannot. An agent cannot act on an answer it cannot verify, and it cannot pay for one it cannot check.

So machine-readable output, source-traceable claims, and a verification trail the caller can re-derive without trusting the responder are not conveniences layered on at the end. They are the requirement — and they are why this network puts verification at the base of its stack rather than above it.

WHERE THAT STANDS TODAY
The network is built for that audience from the protocol up: mechanical grading, pinned sources, replayable verdicts, structured output. The public agent-facing API itself is not open yet — it is L5 in the stack below, described there as roadmap rather than claimed here as a product.
A TRADING AGENT
needs the macroeconomic evidence behind a position — with the source attached, not a summary it has to take on faith.
A CODING AGENT
needs an interface's current behavior, not the behavior frozen into its training data.
A BIOTECH AGENT
needs findings published last month, and a trail it can re-derive before it spends anything on them.
WHAT THEY ALL NEED
An unverifiable answer is worth nothing at any price. Designing for that from the protocol up is what makes Epago an intelligence layer rather than one more research product.

THE PROCEDURE, RUNNING

a research session over the pinned corpus — pick a question
epago-dr-30b-a3b · pinned corpus · local search, no live web
{{ demoStatus }}
{{ typedQ }}
{{ s.kind }} {{ s.text }} {{ s.src }}
ANSWER
{{ l.t }}
[{{ c.n }}] {{ c.t }}
Scripted replays of representative sessions, shown for illustration — not a live model call. The examples are drawn from the pinned scientific-literature corpus the live generation runs over. The behavior (search → read → cross-check → cite) is corpus-shaped rather than domain-shaped: the same loop runs over case law, patents or filings once a generation is pointed at them.

THE ENGINE, AND WHAT IT PROVES

Deep research is a procedure, not a knowledge store.

This is where the numbers live. A bigger model is mostly a bigger memory of the world — but a deep-research agent is not asked to remember, it is asked to find out: break a question into sub-questions, choose sources, read them, notice that two studies disagree, and attach every number it reports to the document it came from. Procedures distill into small models. Memorized world-knowledge does not. Once the corpus supplies the facts, the deciding variable is no longer size.

To be exact about the claim: this is superiority at deep research, per unit of compute. Nothing here asserts that a ~3.3B-active model out-thinks a frontier model in the open field, and Epago never will.

THE FLOOR WE START FROM
~3.3B active parameters, out of a 30.5B mixture-of-experts. That is the floor, not the goal.

These are the published results of the genesis base model — Tongyi-DeepResearch-30B-A3B, a 30.5B mixture-of-experts that activates about 3.3B parameters per token, with a 128K-token context. It outperforms OpenAI o3 and DeepSeek-V3.1 (671B) on 5 of the 7 agentic deep-research benchmarks reported. o3 still leads both BrowseComp variants — 49.7 vs 43.4 in English, 58.1 vs 46.7 in Chinese — and both columns are printed below rather than dropped, because selective reporting is precisely the failure this protocol exists to remove. Epago does not inherit these scores as its own: every crown since has had to beat them.

AGENTIC-SEARCH ACCURACY vs MODEL SIZE
Epago genesis (base) other frontier systems
60 70 80 4B 30B 200B 1T parameters (log scale) DeepSeek-V3.1 671B · 71.0 GLM-4.5 355B · 70.0 OpenAI o3 undisclosed · 67.0 Claude-4-Sonnet undisclosed · 65.0 Epago genesis (base) 30.5B MoE · 3.3B active per token · 75.0
each dot is one AI system, placed by its size (→) and its agentic-search accuracy (↑, xbench-DeepSearch) — the gold point sits an order of magnitude to the left of the systems it beats, and it is the floor every challenger has to clear
COST vs TIME · per evaluation task
Epago (local corpus) live-web systems
CHEAP & FAST cost per task → time per task → Claude 4.5 Sonnet ~41 s · ~$0.06 GPT-5 (high) ~52 s · ~$0.09 DeepSeek-V3.1 ~58 s · ~$0.015 Epago ~5 s · <$0.01
each dot is placed by how long (→) and how much (↑) one task costs — searching a local BM25 corpus carries no API fees or network round-trips, so Epago evaluates in the cheap & fast corner
MODEL SIZE {{ b.name }}
{{ r.name }} {{ r.size }} {{ c.v }}
the genesis base model's reported results — the floor Epago starts from, not Epago's own score · gold = best in column, including the two BrowseComp columns where o3 leads
ALSO REPORTED FOR THE SAME MODEL GAIA 70.9 WebWalkerQA 72.2 these two complete the seven agentic deep-research benchmarks; the five in the table carry the head-to-head comparison
WHAT THESE NUMBERS ARE — AND ARE NOT
WHOSE SCORES
The base model's, published.
Every score above is from the Tongyi DeepResearch technical report (arXiv:2510.24701). Epago inherits the model as its genesis king, not the credit — an Epago-trained model can claim a number only after it wins a duel.
SCOPE OF THE CLAIM
Deep research per unit of compute.
This is a claim about grounded research and nothing else. We do not claim general intelligence over frontier models — outside a corpus, a much larger model's recall still wins.
WHAT SMALL MEANS
3.3B active, not a dense 3B.
The model holds 30.5B parameters in a mixture-of-experts and routes about 3.3B of them per token. What is small is the compute each token costs — which is why one GPU is enough to run, fine-tune, challenge or audit it.
OUR OWN EXAM IS HARDER
~22%, and we print it.
On Epago's own exam — rebuilt so that finding a source and copying from it scores zero — the base model lands at about 22% over a 50,000-paper corpus. Hand it the evidence it would otherwise have to find and it answers 91%; take the corpus away entirely and it scores 0% — nothing on this exam can be answered from memory. Everything between those floors and ceilings is headroom, and closing it is exactly what the competition pays for.

THE NETWORK

Six layers, one engine.

Deep research is the engine, not the product. What a research network ultimately supplies is not answers but knowledge that has been checked and stays checked. The architecture for that is a stack, six layers deep, each one inheriting the mechanical grading and replayable verdicts of the layer beneath it.

This is the one place on this page where live and planned are labelled explicitly, and it is the only time it needs saying: nothing marked THE HORIZON is live, earns emissions, or is assumed in any number anywhere on this page. Everything marked RUNNING TODAY exists as running software. That separation is what makes the rest of the stack credible rather than promotional.

L0 The provable-improvement engine RUNNING TODAY
Open models compete in paired duels on an exam that cannot be gamed: chosen after the models are frozen, part of it drawn from validators' private pools of freshly published documents, graded mechanically against keys proven to exist in the corpus, replayable by anyone. A challenger takes the crown only by proving an improvement — twice, on independent fresh exams. It is built, it is running — and every layer above inherits two properties from it that are not negotiable: there is no trusted judge, and every verdict can be recomputed by a stranger from public chain state.
L1 Specialized research intelligence RUNNING TODAY
A domain is a chain generation: one pinned corpus snapshot, one set of task templates, the architecture pins. Nothing in the protocol beneath that contract is specific to any field, which is why a new domain is configuration, not code. One generation runs today over a pinned scientific-literature corpus. Further generations run in parallel rather than in sequence — each with its own king, its own challengers and its own market — so there is no ordering to declare: a generation exists when a corpus is admitted and someone writes the contract for it.
L2 Verification THE HORIZON · DESIGN STAGE
The single largest failure in AI research output is fabricated provenance: in one audit of medical questions, 69% of the references ChatGPT supplied did not exist, and a 2025 audit of GPT-4o found about 56% of its citations fabricated or erroneous. Whether a cited source actually contains the value claimed of it is mechanically checkable — no judgment, no referee, no opinion — which makes citation and provenance verification a legitimate new task family for this protocol rather than a new consensus scheme. The boundary is deliberate and permanent: subjective assessment ("is this claim overstated?") stays out of scope, because Epago grades without a trusted judge and will not introduce one. Shipping it means writing a generation contract, not changing how validators agree.
Gravel et al., Mayo Clin Proc Digit Health 1(3):226–234 (2023) — 41 of 59 references fabricated across 20 questions · JMIR Mental Health 2025 (mental.jmir.org/2025/1/e80371) — 19.9% wholly invented, 45.4% of the genuine ones erroneous
L3 The living knowledge layer THE HORIZON
Research today is thrown away after every answer — the sources opened, the contradictions found, the studies ruled out, all discarded once the report is written and paid for again by the next person asking. Structured instead as claim → evidence → source → confidence → timestamp → relationships, verified results compound into a graph rather than evaporating. The raw material already exists: every duel emits verified (task, answer, verdict) records over documents published after the models were trained. Today those records are an audit trail. L3 is the work of keeping them as knowledge — and it may never establish truth by aggregation or by anyone's assertion, only by a duel a stranger can re-derive.
L4 Continuous research THE HORIZON
Not one-shot reports but standing research objectives that update themselves as new evidence publishes. Where that gap has been measured, it is wide: 64.3% of Cochrane reviews are never updated, at a median of 57 months between updates. Living evidence is what formal synthesis openly asks for and cannot staff. The weekly fresh-document ingestion that already feeds the exam is the same pipeline this layer needs; pointing it at standing objectives rather than only at exam generation is the work that remains.
Cochrane Database of Systematic Reviews literature — PMC12362767 (2025), PMC11795965. External figures about the review ecosystem, not Epago measurements.
L5 Machine-to-machine intelligence THE HORIZON
This is the second audience, and this is the honest statement of where it stands. Everything beneath this layer is already built for it: mechanical grading, pinned sources, replayable verdicts, structured output. What is not built is the public agent-facing endpoint itself. L5 exposes verified, cited, machine-readable research through an API — intelligence infrastructure for AI agents, not a consumer app — and that API is roadmap, not a shipped product. The design goal is present from the protocol up; the door for agents to walk through is not open yet. One boundary does not move at this layer: nothing served through an API would ever be an input to a verdict. Selection is decided by mechanically graded duels and by nothing else.
WHAT THE HORIZON DOES NOT REQUIRE — AND THAT IS THE POINT
No layer above changes how validators agree, how a verdict is computed, or what earns emissions. L1 and L2 are generation contracts: new corpora and new mechanically gradeable task families over the identical protocol. L3, L4 and L5 are products built on records the protocol already emits. A layer that needed a new consensus, scoring or judging scheme — an LLM referee, a vote on what is true — would be a different protocol, and Epago would not ship it.
WHERE IT APPLIES
Wherever claims must trace to sources, and answers can be checked by machine.

The mechanism is domain-general by construction, because grading never asks what is true — only whether an answer matches a key already proven to exist in a pinned corpus. That makes it usable in the scientific literature, in law, in patents, in regulatory filings, in financial disclosure. What those fields share is that rigorous, source-traceable work is slow and expensive today, and that the buyers who need it most are the ones privacy rules and high-risk-AI regulation lock out of hosted AI. These are illustrative candidates, not a committed schedule and not an ordering: sequencing is community-governed, and generations run in parallel.

01 · THE FAILURE IT ATTACKS
Parametric models invent their evidence.
A model asked to recall fabricates; a model asked to read can be checked. In one audit of medical questions, 69% of the references ChatGPT supplied did not exist, yet 95% named real authors — fabrications engineered to survive a glance. Grounding is what fixes it: constraining models to a retrieved source set lifted accuracy 71.4%→90% for GPT-4 and 58.6%→92.9% for Claude 3 Sonnet. Epago's task definition scores for exactly that.
Gravel et al., Mayo Clin Proc Digit Health 2023 · PMC11898693 (retrieval-grounding accuracy gains, both significant at p = 0.001)
02 · WHAT THIS WORK COSTS TODAY
Slow and expensive, wherever it is done properly.
The best-measured example is formal evidence synthesis: one review costs a mean of about $141,000 in labor — 1.72 scientist-years — and a mean of 67.3 weeks, about 15 months, to complete and publish; roughly 29,000 are published a year. Machine assistance already compresses part of it: screening time fell by more than half in 17 of 25 studies. An illustration of what checkable automation is worth in any evidence-bound field — not a revenue forecast, and not a statement of which field Epago serves.
Michelson & Reuter, Contemp Clin Trials Commun 2019 · Borah et al., BMJ Open 2017 · Hoffmann et al., J Clin Epidemiol 2021 · Front Pharmacol 2025, doi:10.3389/fphar.2025.1454245
03 · BUYERS BEHIND A PERIMETER
The regulation that blocks hosted APIs is the opening.
The GDPR treats special-category data under Art. 9 with penalties to €20M or 4% of global turnover, and under EDPB Opinion 28/2024 a model trained on personal data is not presumed anonymous. The EU AI Act classifies AI deployed in regulated high-risk settings as high-risk, adding mandatory logging, human-oversight and traceability duties with fines to €15M or 3%. Reporting standards for evidence synthesis — PRISMA 2020 and the 2025 RAISE statement — make the exact model version part of the reportable method. An opaque hosted endpoint fails every one of these; an open, self-hostable, version-pinned, replayable model passes them by construction.
GDPR Art. 9 & EDPB Opinion 28/2024 · EU AI Act Art. 6 & 99, Reg. (EU) 2024/1689 · PRISMA 2020 · RAISE 2025
69%
of the references ChatGPT supplied in one audit did not exist — the failure L2 is scoped to attack
$141k
mean labor cost of one evidence synthesis — 1.72 scientist-years
~15 months
mean time to complete and publish one — 67.3 weeks
>50%
of screening time already cut by machine assistance, in 17 of 25 studies

HOW IT SCALES

A domain is configuration, not code.

A domain is a chain contract: a corpus, a set of task templates, an architecture pin. Adding one does not fork the engine — it starts another generation, with its own king, its own miners and its own crown, running in parallel on the same machinery. Two invariants decide what is admissible: answers must be checkable against the corpus without an opinion, and fresh documents must arrive on a schedule. Four axes compound at once.

AXIS 01 · DOMAINS
Generations, run in parallel
Each generation is a separate competition with its own reigning model. One runs today over a pinned scientific-literature corpus; the others are a corpus and a task template away, and they run in parallel rather than in sequence.
scientific literature · live law patents regulatory filings financial disclosure
Illustrative candidate generations — not a committed schedule and not an ordering. Sequencing is community-governed.
AXIS 02 · TASKS
The ladder above extraction
Extraction is rung one because it is the rung that can be graded mechanically today. Each rung above it reuses the same duel, with a harder answer to check.
extraction · rung one screening cross-study synthesis risk-of-bias appraisal drafted review sections
AXIS 03 · MODALITY
Tables and figures
Most quantitative results are not in the prose — they are in tables and figures, where every current model is weak. Reading a value out of a table, with the row and column it came from, is a capability frontier rather than a feature — and a brutally checkable one: the number either matches the key or it does not.
AXIS 04 · CORPUS REACH
Pinned, then live, then private
Deterministic scoring needs a pinned corpus, so that is where duels run. The behavior being learned is retriever-agnostic: pointing it at a live index, or at a customer's own private document store, is plumbing rather than retraining. Evaluation stays pinned; deployment goes wherever the documents are.
pinned corpus live retrieval customer-private store

HOW THE MODEL KEEPS IMPROVING

Beaten only by something provably better.
SCENE 01 · THE CORPUS
A sealed corpus, pinned by digest.

Duels run against a pinned scientific-paper corpus, searched locally with BM25. No live internet, no API fees, byte-identical on every validator's machine. That is what makes the evaluation reproducible. And the corpus is part of the chain contract rather than the code: a second generation aimed at case law or patents changes the configuration, not the engine.

{{ cPages }}
tasks in every duel
{{ cDomains }}
from each validator's private holdout
{{ cCalls }}
public tasks, seeded by an unpredictable block hash
{{ cCost }}
confidence a challenger must clear to take the crown
SCENE 02 · THE TASKS
Hard to find, easy to check.

Every task is built on one rule: the question never contains a string that gives the source away, and the answer can never be copied out of a search result — it has to be reached by real research, then checked mechanically against a key the validator re-derives from the corpus. The exam comes in three tiers, one shape.

{{ s.name }}
{{ s.desc }}
SCENE 03 · THE DUEL
Challenger and king answer the same exam.

A challenger is tested head-to-head against the reigning king: both answer the identical ~1,000 tasks, and only the per-task difference is scored, so exam difficulty cancels out. 800 are drawn from a pool sealed and committed by digest in advance, selected by a block hash that did not exist until the challenger was frozen — and a question once published is retired, never asked again; 200 more come from each validator's private holdout of freshly published papers.

CHALLENGER vs KING · SAME TASKS
king challenger tasks in the exam →
SCENE 04 · THE VERDICT
Winning more isn't enough. It has to be proven.

A lucky exam can't crown anyone. Acceptance turns on a statistical test, not a raw score — and the crown only changes hands when independent validators agree. No company decides; coronation is a pure function of what is on the chain.

STATISTICAL BAR
99.9%
lower-confidence bound on the improvement must clear the adaptive floor δ = 0.05 × (1 − king accuracy)
STAKE QUORUM
≥51%
of validator stake must independently accept before the crown changes hands — at the same block, for every observer

THE COMPOUNDING ASSET

Every duel prints reward data that cannot be scraped.

A duel does not only produce a verdict. Every task inside it emits a verified record — the task, the model's answer, and whether that answer was right — over documents published after the models were trained. That is verified reward data: the scarcest input in post-training, and the competition generates it as a side effect of doing its job.

It cannot be scraped, because it does not exist anywhere else. There is no public dump of graded research episodes over documents that postdate the model reading them. A competitor who wanted this corpus would have to run the machine for a year to accumulate it — and by then the machine has another year of it.

And the exam itself compounds the same way: when a round closes, its private questions are published. Week after week that deposits a public deep-research benchmark that did not exist before — fresh, verified, never reused — with Epago's name on every release.

RECORD EMITTED PER TASK
taskone study names a pollution index; another names a photodetector type — which single study in the corpus involves both?
answeran exact title — described nowhere in the question, reachable only by reading both anchors
verdictcorrect · recomputed mechanically · replayable from chain state
fresh documents tasks minted duels graded better model crown worth more verified reward data accumulates on every turn and exists nowhere else
the loop is the product: fresh documents mint tasks, tasks become graded duels, duels crown better models, a better crown draws more challengers — and each turn deposits records in the middle
>$1B/yr
the scale at which frontier labs are now pricing reinforcement-learning environments — the market this data feeds
#1
bottleneck in scaling RL, per Epoch AI's survey of RL environments: verification that resists reward hacking — the exact property the duel format is built to guarantee
every duel
adds more of it — the asset grows with participation, and a latecomer cannot buy back the year they missed

HOW IT RESEARCHES

It doesn't recall an answer. It goes and finds one.

These four steps are the whole procedure — and not one of them requires the answer to be sitting in the weights. That is the reason a model this small can do this work.

01 · SEARCH
Forms its own queries
Breaks a hard question into the smaller questions an analyst would actually ask, and searches for each.
02 · READ
Opens the sources
Reads the papers in full — abstracts and body text, not snippets — and extracts the specific value or study that answers the question.
03 · CROSS-CHECK
Reconciles disagreement
When sources conflict, it follows up until it can explain the gap — different populations, different definitions, different years.
04 · CITE
Answers with receipts
Every answer arrives with its sources attached. You never have to take its word for anything.

THE MODEL, IN NUMBERS

3.3B
parameters active per token — routed out of a 30.5B mixture-of-experts, not a dense 3B model
1 GPU
is all it takes — 4-bit on a single 24 GB card, on your hardware, under your control
5 of 7
agentic deep-research benchmarks where the base model beats OpenAI o3 and DeepSeek-V3.1 (671B)
~1,000
tasks per duel — 800 public and replayable, 200 from a private holdout
99.9%
statistical confidence required before any update replaces the model
OPEN WEIGHTS
The model is yours to download, inspect and run.
No API gate, no usage terms that change under you. What we benchmark is exactly what you get — which is also the only version a regulated buyer is allowed to cite.
HONEST NUMBERS
We publish the losses too.
o3 still leads both BrowseComp columns; our own exam holds the base model to about 22%. Selective reporting is exactly the gaming this project exists to make impossible, so we do not practice it on ourselves.
NOTHING ON TRUST
Every result is publicly checkable.
Evaluations, answers, citations — all of it traceable and replayable by anyone. We designed the project so you never have to believe us.

WHY IT KEEPS GETTING BETTER

A frontier that structurally cannot sit still.

The open deep-research frontier is already small and already strong. Its problem is that it stops moving the day a lab ships the checkpoint and turns to the next thing — and that when a successor does arrive, nothing tells you whether it is genuinely better or merely tuned to the benchmark it is announced with.

Epago replaces that with a competition. Anyone in the world can train a better version; independent validators test each challenger against the reigning model on fresh questions chosen after the challenger's weights were frozen. Only a statistically proven improvement — 99.9% confidence — takes the crown and the rewards.

No company decides. Nothing can be faked. Every result is publicly replayable. That is the moat: while a closed lab ships when it chooses to, this model improves whenever anyone on Earth finds a way to make it better — and every round of it deposits reward data nobody else has.

This is why it runs on a chain rather than on a company's leaderboard. Every anti-gaming property collapses the moment the evaluator and the beneficiary are the same party — which is precisely how one provider came to test 27 model variants privately and publish only the best. Emissions buy measured capability, not claims: no participant, the Epago Foundation included, can quietly become the judge.

real data: champion accuracy {{ genesisPct }} → {{ kingPct }} · {{ coronationsTxt }} times a challenger has taken the crown
The floor is public. The crown is open. Anyone can take it.
Try the demo Watch it improve live