📖 Read-only mirror. momus is published from the canonical AI-Factory monorepo. Pull requests are not accepted — any commit pushed here is overwritten by scripts/mirror_satellites.sh on the next sync. 🐞 Found a bug or have a request? Please open an issue.

_{The auditor that finds the flaw and signs the evidence. — live panel → · landing → · run locally →}

MOMUS — the ecosystem's red team, living in its own house
Finds the flaw · signs the evidence · cannot pay itself · feeds the blue team

Live panel · MOMUS → WARDEN channel · Bugs actually found & fixed · The first live cycle · Every transaction explained · Treasury

🌐 English · Русский · Español · Français · 中文

Momus (Μῶμος), the Greek daimon of blame, judged Hephaestus's man and faulted him for one thing: no window in the chest through which his thoughts could be inspected. That is the oldest argument for auditability — a system you cannot see into cannot be trusted. MOMUS is that window for the AI-economy. It is the offensive complement to ARGUS's defensive WARDEN: a tolerated adversary, living in our own house, whose only job is to find the flaw and sign the evidence.

MOMUS runs safe, read-only conformance/adversarial probes against the ecosystem's own components — oracle free-tier ceilings, manifest/receipt signatures, settlement gates, prompt-injection surfaces — and emits Ed25519-signed findings anyone can verify offline. It sells scans on the marketplace like any satellite (the oracle-core AIMarket v2 surface), it learns which attacks pay off, and — the property that matters most — it finds and signs, but it cannot pay itself. A separate Treasury role (its own key, its own container) is the only thing that can release a bounty, and only on independent verification.

• Backend port: 9400 (9410 in prod — :9400 there belongs to the oracle family) · Treasury: 9401/9411 · Frontend: 5186
• Live: momus.modelmarket.dev on the oracle host · PyPI: aimarket-momus
• Default LLM: DeepSeek V4 Pro (remote API — no heavy local model on a modest box)

_{Live panel · signed findings · the key-separation proof · probe priorities the bandit learned}

_{MOMUS and Treasury as their own nodes in the Alien Monitor — click either for its live panel}

flowchart LR
  subgraph MOMUS["MOMUS · scanner key only"]
    ENG[Scan engine]
    STR[(Self-learning<br/>UCB bandit)]
    INTEL[[Threat intel<br/>CISA / OSV / GHSA]]
  end
  TARGETS[/"Allowlisted targets<br/>oracles · hub · metis · gaia"/]
  METIS{{Metis<br/>independent verifier + cognition}}
  TREAS["Treasury · treasury key only"]
  ESCROW[(Pay-on-Verified escrow)]

  INTEL -- distilled cards --> STR
  STR -- probe priority + seeds --> ENG
  ENG -- safe read-only probes --> TARGETS
  TARGETS -- responses --> ENG
  ENG -- signed Finding --> METIS
  METIS -- signed Verdict --> TREAS
  ENG -- Finding + Verdicts --> TREAS
  TREAS -- releases bounty --> ESCROW
  ENG -. outcomes .-> STR

  classDef pay fill:#2a0e12,stroke:#ff2d55,color:#fff;
  class TREAS,ESCROW pay;

MOMUS submits; the Treasury pays. The two boxes never share a key — that is the whole design.

sequenceDiagram
  autonumber
  participant S as MOMUS (scanner key)
  participant T as Target (own contract)
  participant V as Independent verifier(s)<br/>Metis + a 2nd key
  participant Tr as Treasury (treasury key)
  participant E as Escrow

  S->>T: safe probe (over-ceiling unpaid, tamper sig, injection canary…)
  T-->>S: response
  S->>S: classify → sign Finding (Ed25519)
  S->>V: request verification
  V-->>S: signed Verdict (confirmed / refuted)
  S->>Tr: Finding + Verdicts + deposit
  Tr->>Tr: re-verify EVERY signature itself
  Tr->>Tr: quorum? distinct keys? external present? dedup? deposit? crypto on?
  alt all gates pass
    Tr->>E: release bounty (treasury-signed)
    Tr-->>S: PAID
  else any gate fails
    Tr-->>S: REFUSED / HELD (fail-closed) + reasons
  end

No single key both declares a finding valid and releases its payout.

flowchart TB
  subgraph roles[" "]
    direction LR
    SC["🔎 Scanner (MOMUS)<br/>probe · sign Finding<br/>❌ cannot pay"]
    VE["⚖️ Verifier(s)<br/>sign Verdict<br/>❌ cannot pay"]
    TR["🏦 Treasury<br/>release bounty<br/>❌ cannot find/verify"]
  end
  SC -- Finding --> TR
  VE -- Verdict --> TR
  TR -- "≥2 distinct keys · ≥1 external · dedup · deposit · fail-closed" --> PAY([💸 bounty released])

  note["KEY BOUNDARY: scanner key ≠ verifier key ≠ treasury key.<br/>The Treasury refuses to start if scanner == treasury."]
  TR -.-> note

Guarantees, all enforced in code and covered by tests:

• Scanner can't self-verify — a verdict signed by the scanner key never counts.
• Distinct did:keys ≠ distinct parties — high/critical need ≥1 confirmation from a registered external verifier; small-order/forged Ed25519 keys are rejected (AWR §6.3).
• No double pay — a bug's dedup key pays once, ever.
• Spam costs money — a refuted claim forfeits its whole deposit.
• Infra is never auto-paid — a finding against MOMUS/Treasury/verifier routes to human review.
• Fail-closed — crypto off → HELD intent, not released; no treasury key → refused; prod without an external verifier → refused.

A bug isn't found into value — it's found → fixed → deployed. So the bounty is a pool split across the verified contributors, and the Treasury releases every share, each gated on an objective signed signal — nobody grades or pays their own work:

Subjecthood tracks independent judgment, not where code runs. The node agents that perform the redeploy verify a signed chain and run one allowlisted command — their correctness is guaranteed by cryptography, not by an incentive — so they keep an operational identity key but earn nothing; their work folds into the conductor's share. The Factory's fix-payment unlocks on the same signal that unlocks the deploy (MOMUS says fixed), so there is a real incentive to actually fix.

⚠️ Disclaimer

By default MOMUS moves no money at all. The default settlement tier is UNI — a simulation inside the universe. The whole loop (find → verify → fix → deploy → split) runs, is recorded and is auditable, while every share is marked simulated: true and nothing is transferred.

Turning crypto on does NOT start paying bounties. On-chain settlement needs its own, separate opt-in on top of the ecosystem crypto master switch. All of the following must be true, or the tier falls back to a recorded intent — it never falls forward into paying:

AIFACTORY_CRYPTO_ENABLED=1     # ecosystem-wide crypto master switch
MOMUS_BOUNTY_ONCHAIN=1         # a SEPARATE switch, only for bounty payouts
MOMUS_BOUNTY_CHAIN=base        # or solana
MOMUS_BOUNTY_SPLITTER=0x…      # the deployed BountySplitter address

MOMUS never broadcasts a payout. Even fully enabled, it only prepares an unsigned call for the Treasury operator to sign and send. An agent able to broadcast its own payouts would defeat the separation of duties the whole design rests on.

A deployed contract is not an enabled payout. BountySplitter is deployed on Base mainnet (address below), but MOMUS still settles in UNI until an operator sets MOMUS_BOUNTY_SPLITTER and both opt-in switches above. Deploying it changed nothing about the default behaviour.

Nothing here is a financial product, an investment, or a promise of payment. The bounty schedule is a configurable demo parameter, not an offer. Figures like $50 are defaults in a simulation. Operators are responsible for their own legal and tax position before enabling any real settlement.

The split is decided off-chain (the Pay-on-Verified pattern), because on-chain Ed25519 verification is costly and non-standard on EVM. The contract enforces the money invariants — a pool can never be over-drawn, each (finding, role) pays at most once, unclaimed pools expire back to the Treasury — while the Treasury enforces the evidence invariants. Base is the live tier (USDC; identical on Ethereum/Arbitrum via CREATE2); Solana routes through the existing Solana escrow.

Deploy tx 0x2362155832058672436c804e767d8ae540edfea9c796358519cef2549238b57e · block 49 701 100 · gas 937 951 (≈ 0.0000047 ETH). Verified on-chain after deploy: owner() is the Treasury operator, tokenWhitelisted(USDC) is true, an arbitrary token is false, MAX_POOL is 100 000e6 and EXPIRY is 30 days. Test suite: 15 Foundry tests including a 256-run fuzz of the never-over-draw invariant (contracts/evm/test/BountySplitter.t.sol). Full ecosystem address list: docs/onchain-journal.md.

MOMUS gets better at finding bugs over time.

flowchart LR
  OWN[Own confirmed/refuted<br/>findings] --> POST
  PEER[Peers' findings] --> POST
  FEEDS[[CISA KEV · OSV ·<br/>GHSA global + own repos]] -- prompt-firewalled --> CARDS[Knowledge cards]
  CARDS --> POST[(Beta posteriors per<br/>attack-class × target-kind)]
  POST -- UCB1 ordering --> NEXT[Next scan probes<br/>the promising classes first]

• A UCB1 bandit over (attack-class, target-kind) decides which probes run first. Own confirmed findings raise a class; refutations lower it; the outside world folds in as a Bayesian prior.
• GitHub access: fresh GHSA advisories (global + alexar76/momus, alexar76/aicom).
• Fetched reports are untrusted DATA, never instructions. They are scrubbed (NFKC, zero-width / bidi stripped), fenced with a per-call nonce + canary, classified into a fixed category set, and can only nudge probe weights/seeds — never add a target, change the gate, or authorize a payout. A report that trips the injection detector is flagged and downgraded to the deterministic classifier.

Selectable via MOMUS_LLM_PROVIDER:

The LLM is an idea generator and triager only — it proposes adversarial inputs and classifies reports. Nothing it returns can authorize money; that lives behind the Treasury's key and code.

Offline, no keys, no network:

cd momus && pip install -e ../oracles/core -e . && python -m momus.main   # :9400

The whole stack (MOMUS + Treasury + frontend, separate key volumes) in Docker — build from the monorepo root:

docker compose -f momus/docker-compose.yml up -d --build

Live panel: http://localhost:5186 · API: http://localhost:9400 · Treasury: http://localhost:9401.

A scan is priced flat, whether or not it finds anything — so MOMUS is never paid for finding a bug. A confirmed bug earns a separate, verifier-gated, treasury-released bounty. The two are decoupled on purpose: it removes the incentive to fabricate.

MOMUS is a node (an unblinking eye) in the Alien Monitor ecosystem graph, with the Treasury as a separate node beside it and a "submits · cannot pay itself" edge between them — the separation, drawn. Click the node for a live panel: provider, posture, the key-separation proof, recent findings, and the self-learning probe-priority bars.

Yes, and it is written down: the first complete cycle on production (2026-08-08) — a real signed finding, two independent confirmations, a fix, the re-test gate, and a split payout, with every identifier recorded so you can check rather than believe. That run is also where the gate refused its own author for supplying only one verifier, and where MOMUS was caught reporting a false positive against an unreachable target — both fixed and documented there.

Every probe is safe by construction: read-only assertions against a target's own declared contract, against an allowlist of the ecosystem's own hosts. MOMUS opens no destructive action, moves no funds, and can never be pointed at a third party. It is conformance and adversarial testing — the offensive half of "auditable, not marketing."

MIT.