Agentic Financial Intelligence (AFI) is an open protocol for turning trading signals — produced by human or agentic analysts — into scored, auditable, replayable evidence. Every signal is normalized to a canonical form, scored by version-pinned rules, and persisted with full provenance, designed so that any independent party can verify exactly how a score was produced.
This repository is the navigation map for the AFI organization: what defines the protocol, what implements it, what may be replaced, what exists today, and where to begin. It is deliberately thin and carries no authority of its own.
Protocol authority lives in exactly three repositories:
| Repository | What it holds | |
|---|---|---|
| 1 | afi-governance | Accepted protocol decisions — protocol law. Start at decisions/. |
| 2 | afi-config | Canonical schemas, registries, conventions, and known-answer tests (kats/). |
| 3 | afi-math | Canonical deterministic math kernels and their golden vectors. |
Authority is exactly what the accepted decisions expressly delegate — self-labeling confers nothing. If anything in this repository disagrees with a governed artifact in those repositories, the governed artifact wins. This map is not the schema home, not the math home, not a runtime, and not a substitute for afi-docs.
AFI separates the protocol from its operators.
- Rules vs operation. AFI Protocol defines the interoperable rules — the governed contracts, schemas, math, identities, and lifecycle. Operating a service confers no authority over them; self-labeling confers nothing.
- Official reference services. AFI Research Institute is designated to operate AFI's official open reference services — a hosted Gateway reference service for structured ingress and an oracle-ingress / CPJ-normalization reference service for message- and source-derived signals (INST-GOV; reference-services spec).
- Non-exclusive. The designation is non-exclusive. Any independent party may run the open-source Gateway, a conforming re-implementation, its own collectors, or its own AFI-compatible infrastructure — conformance is defined by the afi-config contracts and KATs, not by who operates.
- No live deployment claimed. These reference services are not deployed; the
implemented pipeline is CI-proven and currently reaches
SCORED(see Implemented today).
signalIdis the canonical join key. Every artifact about a signal — canonical form, score, evidence record — joins on it.- Strategy identity is the triple
analystId + strategyId + strategyVersion(object-identity-v0.1). - Canonical inbound signal form is the Universal Signal Schema (USS) v1.1
(
schemas/usignal/). - Scoring runs the governed UWR (Universal Weighting Rule) engine. The
protocol-recognized profile is version-pinned (testnet-provisional) in
registries/uwr-profiles/. Analysts compose the five analysis categories (technical,pattern,sentiment,news,aiMl) into registered pipelines via the delegated afi-config registries (factory-configurable-pipelines-v1). Analysts may configure and run conforming UWR profiles, decay/Greeks surfaces, and strategy pipelines without bespoke governance permission; a profile becomes protocol-recognized only when registered and version-pinned in afi-config — a governed registry change. - Persistence — the current canonical chain is Gateway → Reactor → afi-infra → MongoDB. Protocol law fixes the single write path: afi-infra owns the sole canonical persistence interface. The Gateway authenticates and routes to the Reactor; it never constructs or writes canonical evidence (guardrail-enforced).
- Lifecycle is the governed state machine
INGESTED → VALIDATED → SCORED → CERTIFIED → QUALIFIED → CHALLENGE_OPEN → [CONTESTED →] FINALIZED → EPOCH_ELIGIBLE. The implemented lifecycle currently reachesSCORED— see Implemented today.
The organization is exactly 18 repositories.
| Repository | Role |
|---|---|
| afi-governance | Accepted protocol decisions (decisions/) |
| afi-config | Canonical schemas, registries, conventions, KATs |
| afi-math | Canonical deterministic math kernels + golden vectors |
| Repository | Role |
|---|---|
| afi-core | Scoring library: governed UWR engine, profile loader, decay surfaces |
| afi-reactor | Scoring runtime: ingest → USS validation → analyst-configurable graph execution (registered strategies compose the five analysis categories through explicitly selected provider instances; all five lanes must succeed) → deterministic join → scoring → UWR resolution → Evidence V3 construction (five bound provider invocation proofs) → submission |
| afi-infra | Canonical evidence store: scored_signal_evidence on MongoDB — afi.scored-signal-evidence.v3 only, hash-verified admission, unique signalId, append-once, transactional supersession |
| afi-gateway | External submission boundary: authenticates and routes to the Reactor; never writes evidence |
| afi-mint | Mint/reward execution home (delegated; not live — see status) |
| afi-token | On-chain token enforcement (86B hard cap; testnet-proven) |
| Repository | Role |
|---|---|
| afi-docs | Documentation hub |
| afi-factory | Pipeline authoring system: template authoring/instantiation, manifest validation, canonical hashing, and an agent-operable capability layer (SDK, CLI, machine-readable catalog, MCP adapter) for authoring pipeline artifacts |
| afi-xerc20 | Vendored xERC20 standard (dependency of afi-token) |
| afi-tiny-brains (private) | Bounded first-party aiMl orchestration service (profile-selected experts, deterministic resolver; scoring-inert) |
| Repository | Role |
|---|---|
| afi-econ | Economic research kit (non-canonical unless promoted by governance) |
| afi-benchkit | Evaluation and benchmark harness |
| afi-artifacts | Frozen, DOI-minted paper reproducibility bundle |
| Repository | Role |
|---|---|
| afi-protocol | This map |
| .github | Organization profile and defaults |
A conforming AFI implementation must:
- Honor the accepted decisions in
afi-governance/decisions/
— object identity (
signalId, USS v1.1, the strategy triple), lifecycle, persistence, scoring pins, math authority, and economic law. - Validate against the afi-config contracts — the
scored-signal-evidencev3 schema (canonical evidence isafi.scored-signal-evidence.v3, which carries a thin composition reference, exactly five credential-safe per-lane provider invocation proofs, and record-levelrecordHash/replayHashcommitments) and its published valid/invalid vectors, the pinned UWR profile values (testnet-provisional), and the KAT vectors. - Match the canonical math kernel outputs for any component computing emissions, decay, or time value — the afi-math golden vectors and the afi-config KAT vectors (scoring and time decay).
Everything else may be replaced: any runtime honoring the same contracts and
handoffs, any store honoring signalId-keyed evidence semantics, any
submission boundary that submits rather than writes, and every support and
research component.
The pipeline through scoring and persistence is implemented and proven by CI against real infrastructure:
ingest → USS v1.1 validation → scoring (governed UWR engine, pinned profile)
→ thin evidence construction → submission
→ canonical store (unique signalId, append-once, transactional supersession)
The implemented lifecycle currently reaches SCORED.
Deliberately not implemented, or not yet governed:
- Post-
SCOREDtransitions (CERTIFIED,QUALIFIED, challenge,FINALIZED) — the single finality writer is defined in law but intentionally unimplemented pending new authorization. - Epoch accounting and rewards — no implemented owner.
- Live mint orchestration — governance-blocked: role weights and the settlement/epoch layer are not yet governed.
- Mainnet settlement — not yet governed.
- A live read/replay/verify API surface over the Atlas or evidence — reserved to ATLAS-GOV and not built. The Atlas foundation (below) is discoverability metadata, not a runtime endpoint.
- The AFI Protocol City (the visual navigation of Atlas truth) and the AFI Participant Gateway (a registration/conformance control plane) — not started.
Exactly two Districts are formally registered (authority-districts-v0.1, Part D); both are active capability domains, and both remain non-production:
- District 1 — Signal Evaluation — the active Signal Evaluation capability domain (district-one-signal-evaluation-capability-v0.1): its current implementation is the live manifest-driven GraphExecutor pipeline hosted in afi-reactor — the sole signal-evaluation executor. The district is an enduring capability and authority domain — implementations may be replaced through accepted authority without retiring the district.
- District 2 — Evidence & Provenance — the data/provenance boundary: it
receives the scored evaluation result from District 1 and owns evidence
construction and the canonical persistence handoff. Its M1 schema family is
homed in afi-config
(
schemas/provenance/), and its M2 surface in afi-reactor is prospectively ratified. The registry keeps District 2’s M1 authorization (R1-GOV D-R1-5, in afi-governance) in force.
AFI is organized into capability Districts — enduring capability and authority domains, not repositories. A single structure may serve more than one District (the Reactor runtime serves both), and a routing structure such as the Gateway may route access without owning the underlying capability.
The District / API Atlas is AFI's canonical discoverability layer
(district-api-atlas-foundation-v0.1,
ATLAS-GOV). Its machine-readable foundation (afi.protocol-atlas.v1, in
afi-config) exposes AFI's current capabilities, interfaces, routes, contracts,
and maturity, keeping maturity separate from visibility and marking honestly
reserved future space. It describes the protocol; it defines no District,
routes no traffic, and exposes no endpoint. The full picture lives in
afi-docs AFI_Full_Architecture.md.
The future AFI Protocol City will visualize Atlas truth, and any AFI-operated participant onboarding will be a reference entrance, not the only one: third parties may build competing conformant onboarding and developer experiences against AFI's shared conformance requirements.
| You are a… | Start with |
|---|---|
| Developer | afi-reactor (runtime) and afi-core (scoring library), then the contracts in afi-config |
| Analyst / strategy author | the UWR profile registry and KATs in afi-config; identity and configurability rules in afi-governance decisions |
| Validator | the scored-signal-evidence v3 schema and its vectors in afi-config; store semantics in afi-infra |
| Operator | afi-gateway (submission boundary) and afi-infra (canonical store); persistence decisions in afi-governance |
| Researcher | afi-docs, afi-econ, afi-benchkit, and the frozen record in afi-artifacts |
Contributions are welcome by fork and pull request in any public repository. Conforming analyst pipelines need no permission — conformance is defined by the contracts, registries, and KATs above.
MIT — see LICENSE.