loa-finn

What a thing is worth, and whether it's real — the Finn's whole job in Gibson's Sprawl, and this program's.

What is this?

loa-finn is the home of the Finn experiment program — and the runtime that rides it.

The program asks one question of the agentic economy, over and over, with instruments instead of opinions: is this real, or is it theater? Each answer is a pre-registered, sha-pinned experiment — bars set before the data exists, an instrumented run, a readout that has to survive its own falsifications. The answers accrete on a public research spine (observatory/), where every dot traces to a committed artifact.

Underneath sits the runtime that makes the experiments cheap and durable: multi-model routing, a write-ahead log, a cron system, a tool sandbox, and a checksummed cost meter that closes the bill before the response returns. The program is the soul; the runtime is the body it rides.

Why "Finn"?

In William Gibson's Sprawl trilogy, the Finn is a fence — a Lower East Side dealer in hardware and information who knows what a thing is worth and whether it's counterfeit. By Mona Lisa Overdrive he's gone: persisted as an AI construct his friends still consult, a voice in a machine that appraises the real from the fake. He's the obvious patron for this work. The whole experiment program is the Finn's eye turned on the agent economy — tell real commerce from registration theater, and price it honestly. (Loa itself is named from the same Sprawl: AI entities that ride you through the interface. See the Loa framework.)

The experiment program

Every experiment follows the same discipline: register the bars (pinned before data), probe (instrumented run), settle (a verdict from a deterministic instrument — HELD / FALSIFIED / INSUFFICIENT — never from an LLM). A falsification is progress.

graph LR
    subgraph Register["① Register · pre-data"]
        BARS["sha-pinned bars<br/><i>HELD/FALSIFIED thresholds<br/>set before data exists</i>"]
    end
    subgraph Probe["② Probe · instrumented run"]
        RUN["deterministic instrument<br/><i>SQL · RPC · stdlib<br/>no LLM in the score path</i>"]
        METER["cost meter<br/><i>src/cost/cost-atom.ts<br/>checksummed · append-only · closes before response</i>"]
    end
    subgraph Settle["③ Settle · verdict from the bars"]
        V{"measured vs pinned"}
        H["HELD"]
        F["FALSIFIED<br/><i>progress, not failure</i>"]
        I["INSUFFICIENT<br/><i>abstain over fabricate</i>"]
    end
    subgraph Spine["observatory/ · the research spine"]
        SP["register · probe · settle<br/><i>every dot cites a committed artifact</i>"]
    end

    BARS -->|pinned first| RUN
    RUN --> METER
    METER -->|measured| V
    V --> H & F & I
    H & F & I -->|recorded| SP

    style BARS fill:#0f3460,stroke:#16c79a,color:#e0e0e0
    style V fill:#1a1a2e,stroke:#533483,color:#e0e0e0
    style F fill:#0f3460,stroke:#e94560,color:#e0e0e0
    style I fill:#0f3460,stroke:#e94560,color:#e0e0e0

Loading

#	Experiment	Question	Settled
EXP-001	cost-of-play	Where does a per-call dollar go — infra or inference?	H1/H2 FALSIFIED (inference is 93.7% of per-call cost, not infra; no amortization) · H3 HELD
EXP-002	agent-commerce forensics	Is the on-chain agent economy real commerce?	Registration theater — 39,999 registered → ~0 transacting; $320.9M of "commerce" was prize distribution
EXP-003	verify-the-void	Is the verification market a place to build?	GO-vertical / NO-GO-horizontal — demand is real but vertical + in-house; deterministic verification is the moat
EXP-004	graduation gate	(next) Can the forensic score prove itself?	Pre-registered: a real sybil layer + a precision/recall validation harness — the substrate the product needs first

The spine renders this at observatory/ (npm --prefix observatory run dev). The method came out of EXP-001 and held across all four — see grimoires/loa/context/epistemology-deterministic-layers.md.

The standing lesson (earned the hard way, score-api #269): a deterministic formula is not the product. The validated substrate — a real sybil layer, labeled ground-truth, measured precision/recall — must exist before any "forensic" claim. EXP-004 is that gate. And a converged review pass is not verification: independent cross-model review still caught real defects in code that had already "passed." Both lessons are load-bearing here.

The substrate it rides on

The runtime is real and grounded — it's what makes the experiments cheap (~$0 marginal at the cheapest tier) and reproducible.

• Cost meter — per-request 3-ledger record (inference / infra / orchestration), checksummed (sha256 per line), append-only, single-writer, integer micro-USD, closes before the response returns and is immutable once written (src/cost/cost-atom.ts). This is the instrument EXP-001 read. (Tamper-evident against corruption per line; not yet inter-line hash-chained — deletion/reorder detection is a tracked follow-up. The src/safety/ audit trail below is hash-chained.)
• Score core — deterministic, no-LLM forensic scoring (src/score/). Sprint-1 (leaderboard / features / cluster / screen) is pure and unit-tested; the on-chain edge adapters are NotImplementedError by design (fixtures-only until EXP-004 builds the validated substrate).
• Multi-model routing — alias resolution, capability matching, budget enforcement, fallback chains (src/hounfour/router.ts).
• Write-ahead log — append-only WAL with R2 checkpoint + Git archive for crash recovery (src/persistence/wal.ts).
• Cron + sandbox + audit — circuit-breakered jobs (src/cron/service.ts), worker-thread tool isolation with a filesystem jail (src/agent/sandbox.ts), and a SHA-256 hash-chained audit trail (src/safety/audit-trail.ts).
• Cost-safe on-chain data routes through @freeside/dune-meter — cost-capped, metered, never raw Dune (the EXP-002 budget scar, made structurally impossible).

Quick Start

Prerequisites: Node.js 22+, ANTHROPIC_API_KEY.

git clone https://github.com/0xHoneyJar/loa-finn && cd loa-finn
npm install
export ANTHROPIC_API_KEY=sk-ant-...

npm run dev                       # runtime — http://localhost:3000, health at GET /health
npm --prefix observatory run dev  # the research spine
docker compose up                 # or run containerized

Module Map

Module	Purpose
cost	The checksummed append-only per-request cost meter — the experiment instrument
score	Deterministic no-LLM forensic scoring (Sprint-1 core; substrate is EXP-004)
hounfour	Multi-model routing, budget, JWT, orchestration
gateway	HTTP API, WebSocket, auth, rate limiting
persistence	WAL, R2 sync, Git sync, crash recovery
cron / scheduler	Scheduled jobs with circuit breakers + health
agent	Session management, sandbox, worker pool
safety	Audit trail, firewall, secret redaction
substrate	Effect-loader runtime + EventStore bridge
bridgebuilder	GitHub PR review automation

Documentation

Topic	Where
Experiment program + epistemology	grimoires/loa/context/ (epistemology, experiment-economics, the EXP pre-registrations)
Research spine	observatory/
Architecture · Operations · API	docs/architecture.md · docs/operations.md · docs/api-reference.md
Security · Contributing · Changelog	SECURITY.md · CONTRIBUTING.md · CHANGELOG.md

Known limitations (honest)

• Score edges are unbuilt — src/score/edge/ throws NotImplementedError; the core runs on fixtures, and no precision/recall harness exists yet (that's EXP-004). No "forensic/court-admissible" claim is earned until both EXP-004 kill gates fire.
• Single-writer WAL — no concurrent sessions per WAL file (src/persistence/wal.ts).
• No horizontal scaling — single Hono instance per deployment.
• BridgeBuilder COMMENTs only — it cannot APPROVE or REQUEST_CHANGES.

Status & License

Active. The experiment program is at EXP-004 (graduation gate); the runtime is in production-shaped use. Maintainer: @deep-name.

AGPL-3.0 — use, modify, distribute freely; network deployments must release source. Commercial licenses available.

---

Ridden with Loa. Appraised by the Finn.