Skip to main content

​
Docker

Three services, orchestrated with Docker Compose:
All three have health checks and a defined restart policy, and share persistent evidence storage (the corpus directory holding captured feed data, the ledger, anchor evidence, and policy snapshots).

​
Environment variables

TISSUE_MODE=live is required. There is no TISSUE_MODE=synthetic — the daemon fails loudly at startup rather than silently substituting fabricated data.

​
Monitoring & observability

  • Metrics — GET /metrics in Prometheus exposition format: stream failures, source-proof failures/pending/verified, SSE client count, and two real latency histograms (tissue_proof_verification_latency_ms, tissue_decision_loop_latency_ms)
  • Health — GET /health (liveness) and GET /ready (readiness — requires at least one verified proof and zero recent failures)
  • Logs — structured JSON events (tissue.source_proof_failed, tissue.proof_circuit_halt, tissue.checkpoint_anchor_failed, etc.), bounded and rotated in the container runtime

​
Security & safety practices

  • Source proofs are fail-closed: a message that can’t be verified is never admitted, never quoted around
  • Drawdown kill and the proof-failure-rate circuit breaker are operator-restart-only — neither latch auto-resumes
  • Persisted local receipts are never treated as source authority; recovery always reverifies against the live TxLINE proof endpoints
  • No transaction-signing tool exists anywhere in the analyst layer
  • Manual private release artifacts include BuildKit provenance, SBOM attestations, and SHA-256 checksums; the release workflow is read-only and cannot itself publish or deploy

​
Testing & verification

pnpm run ci (dependency audit, lint, typecheck, the full default test suite, build, compiled-runtime verification, and a determinism-checked replay) runs on every build and covers 322 tests across the daemon, dashboard, and analyst — including adversarial and malformed-input suites that run by default, not as an opt-in extra:
  • Malformed/adversarial TxLINE input (apps/daemon/src/ingest/ingest.test.ts, apps/daemon/src/exec/exec.test.ts) — deliberately corrupted feed payloads (non-numeric timestamps, negative/huge scores, NaN-producing odds prices, tampered proof responses) asserting the ingest pipeline fails closed rather than silently accepting garbage.
  • Adversarial analyst input (apps/analyst/src/adversarial.test.ts) — prompt injection, spoofed tool names, malformed MCP arguments, and “terrible user prompts” driven through the real MCP client/server pair and the real agent loop, asserting no path can ever reach a write/execute action (none exists) regardless of what a compromised model attempts.
Three further suites are opt-in — they need external infrastructure this project doesn’t assume is always running, so they’re excluded from default CI but real and runnable:
1

Local Solana anchoring tests (Surfpool)

Real transaction-level tests — successful confirmation, insufficient-balance failure, unreachable-RPC failure, concurrent same-keypair submissions, and independent on-chain verification of a confirmed signature — against a real local Solana validator instead of racing public devnet’s rate limits.
2

Dashboard E2E (Playwright)

Real Chromium against a real Next.js server, driving every page under every desk status (starting, verifying, quoting, watching, halted, error) via a fake daemon HTTP process that speaks the exact API shape the real daemon does.
3

Real process-level chaos drills

Spawn the compiled daemon as a real OS process against real (proxied) TxLINE/Solana endpoints. drill:restart SIGKILLs it mid-stream and asserts the persisted hash chain survives a hard crash. drill:streamdrop severs the SSE connections without killing the process and asserts the daemon detects the drop and reconnects — a distinct fault class, requiring only real credentials, not a message that manages to pass TxLINE’s live proof verification.
Real, honest limitation: replaying an old captured corpus against TxLINE’s real proof endpoints during a restart drill doesn’t reliably produce successful admissions — TxLINE’s proof service isn’t guaranteed to serve proofs for historical/replayed messages. This blocks a live SIGKILL-during-anchor-submission scenario specifically, which is why drill:streamdrop exists as an independently valid fault class that doesn’t depend on any message passing proof verification. See Feedback & Roadmap for the full finding.

​
Scaling notes

Tissue’s default scope is one-fixture-focus, matching the PRD’s design intent, but the World Cup schedule stages multiple knockout ties concurrently. Portfolio-level exposure and drawdown caps (see Core Concepts) exist specifically so a loss on one concurrently running fixture halts every fixture the desk is running, not only the one that tripped it.
For judge evaluation, pnpm run ci alone (see Testing & verification above) is the fastest way to confirm the core claims — the opt-in suites are there for deeper verification, not required for a first pass.