Seiche — funding-stress terminal

sealed record

A seiche is a standing wave in an enclosed body of water — invisible from the shore, until it sloshes over the edge. Funding stress behaves the same way.

Seiche is a free, open source (AGPL-3.0) funding-stress, positioning and divergence terminal for the dollar funding system — US money markets, the Treasury capital-market complex, the global basins connected to them through the swap lines, and the offshore-dollar crypto basin moored to the T-bill market through stablecoins. Zero data cost: built entirely on free, keyless public APIs (FRED, NY Fed Markets, OFR STFM, Treasury FiscalData, CFTC, ECB Data Portal, DeFiLlama, Coinbase Exchange).

Every 2025–26 stress event (Sep 15 2025 tax-date squeeze, Oct/Dec 2025 record SRF draws, Apr 2025 basis unwind) was front-run by plumbing signals while price screens looked calm. Incumbent tools either have the data with no opinion (Bloomberg, $32k/yr) or authority with no synthesis (OFR/NY Fed dashboards). Seiche is the opinionated fusion layer: forward-looking, alerting-ready, provenance-honest — and v2 adds the layer none of them have: honest evidence about itself.

v2 "Deep Water" — twenty-two engines, eight analytics layers, twelve tabs

v2.3 "Letters of Marque" (built in tandem across two sessions) adds the forecast layer and the layer that makes every other layer accountable: Undertow (critical slowing down — the basin's damping, measured on ordinary days), the Swell Forecast (the funding-stress forward curve — P(pop ≥ x bp) by date, six weeks out, from the public forcing calendar), The Stack (walk-forward ensemble of every event forecaster — rule, ML, analogs, Swell — plus The Tell, with a disagreement gauge), The Book (HELM tab — explicit daily positions on 2y/10y duration proxies, S&P and BTC over a T-bill base, walk-forward P&L with costs, block-bootstrap Sharpe CIs and mandatory benchmarks, verdict printed even when it loses), a hash-chained as-published track record shipped inside the static publish (nobody, including the operator, can quietly rewrite a bad month), and the Far Basin — Palimpsest's censorship-fear channel (palimpsest.info), a policy confession signal no market data vendor carries, honestly quarantined until it accrues testable history. Strategy doc: docs/STRATEGY.md.

★ = methods invented for this tool.

The desk assistant: seiche ask "why is the index elevated?" (or the Ask box on BOARD) answers strictly from a deterministic context pack of the live board — every number cited to its engine and as-of date, "not on the board" instead of improvisation. Routed through free-llm-router's free tiers, or any OpenAI-compatible endpoint via SEICHE_LLM_BASE_URL; with neither configured it returns the context pack itself.

The analytics layers on top:

• The Tell — plumbing percentile minus market-priced-stress percentile (VIX, HY/IG OAS, rates vol). Positive = the basin is sloshing and the screens haven't noticed. The whole thesis in one tradeable number.
• The Navigator — an LLM forecaster made accountable: one committed P(funding event, 5bd) per data-day, grounded strictly in the live board, written into the hash-chained record. An LLM cannot be honestly backtested (it has read the history), so its FORWARD record is its only evidence and its weight stays zero until that record earns a hearing. seiche navigator.
• The Communiqué — FOMC statements read as vintage-stamped data: frozen deterministic lexicons score policy direction, balance-sheet bias and funding-stress vocabulary per statement; the change vs the previous statement is the signal, and the Time Machine replays text as it stood.
• The TED bridge — the ML Lab pretrains on the TED spread's 1990–2018 funding-stress record (2008/2011/2016) in the same feature slots, down-weighted, and publishes the transfer gain vs the SOFR-only model either way.
• The Stack + The Book — the rule index, ML Lab, Tide Tables analogs, the Swell curve and Bathymetry's first-passage odds all emit P(funding event, 5bd); the Stack calibrates and blends them walk-forward (publishing the equal-weight mean whenever the fitted blend can't beat it), publishes member dispersion as a first-class ambiguity signal, and the Book converts the result into explicit daily paper positions with costs, benchmarks and bootstrap CIs. Every view's daily forecast and every position is appended to the hash-chained PIT record — a track record no reconstruction can polish.
• Turn Barometer — forecasts the next month/quarter-end turn's severity with leave-one-out cross-validation, always benchmarked against a naive forecast. When the model can't beat naive, it says so and publishes naive instead.
• Playbook — what S&P/VIX/OAS/yields did the last N times the board looked like this, in native units, with n printed. Decision support, not advice.
• PROOF — the backtest lab: the index rebuilt with expanding-window statistics only (no look-ahead — enforced by a unit test), recall/precision with Wilson 95% intervals, run-level precision (alert days are serially correlated; runs are the honest trials), episode-by-episode lead times including the ones it missed — and the orthogonal signal test: the same event-capture with the target's own variable family (spread/tails) removed from the signal. At build: orthogonal recall 69% [CI 42–87] vs 62% full, with the structural components alone at the 98th–100th percentile 42 days before the Sep/Dec-2025 squeezes. The claim is causal structure, not autocorrelation.
• Time Machine — replay the whole board as of any date since ~2018. Replayed to Sep 12 2019, the board reads EROSION with reserves $576B below the kink and flags Sep 16 2019 — the exact day the repo market broke — as a crunch window.

Principles: no naked numbers (every value carries source + as-of + staleness), fail-loud (a dead feed shows as DEAD and reduces published coverage — it never silently vanishes), honest lags (COT is T+3 by construction; shown, not hidden), honest scope (markets without a qualifying free feed — Japan, China, Russia, Africa — are stated as out of scope, not faked in).

Run it

# backend (Python 3.11+)
cd backend
python3 -m venv .venv && .venv/bin/pip install -e .
.venv/bin/uvicorn seiche.api:app --port 8787

# frontend (dev)
cd frontend
npm install
npm run dev          # http://localhost:5173 (proxies /api to :8787)

# or production single-process: npm run build, then uvicorn serves frontend/dist at /

First load is slow (cold fetch of several years of history); everything after is served from the SQLite cache with cadence-aware TTLs.

The operator CLI

seiche pull               # force-refresh, print the index line
seiche brief --save       # this morning's desk note (markdown, archived to data/briefs/)
seiche alert              # evaluate alert rules once (cron/launchd-friendly; exit 2 = fired)
seiche watch -i 1800      # pull + alert on a loop
seiche replay 2019-09-12  # Time Machine in the terminal
seiche backtest           # PROOF summary
seiche ml                 # ML Lab: event probability + honest validation
seiche analogs            # Tide Tables: nearest historical analogs + forward fan
seiche swell              # the funding-stress forward curve, 6 weeks out
seiche physics            # the physics board: floor, modes, determinism, tail law
seiche bathymetry         # the basin floor in detail: potential, spectrum, entropy, first passage
seiche book               # the Book: today's positions + walk-forward P&L verdict
seiche ask "…"            # desk assistant, grounded in the live board
seiche serve              # API + UI
seiche mcp                # serve the board to AI agents over MCP (stdio)

For AI agents (MCP)

Seiche is also a Model Context Protocol server — any MCP-capable agent (Claude Code, Codex, your own) can read the live board as tools. Where a data feed hands an agent raw macro numbers, Seiche hands it the conclusion: a regime read, forward event odds, historical analogs, and an honest backtest. Stdlib-only, no new dependencies.

claude mcp add seiche -- seiche-mcp          # Claude Code, local (stdio)
SEICHE_MCP_PUBLIC=1 seiche-mcp               # free surface only

Or, zero-install, over HTTP: the same tools are served at /mcp on the API (https://api.seiche.info/mcp) — the full tool surface, free for everyone, no token needed (rate-limited per caller so one client can't crowd out the rest).

Full setup, the tool catalogue, client config, and metering: docs/MCP.md.

Want the board in your pocket instead of a client config? The Hermes desk-agent kit turns hermes-agent into a Seiche desk agent on Telegram/Discord/Slack: a scheduled morning brief, regime alerts with anti-noise rules, point-in-time episode replays, and a PROOF-grounded answer to "can I trust this". Walkthrough: docs/HERMES.md.

Where this is heading on the crypto side (stablecoin reserves, tokenized Treasuries, DeFi rates all sit on the market Seiche reads): docs/CRYPTO.md.

Alerts dedupe per state in SQLite, notify via macOS notification and optional webhook (SEICHE_WEBHOOK_URL — Slack/Telegram/ntfy style {"text": ...}). A launchd template lives in ops/com.seiche.watch.plist.

Deploying on a VPS (Hetzner etc.)

# first time, as root — clones to /opt/seiche, runs the test gate, builds,
# installs systemd units (API on 127.0.0.1:8787 + a 30-min alert timer)
bash ops/deploy/install.sh

# every release after that — pull main, re-gate, rebuild, restart
bash /opt/seiche/ops/deploy/update.sh

Put a TLS reverse proxy in front. The box already serving another site (e.g. Palimpsest) just adds a vhost — nginx:

server {
    listen 443 ssl http2;
    server_name seiche.example.com;   # certbot --nginx -d seiche.example.com
    location / { proxy_pass http://127.0.0.1:8787; proxy_set_header Host $host; }
}

Auto-deploy on every merge to main: add two repo secrets — HETZNER_HOST (the server IP) and HETZNER_SSH_KEY (a private key whose public half is in root's authorized_keys) — and .github/workflows/deploy-hetzner.yml runs the test-gated update.sh on the box after each push to main. Without the secrets it skips cleanly. The SQLite cache AND the PIT/Navigator as-published record live in /opt/seiche/backend/data — back that directory up; it is the track record. LLM keys for the desk assistant/Navigator and Telegram alert credentials go in the [Service] environment of ops/deploy/seiche.service.

Tuning the editorial voice

backend/seiche/config.py quarantines every judgment call: composite weights, regime thresholds, resonance/turn/tell parameters, alert rules, the episode library, contract DV01s. The math never hides an opinion.

Non-goals

No paid data, no auth, no intraday ticks. Daily cadence + operation results is the honest granularity of the free stack. Backtests use final-vintage data (weekly H.4.1 aggregates are lightly revised; the caveat is printed on the PROOF page). From v2 onward Seiche also accrues a true as-published point-in-time record (/api/pit) that no reconstruction can be accused of polishing. Not investment advice.