No raw prompts or outputs. Ever.
Only derived, non-reversible features cross the boundary — token counts, a JSON-valid flag, SHA-256 hashes. Raw text is never stored or transmitted.
Providers ship silent model changes with no version bump and no changelog. Same 200 OK, same latency — but JSON fidelity slips and outputs drift. Standard monitoring is blind to it. I build the early-warning layer that isn't.
Open-source engine · Apache-2.0 · privacy-preserving by design · you own every result.
Not a screenshot and not a mock-up. Every figure below is fetched from the public gateway on a schedule and recomputed here. Where the data isn't yet deep enough to answer a question honestly, the answer reads unknown — never a comfortable green.
Two-person teams pitch with a deck. I pitch with a reproducible, seeded backtest you can re-run yourself.
In the backtest, detection occurred during the 0.8% misrouting window — 38 days before the official postmortem and 19 days before the escalation became visible to users. The methodology is hash-committed, so the next call is a prediction, not a postmortem.
Evidence: live dashboard · open-source engine · dev.to writeup · pip install seismograph-probe · how-to guide
Apache-2.0, open source 325 automated tests, CI-gated CodeQL: 0 open alerts PyPI releases via OIDC Trusted Publishing — no tokens Citable: DOI 10.5281/zenodo.21045517
Leibniz published binary arithmetic in 1703 and, in the same paper, mapped it onto the I Ching's broken and unbroken lines. That is the whole job here: take a continuous signal and reduce it to one honest verdict. The six-line figure on each leg above is not decoration — each line is a real check, drawn unbroken for yes, broken for no, dotted where the honest answer is not yet knowable.
The same doctrine governs the engine: status precedence is DRIFTING > STALE > STABLE, so a leg that has stopped reporting can never publish as healthy. A monitor that shows green because it cannot hear anything is the exact failure this project exists to prevent.
One email listing the LLM APIs you depend on. No call, no system access, no NDA needed.
Your top-3 silent-drift exposures, ranked, with the exact metrics that would move first. Yours to keep either way.
If the map warrants it: canary probes on your real model tuples, tuned thresholds, alerts wired into Slack or PagerDuty. Scoped together with you.
Only derived, non-reversible features cross the boundary — token counts, a JSON-valid flag, SHA-256 hashes. Raw text is never stored or transmitted.
Numeric signals carry Laplace DP noise with a tracked epsilon budget. Aggregated, anonymised, defensible.
Probes deploy inside your perimeter. Your data never depends on my infrastructure. Audit-grade evidence export for compliance.
I don't sell you a model, so I have no reason to tell you it's fine. The engine is Apache-2.0 and auditable line by line.
Security posture. Signals are signed with per-org Ed25519 keys. The codebase ships with CodeQL static analysis (0 open alerts), a 325-test CI gate, 2FA-protected accounts, and tokenless PyPI releases via OIDC Trusted Publishing. Reproduce every claim yourself — the backtest is seeded and open.
The Anthropic routing bug degraded output for weeks at 0.8% of traffic while every dashboard stayed green — a seeded backtest flags it 38 days before the postmortem. Your free scan takes one email and ~48 hours.
Silent changes happen behind stable model names — in the Anthropic Aug–Sep 2025 case, a seeded backtest flags it 38 days before the postmortem. Leave your email and you'll be notified when drift alerts open up for the models on the public board.
No form? Request an alert by mail. One email per confirmed drift event — no newsletter, no sharing.
App-level observability tells you your outputs got worse. It can't tell you whether the model itself changed — and it's often channel-conflicted (it sells to the providers it would have to indict). SEISMOGRAPH watches the provider side, neutrally, across time.
The free scan: your top-3 drift exposures within ~48h. If it's worth going further: live canary probes on your own model tuples, tuned thresholds, a written report and an alerting plan.
Raw prompts and outputs never leave your perimeter — only DP-noised aggregate features. Probes run in your VPC. You own every artifact. Full technical detail is in the open-source repo.
Because it's falsifiable. The 38-day Claude Sonnet 4 call was made on a seeded, reproducible backtest you can re-run. The methodology is hash-committed up front.
The public gateway's /v1/weather endpoint, fetched on a six-hour schedule by a scheduled job and committed next to this page, so your browser reads it instantly and no request of yours ever touches the gateway. The baseline, z-score and CUSUM statistic are recomputed from the observations accumulated since that job started — which is why some lines read unknown until enough points exist to answer honestly.
Honest answer: the public network effect (cross-org quorum) compounds with adoption and is early. Today the value is the in-VPC early-warning layer and the neutral monitoring — which stand on their own. I'd rather tell you that than oversell it.
Send me the model APIs you depend on. Within ~48 hours you get your top-3 silent-drift exposures and the metrics that move first.
or email me directly: tatyan.radchenko@gmail.com
Early-observer program: the first 10 design partners get the scan free in exchange for honest, publishable feedback.