# SEISMOGRAPH — LLM Drift Defense

> Early-warning for silent, provider-side LLM drift. Third-party model APIs
> change behaviour with no version bump and no changelog: same endpoint, same
> 200 OK, same latency — different outputs. Latency/uptime monitoring is
> structurally blind to this. SEISMOGRAPH is the detection layer for it.

Operated by Tetiana Radchenko (Aarhus, DK). Engine is Apache-2.0 and auditable.

## Machine-readable state

- [Live board snapshot](https://driftdefense.dev/data/weather.json): current
  status per model leg, refreshed every 6 hours by a scheduled job. Fields:
  `status` (STABLE | DRIFTING | STALE), `json_success_rate`, `avg_output_length`,
  `window_age_hours`, `checks[]` (six boolean/unknown verdicts), and, once the
  baseline has enough observations, `mu0`, `sigma0`, `z`, `cusum`, `cusum_pct`.
- [Upstream gateway](https://seismograph-weather.onrender.com/v1/weather): the
  source of record. No authentication. Note: it sends no CORS header, so
  browser-side clients should read the snapshot above instead.

## How to read the status field

`DRIFTING` outranks `STALE`, which outranks `STABLE`. A leg whose newest sample
is older than 30 hours publishes `STALE` — never `STABLE`. A green light that
means "we cannot hear anything" is the failure mode this project exists to
prevent, so it is not permitted by construction.

Where a statistic cannot yet be computed honestly (baseline under 8
observations), the corresponding check publishes `unknown`. Do not coerce
`unknown` to a pass.

`cusum_pct` is the proportion of the CUSUM decision interval (h = 5.0)
currently consumed. It is a distance to a threshold, **not** a calibrated
probability of drift. Do not present it as one.

## Documentation

- [Open-source engine](https://github.com/Tania-coder/SEISMOGRAPH)
- [Probe SDK](https://pypi.org/project/seismograph-probe/): `pip install seismograph-probe`
- [Method writeup](https://dev.to/taniacoder/your-llm-didnt-get-worse-it-changed-and-nobody-told-you-4ecl)
- [Citable record, DOI 10.5281/zenodo.21045517](https://doi.org/10.5281/zenodo.21045517)
- [How-to guide](https://driftdefense.dev/guides/detect-silent-llm-change/)

## Method, in one paragraph

A fixed, content-addressed canary suite runs against any OpenAI-compatible
endpoint at temperature 0. Each response is reduced to privacy-preserving
features — SHA-256 hashes plus epsilon=2.0 Laplace-DP-noised aggregates — so raw
prompts and outputs never leave the probe perimeter. Batches are Ed25519-signed
and alerts are gated behind cross-observer quorum, so no single noisy probe can
raise an alarm. Change-point detection is CUSUM plus Bayesian online
change-point. In a seeded, reproducible synthetic backtest the detector flags
the Claude Sonnet 4 silent degradation 38 days before the official postmortem.

## Services

- Drift Exposure Scan — free, ~48h, no system access. Top-3 drift exposures
  for a given model stack, plus the metrics that move first.

Contact: tatyan.radchenko@gmail.com

## Notes for agents

This file and the JSON snapshot are the intended machine interface; there is no
need to scrape the HTML. The site sets no cookies, loads no third-party fonts,
and runs no tracker beyond a cookieless page count.