# Pulse Desk

> Paste any time-indexed numeric series - a KPI, a sensor log, a latency or error-rate export, a price
> history - and get its exact statistical shape computed free in the browser: an OLS trend with a
> significance test, autocorrelation-based seasonality, a single best changepoint, z-score anomalies, a
> volatility-regime read, and a naive forecast. An AI pass, held to those exact numbers, explains what
> the shape means and can turn it into a decision brief for a stated purpose.

- URL: https://pulse-desk.skillsafe.ai/
- API docs: https://pulse-desk.skillsafe.ai/api.html
- Token panel: https://pulse-desk.skillsafe.ai/tokens.html
- Platform: SkillSafe (https://skillsafe.ai/)
- Model: `gpt-terra` (the balanced OpenAI tier alias)
- Price: the app itself is free; each run costs the signed-in user credits for the model call

## What it is for

One user, in one sitting, holding one pasted series they want understood: is it trending, cyclical,
stable, or breaking a regime; is anything about the pasted numbers themselves worth flagging; and - if
they want it - a short decision brief tailored to what the series actually is (an ops metric, a business
KPI, a quality-control measurement, a capacity-planning input). Pulse Desk is not a charting tool, a
database, or financial/medical/investment advice; it describes statistical shape.

## The one work object

A **time-indexed numeric series**: one point per line, `timestamp,value` (comma, tab, or whitespace
separated) or a bare value with no timestamp (a sequential index is assumed). A header row such as
`date,value` is detected and skipped automatically. Missing values (`NA`, `null`, a blank field) are
excluded from the numeric series but their timestamps are still counted. An optional `value_label` (free
text describing what the metric is) helps the AI pass; a `purpose` field (ops monitoring, business KPI,
quality control, capacity planning, or other) steers only the `brief` lane's language.

## The two lanes

Both read the same series and return the same envelope; only the `body` differs.

1. **`diagnostics`** — trend direction and strength, whether a season repeats, whether the mean shifted,
   how volatile the series is, and a data-quality note. This is the inspect-and-decide stage, and it is
   the one that usually hands off to `brief`.
2. **`brief`** — the same series turned into a short, decision-oriented brief for a stated purpose:
   ranked priority actions, what to watch next, and longer-term considerations. Reads `prior_diagnostics`
   (the `diagnostics` lane's own prior output on the same series) as established context when the user
   follows the handoff button, rather than recomputing everything from scratch.

## What runs free, in the browser, with no account

- **Parsing** comma/tab/whitespace-separated series, with or without a header, with or without
  timestamps, missing-value tokens handled explicitly.
- **Basic stats** — n, mean, median, std, min, max, first, last, total change.
- **OLS trend** on point index — slope, intercept, r², a t-statistic and significance test, and a
  direction (`up`/`down`/`flat`) that only fires when the slope is statistically significant.
- **Seasonality** — autocorrelation on the detrended residuals, searched across a range of candidate
  lags, and required to form a genuine local peak (not merely the largest value tested) before it counts
  as detected - this specifically avoids mistaking an unmodeled level shift for a repeating cycle.
- **A single best changepoint** — an exact O(n) search (via prefix sums) over every possible split point
  for the one that most reduces sum-of-squared-error versus a single whole-series mean, reported with the
  before/after means and the fraction of variance it explains.
- **Anomalies** — z-score on the detrended (and, when a season was found, deseasonalized) residuals,
  threshold |z| ≥ 3.
- **Volatility regime** — first-half vs second-half standard deviation, reported as increased, decreased,
  or stable.
- **A naive one-step-ahead forecast** — linear-trend extrapolation, plus a seasonal-naive adjustment when
  a season was found, with an indicative (not statistically rigorous) ±1.96σ range.
- **Flags**: truncated input, unparseable lines, missing values, duplicate or non-monotonic or
  irregularly-spaced timestamps, a constant series, insufficient data, a detected changepoint, detected
  seasonality, detected anomalies, and a volatility shift.
- **Exports**: the parsed series and the anomaly list as CSV, and the read-out as Markdown - all with no
  run and no charge.

## How the model is held accountable

The engine's facts go **into** the run input as `prescan`, and the model's answer is checked back
**against** them:

- Every flag the browser raised must come back in `reconciliation`, one entry per flag, marked
  `confirmed`, `noted`, `set-aside` or `superseded`. A flag the model ignored is displayed as ignored.
- Every timestamp the model cites anywhere in its reply - in a finding, a key signal, or a priority
  action - is checked back against the actual parsed series. A timestamp that does not appear in it is
  still rendered, but marked `ungrounded` rather than trusted silently.
- Every numeric metric value the model cites is expected to match what `prescan` already computed; the
  house rules in the system prompt forbid inventing or altering it, and forbid asserting a *cause* for a
  changepoint the pasted text does not itself state.

## Deliberate limits

- **A true period of 2 can never be confirmed** by the local-peak seasonality test (it has no interior
  neighbour on its low side to compare against) - a stated, accepted trade-off for rejecting the far more
  common false positive of a slowly-decaying, non-periodic residual.
- **The changepoint search finds exactly one split**, not a full multi-regime segmentation - a series
  with two real regime changes will report only the single split that explains the most variance.
- **z-score anomaly detection can be masked by a single very large outlier** inflating the residual
  standard deviation used to score every other point - a known limitation of the method, not a bug.
- **A paste longer than one run can carry is cut on whole-line boundaries**, with the number of lines
  profiled locally versus sent stated to both the user and the model.
- **This is not financial, medical, investment, staffing, or engineering advice.** Every finding is a
  prompt to verify with the people who own the system that produced the numbers, never a decision already
  made.

## Persistence

Runs are saved to the signed-in user's SkillSafe account in a declared `pulseruns` collection, mirrored in
`localStorage` for instant paint and offline fallback. The mirror is scoped per subject, so two people
sharing a browser never see each other's series, and signing out purges this device's copy.

## Sources

Pulse Desk is an independent derived work inspired by the published agent skill `@k-dense-ai/aeon`
(`k-dense-ai/scientific-agent-skills`, MIT licence). It reimplements a classical, exactly-computable
subset of that skill's task surface - trend, seasonality, anomaly detection, a single changepoint, and a
naive forecast - as deterministic browser arithmetic. It does not use and is not affiliated with,
endorsed by, or maintained by k-dense-ai or the authors of aeon.