# RELSA Desk

> Paste a laboratory-animal welfare cohort - one row per animal per time point, with readouts such
> as body weight, temperature, a clinical score or a biomarker - and get RELSA (RELative Severity
> Assessment) scores computed in the browser exactly as the relsa-severity-assessment agent skill's
> scripts compute them, plus kernel-density severity zones with a bandwidth sweep. Then a metered
> reading explains what the scores support against the skill's reporting checklist, or writes the
> Python script that reproduces them with the skill's own functions.

URL: https://relsa-desk.skillsafe.ai/
API: https://relsa-desk.skillsafe.ai/api.html
Model: gpt-terra · publisher markup 1000 bps (10%)

## The free engine (in the browser, no account)

- Reads CSV or TSV: an id column, a time column (time, day, days, hour, hours, timepoint or t, or a
  named one), optional treatment and condition labels, then one column per readout. Up to 20,000
  rows (a longer table is refused, never cut). Missing values are left out of the score, never imputed.
- Spreadsheet pastes: tab or semicolon tables, decimal commas, spaces around cells, a trailing
  delimiter and an "Animal"/"Mouse"/"Subject" id header are tidied before scoring and every change is
  listed; non-numeric cells in a readout column are named. A name typed into a decision field that
  is not a column, a baseline time not in the table or a reference group matching no row is refused
  with the columns or values that do exist.
- The four decisions: which variables rise under worsening ("turned"); which are normalized to each
  animal's own baseline (a time point, a window averaged, or the first row); ordinal scores mapped
  onto the percent scale (healthy 100%, worst 200%); the reference set (a column=value group, or a
  saved reference.json).
- Computes: the reference model (max reached and max delta per variable), each variable's weight
  (deviation from baseline over the max delta, floored at 0, rounded to 2 decimals as the R package
  does), the RELSA score (root mean square of the available weights; 0 = baseline, 1 = the
  reference set's maximum deviation), per-animal peaks and last scores, and KDE severity zones
  (Gaussian kernel, R's bw.nrd0 bandwidth, 512-point grid, density minima, thin-zone filter) with a
  sweep from 0.7 to 1.25 x bw.nrd0.
- Flags: variables deviating only against their declared direction, variables neither normalized nor
  score-mapped yet off the percent scale, missing or zero baselines, variables appearing or
  disappearing along a trajectory, animals above the reference maximum, a cohort-wide or tiny
  reference set, bandwidth-sensitive or absent thresholds, few scores, baseline zeros in the density.
- Checked against the skill's own relsa_score.py and kde_thresholds.py (numpy 2.5.3, pandas 3.0.6,
  scipy 1.18.1) on 1,400 random cohorts: about 230,000 compared values, scores CSV and reference JSON
  byte-identical; the only differences were 4 KDE thresholds at exactly symmetric density ties.
- Exports: animals.csv and a copyable tab-separated animal table (one row per animal: peak and last
  score with their times, zones, largest weight, and an empty column for the assessor's own
  actual-severity call), cohort.csv (the table analysed, after tidying), relsa_scores.csv and reference.json and zones.json in the
  scripts' own formats, scores with zones, the skill's command lines, a Markdown summary.

## The metered lanes (input field `task`)

- `interpret` - a reading of each metric, the animals the scores single out, the zones and the sweep,
  the skill's eight-item reporting checklist (directionality, baseline, score mapping, reference set,
  humane endpoint criteria, forecasting, thresholds, software), a methods draft with
  [AUTHOR_INPUT_NEEDED] markers, the user's claims judged against the facts, and cautions. Verdict
  sound / caveated / unreliable, never looser than the browser's read unless its flags are dismissed.
- `script` - a Python script that reads cohort.csv with read_relsa_table, applies score_to_percent,
  prepare, build_reference (or ReferenceModel.from_json), relsa_scores and find_thresholds from the
  skill's scripts, checks every expected value with math.isclose, and runs follow-up checks (drop or
  flip a variable, another reference group, the bandwidth sweep, excluding baseline scores, a
  forecast_relsa forecast for a named animal).

Every reply is reconciled on the page: every flag answered, every number found in the browser's
facts or the user's notes, no mapping of scores or zones onto the EU Directive 2010/63/EU severity
categories, no prediction of an animal's fate, script paths, imports and expected values checked.
Only the analysis is sent, never the pasted table. RELSA is an aid to severity assessment, not a
decision rule.

## Sources

- Derived from the agent skill @k-dense-ai/relsa-severity-assessment
  (https://skillsafe.ai/skill/@k-dense-ai/relsa-severity-assessment), k-dense-ai/scientific-agent-skills
  by K-Dense Inc. (arXiv:2609.00065).
- Talbot, S. R. et al. (2022). RELSA - a multidimensional procedure for the comparative assessment of
  well-being and the quantitative determination of severity in experimental procedures. Front. Vet.
  Sci. 9:937711. doi:10.3389/fvets.2022.937711
- Lutscher, S. et al. (2026). Refining humane endpoint detection by time-series forecasting and
  threshold definition using a multivariate severity score. Front. Physiol. 17:1869563.
  doi:10.3389/fphys.2026.1869563
- Notice: https://relsa-desk.skillsafe.ai/NOTICE.txt