See how an agent evolved

honed report artifacts/graphify -o report.html --reflections artifacts/reflections

Runs, candidates, merges and releases come from the workspace's branches — candidate/<run>/<name>, merge/<id>, release/<version> (and curation/<id>) — and the journal: a merge's entries name the candidates it combined.

report.html is one self-contained file: open it in a browser, keep it with the run, or share it. It combines every branch of the workspace (ratings recorded on one, a curator's changes on another) and follows sections across renames by their lineage ID.

View Shows
Evolution graph The workspace's history as the agent's evolution, newest first: each change commit a row (section changed, reason, branch, author), candidates as lanes forking from their base, merges joined to the candidates they drew on, releases as tags. The merged-from edges are dashed: they come from the journal, since git records a merge as a line from the base. Bookkeeping commits and dead branches can be shown or hidden. Select a row for its reason, diff, and — for a merger commit — the candidate commits it drew on. A run that forked from an older version than the latest release is flagged.
Section versions One section's text at the base, in each candidate (new words in bold), and merged — coloured by where each phrase came from: a candidate, several, or the merger itself.
Changes Every committed change per branch, with its reason and diff.
Ratings A grid of every section against every revision, coloured by its ratings there (in the scale's colours) and marked with what happened (✎ modified, → moved, + added, ✕ removed, ✓ kept). Select a section for its history: a chart, then each change with its reason, commit, and diff, and each rating with the reflection that gave it.
Documents Any configuration file at any revision, compared with another, with each heading's ratings beside it.

From Python, with your scales for exact label order and meaning:

from honed.resources.report import write_report
from honed.resources.evolution import evolution

write_report("artifacts/graphify", "report.html", scales=[USEFULNESS], reflections="artifacts/reflections")
data = evolution("artifacts/graphify")   # the same data as JSON, for your own tools

A section's chart

Selecting a section shows its ratings and changes over the rounds that judged it, three ways:

  • Rounds: a bar per round, with a lane underneath for the changes that round made (a dashed one was never released) and a ✓ when the round kept the section as it was.
  • Versions: the rounds sit in bands for the version they judged, and the changes that made the next version are numbered gates between the bands, listed with their reasons below. A release no round has judged yet shows as an empty band.
  • Score: each round as a dot at its mean rating, sized by its number of ratings, with the score before and after each gate.

On a diverging scale (harmful … neutral … helpful) good labels rise above the baseline and bad ones hang below it; on a sequential one (meaningless … critical) every label rises. A scale's labels take their order and colours from the scale (see Reflections). The report remembers the view you chose.

Token usage

Every reflector, curator, and merger run records what it used, as its harness reports it: uncached input, cache reads and writes, output, and — where the harness reports it — the cost in USD. Claude Code and Pi report cost; Codex reports tokens only. A reflection keeps its usage in its record (reflector.usage); a curation or merge stamps it on each entry it records, and readers count it once per curation or merge ID.

The Token usage tab sums them by kind, by round, by harness and model, and lists every run; the header shows the total. In code, Reflection.reflector["usage"], Curation.curator["usage"] and Merge.merger["usage"] hold the same numbers, and honed.reflection.Usage.from_dict reads them back. Runs made before honed recorded usage show none.