aot-evals 0.2.0

• aot-evals run --resume RUN_ID carries on a run that stopped or died partway, in place, instead
  of starting over. Every (case, trial) with a final attempt is skipped, and the rest are appended
  to the same outcomes.jsonl. Resume is refused if the run is finished, or if the agent commit,
  the suite's commit, the env, the judges or the case set changed since the run started. A
  truncated last line is dropped, and a case interrupted mid-retry is run again. The manifest
  gets a resumed list (at, pairs_done_before) and a judges_enabled field. Counts, cost and
  observed models are now always recomputed from outcomes.jsonl. A stopped run prints the
  resume command.
• In outcomes.jsonl, an attempt's check lines now come before its attempt line, so a crash
  between them can't leave a final attempt without its checks.

aot-evals 0.1.1

• The discovery report's pillars read "Safe testing environment" and "Golden test cases"
  (were "Safe testing env" and "Golden use cases"). Reports written with 0.1.0 still render
  as before.

aot-evals 0.1.0

The first public release.

• Skills that take a team from no evals to a measured, comparable suite for their own agent:
  /aot-evals:start is the one command to remember: it runs discover (Phase 1) when there
  is no discovery report yet, and otherwise status, which routes to scope, build,
  calibrate, run and report (Phase 2, commands C0–C18). Only start, discover and status
  appear in the / menu; Claude loads the step skills as it reaches them.
• A discovery page that opens with an overview: what AOT Evals does, a bottom line, three key
  numbers, and the four things good evals rest on (business context, agent architecture, safe
  testing env, golden use cases), each with what discovery found and what's still open. The
  details follow, grouped by those four, then the next steps.
• The aot-evals CLI: gates, import, allocation, screening, runs with per-trial outcomes and
  operational counters, judge labelling (with a browser labelling page), calibration and
  reliability, paired comparisons, coverage, redundancy, report.json, checkpoints and an
  activity log. aot-evals demo sets up a toy agent with a finished suite; aot-evals doctor
  checks the environment.
• Judge backends: System One (POST /v1/systemone), Anthropic, and any command of your own.
• Several agents per repository (evals/agents/<agent>/), isolated runs that keep the agent
  away from the cases, run pacing, and a stop when the agent's provider goes silent.
• Pseudonymous milestone telemetry (Phase 1 started, discovery written, checkpoints, report built,
  demo run), off whenever Claude Code's own telemetry is off, on cloud providers and in CI.
  aot-evals telemetry status|show|off|on. What's sent and the opt-outs: SECURITY.md#telemetry.
• The method reference, docs/method.md.