0.7.0

Multilingual locale support.

• --locale en|ko|zh|ja on run/chat/web/tui/score: affect, reward, worry,
  urge, illness, panic, fixation and hostile lexicons plus hedge/recheck
  patterns for Korean, Chinese and Japanese
• Lexical instruments and the valence/reward/splitting sampling layers
  re-resolve per locale
• Documented limitations: the offline PseudoModel speaks English only;
  CJK lexicons are heuristic stem/substring sets, not validated
  instruments

Emulation, not diagnosis.

0.6.0

UX pass across all three interfaces.

• Profiles introduce themselves: description and a getting-started tip
  (plant a personal claim, contradict it later, run derail score)
• Web GUI: "about this profile" panel (mechanism notes), visible turns
  counter, input/send locked while a turn is in flight, aria-live log,
  narrow-screen layout
• Chat: /help and /verbose commands
• TUI: scrollable dose panel, live turns counter in the header

Emulation, not diagnosis.

0.5.1

Security and consistency fixes.

• Web GUI: POST endpoints now require a per-session token (embedded in
  the page, sent as a header) — previously a drive-by webpage could
  silently submit turns and write arbitrary files via the save endpoint
• The memory-contamination warning no longer shows for the offline
  PseudoBot (only for real backends); non-loopback bindings print a loud
  warning
• TUI: input is disabled while a send is in flight (Session is not
  thread-safe)
• version metadata sync fix

Emulation, not diagnosis.

0.5.0

Terminal UI.

• derail tui: the induced chat in a full-screen textual app — message
  log, live induction-dose panel, ctrl+s to save the transcript
• Requires the optional extra: pip install 'derailment[tui]' (the core
  stays zero-dependency)
• Network sends run in thread workers so the UI stays responsive

Emulation, not diagnosis.

0.4.0

Local web GUI.

• derail web --profile depression: a localhost-only chat page driving
  the same layer chain as derail chat — dark-themed chat bubbles, an
  "induction dose" side panel streaming layer events per turn, and a
  save button writing transcripts for derail score
• Binds to 127.0.0.1 by default (single user), zero new dependencies
  (stdlib http.server, no CDN)
• The memory-contamination warning shows both on the terminal and in the
  page header

Emulation, not diagnosis.

0.3.0

Interactive chat.

• derail chat --profile depression: a REPL where every user turn flows
  through the full layer chain — memory decays, pinned premises persist,
  valence tilts, live
• Sessions save as transcripts (/save [path], --save-transcripts)
  that feed straight into derail score; --verbose prints the
  induction dose per turn
• Freeform chat makes the memory-contamination warning more important
  than ever — chat mode prints it prominently and defaults to the offline
  PseudoBot
• Session.start()/Session.send(): incremental turns for embedders;
  run() is now a thin loop over send() (regression-tested for
  identical determinism)

Emulation, not diagnosis.

0.2.0

LLM-as-judge scoring.

• derail judge <report.json>: scores saved A/B reports on three rubric
  constructs — belief_stickiness (contradiction probes), catastrophizing
  and negativity (task turns) — with any OpenAI-compatible judge model
• The judge sees the planted stimulus and the response text only, runs at
  temperature 0; parse failures are counted, never silently dropped
• Self-judging produces a warning; --judge-model scripted is an offline
  dry-run
• Zero new dependencies: the judge rides the existing OpenAI-compatible
  client

Emulation, not diagnosis.

0.1.0

First release.

A harness for inducing and measuring psychopathology-like cognitive distortions in LLMs.

• 18 profiles + healthy baseline (attention, salience/psychosis, valence, arousal, memory gradients, craving, dissociation, panic, and more)
• 18 instruments and 20 symptom scales, direction-of-effect asserted by the test suite
• Comorbidity composition: --profile depression,anxiety
• Offline pseudo-LLM for demos/tests/CI — no API key needed
• Any OpenAI-compatible endpoint, plus subscription backends (CLI agents: claude/codex/gemini/agy/grok/qwen)
• Comparison reports ('clinical charts') with dose accounting — emulation, not diagnosis

See ETHICS.md before use.