unslop

[0.5.1] — 2026-04-21

Research sync release. Follows the April 2026 update to
docs/research/ (20 categories refreshed, 20 new synthesis files, 14
implementation-trace rows added) and ports the new findings into code,
rules, and documentation. No behavioral break: every previously working
CLI flag, hook, and skill still works identically. AI-ism reduction at
balanced moves from 89.1% to 92.0% on the nine-fixture benchmark.

Added

• Persona-drift reinforcement in hooks/unslop-mode-tracker.js. A
  per-session turn counter at $CLAUDE_CONFIG_DIR/.unslop-turn-count
  (symlink-safe, size-capped, whitelist-validated) fires an expanded
  drift-check banner at turns 8, 16, 24, 32, then every 16 turns
  thereafter. Calibrated against RMTBench (>30% persona degradation at
  turns 8–12) and HorizonBench (arXiv 2604.17283, Apr 2026).
  hooks/unslop-activate.js resets the counter on session start, and
  "stop unslop" also clears it. Three new integration tests in
  tests/test_hooks.py lock the turn-8 banner, the session-start reset,
  and the stop-phrase reset.
• AI-ism vocabulary expansion in unslop/scripts/humanize.py
STOCK_VOCAB and unslop/scripts/validate.py AI_ISMS: meticulous(ly),
  bustling, paradigm shift, game-changer/changing,
  revolutionize(s/d/ing), transformative, unprecedented in
  connective-adjective contexts, a myriad of / myriad,
  a plethora of, uncharted territory/waters/ground/area/domain,
  nuanced as connective filler, synergy/synergies/synergize(s/d/ing).
  Each has a corresponding validator pattern so a rewrite cannot silently
  reintroduce them. Six new tests cover the rewrites, including the
  factual-context guard for unprecedented (e.g. "unprecedented drought"
  survives; "unprecedented opportunity" does not).
• LLM-as-judge bias mitigations in evals/perceived_humanness.py.
  --judges accepts a comma-separated multi-model jury (Claude + OpenAI)
  to counter self-enhancement bias; --counterbalance (default on)
  runs each fixture in both A/B orientations to average out position bias
  (~40% position inconsistency baseline per arXiv 2411.15594);
  length_delta_chars is tracked and reported to surface verbosity
  inflation (~15% per arXiv 2410.02736); per-judge win rates are broken
  out in the summary. Four new tests cover counterbalancing, multi-judge
  juries, all-judges-unavailable, and bias-note reporting.
• Detector-feedback ladder exhaustion recommendation in
  unslop/scripts/detector.py. When the escalation ladder (balanced →
  full → full+structural+soul) can't reach the target AI probability,
  reason_stopped now prints a structured cross-model-paraphrase
  recommendation, names TempParaphraser (EMNLP 2025) and Adversarial
  Paraphrasing (NeurIPS 2025) as the research basis, and explicitly warns
  against watermark removal (EU AI Act Article 50).
• Style-memory security hardening in unslop/scripts/style_memory.py.
  File-size cap at 64 KB on load (StyleProfile JSON is ~1 KB; anything
  larger is either corrupt or adversarial). Expanded module docstring
  documents the OWASP Top 10 for Agentic Applications 2026 memory-risk
  class (InjecMEM, memory control flow, semantic drift), and cites MIT /
  Penn State CHI 2026 for the sycophancy × memory link that drives the
  design constraint of persisting only numerically measured signals (no
  free-text preferences).
• Commercial humanizer tool coverage in README's detector section:
  names Ryter Pro, Walter Writes AI, and GPTHuman.ai alongside the
  existing list; cites Turnitin's August 2025 anti-humanizer update and
  Chicago Booth 2026's audit of twelve services (median ~6-point drop
  vs. claimed 40+). README also positions Anthropic Custom Styles (Nov
  2025) and OpenAI's style-steering as the real first-party comparison,
  and references Antislop (ICLR 2026) as the learned per-pattern framework
  this repo implements.
• SKILL anti-detector step 7: "re-anchor after long contexts". Cites
  RMTBench and HorizonBench directly and lists the re-anchor checklist
  so agents that miss the hook output still self-correct.
• IMPLEMENTATION_TRACE.md rows for every change above, each linking
  the research file, the code location, and (where applicable) the tests.
• Benchmark baseline benchmarks/results/baselines/post-v050-vocab-expansion-20260421.json
  pinned at 92.0% reduction.

Fixed

• Type-check failures in unslop/scripts/structural.py (_SplitCandidate
  now uses collections.abc.Callable instead of a string literal) and
  unslop/scripts/detector.py (explicit str() cast on tokenizer.decode()
  results to satisfy mypy's inferred str | list[str] union).
• Lint failures (ruff) across unslop/scripts and benchmarks from
  earlier mixed-ordering imports and redundant __future__ annotations.
  ruff check unslop/scripts benchmarks and mypy --config-file unslop/pyproject.toml unslop/scripts both clean.

Changed

• evals/perceived_humanness.py --judge-model is deprecated in favor
  of --judges. Backward compatibility is retained for the duration of
  the 0.5.x line; a deprecation warning points to the new flag.
• tests/unslop/test_perceived_humanness.py migrated to the new API;
  tests that assert exact vote counts now pass counterbalance=False
  to keep the prior one-vote-per-fixture accounting.