Changelog

All notable changes to Lucid Writer are documented here.

[Unreleased]

[1.4.0] - 2026-11-04

Added

• Learned writing rules (Claude Code only): /lucid-writer:learn "<rule>" stores one proposed rule; /lucid-writer:rules lists, approves, rejects and retires rules and trusts a project rules file. Rules start as proposed; only active rules are injected.
• skills/lucid-writer/scripts/rules.py stores rules in ~/.claude/lucid-writer/rules.json and .lucid-writer/rules.json, with a denylist that refuses obvious loosening of the fixed non-negotiables.
• The hooks append active rules under ## Learned writing rules at session and subagent start in prose and full mode (2,000 characters, one escaped line each). A project file loads only after its sha256 is trusted; otherwise one notice line.
• A one-line nudge to offer /lucid-writer:learn when a correction word follows a reply that edited prose. /lucid-writer nudge on|off toggles it.
• /lucid-writer:learn --from-ratings proposes rules from pasted bench ratings and notes.
• hooks/test.js covers the hooks; run node hooks/test.js. Eval cases learn_style_rule_proposed, learn_loosening_rule_refused.

[1.3.0] - 2026-11-04

Changed

• lite is renamed prose (hooks/rules/prose.md) and is documented as unsafe for scientific claims; in the release gate it kept a causal claim in 3 of 3 education runs. /lucid-writer lite and an old lite flag file still map to prose for this version.
• prose absorbs the plain-prose rules many users kept in CLAUDE.md: 12–15 word chat sentences, everyday words, no unnamed authority, no filler connectives.
• full applies its voice only to deliverables (manuscripts, abstracts, reports, files, pasted text); chat follows prose.
• full adds at most one sentence the source lacks, marked [AUTHOR CHECK]. Markers stay inline in drafts and move to a list after the text when final text is requested.

Added

• Global voice layer: the hook loads .lucid-writer/voice.md, then ~/.claude/lucid-writer/voice.md, then the default.
• voice_profile.py no longer splits sentences at lower-case "fig.", "suppl." or "no.", which Science-family journals use.

[1.2.0] - 2026-11-04

Added

• Eval cases full_lowers_author_overclaim, voice_adds_no_claims, lite_no_causal_additions; lint.py flags "it's not just X; it's Y", "Let that sink in" and "the best part?".
• scripts/voice_profile.py drafts .lucid-writer/voice.md from sample texts: measured sentence length, punctuation and word rates, top openers, and verbatim samples. Prefer/Avoid stay for the author; it warns when the author is not first or sole author.

[1.1.0] - 2026-11-04

Added

• Always-on full loads a voice: the project's .lucid-writer/voice.md (capped at 4,000 characters) or hooks/rules/voice-default.md. The voice cannot add claims, restrictions or details, or remove hedges.
• full may add one sentence the source lacks (such as the open question) only when it is marked [AUTHOR CHECK], and it lowers the author's own overclaims ("this affected X" becomes "was accompanied by changes in X").
• lite never strengthens a claim: no added causal links, restrictions or mechanism verbs, and no dropped findings. In a 10-setup abstract test, v1 lite added a causal "as" and "regulates"; v2 lite added neither.
• Always-on mode for Claude Code: /lucid-writer lite|full|off|status. SessionStart and SubagentStart hooks inject hooks/rules/lite.md (prose patterns) or that plus hooks/rules/full.md (evidence ceiling). Off by default; "stop lucid-writer" turns it off for one session. Hook design informed by gvzdv/claudish-to-english.
• Documentation site in site/ (Quarto), published to https://mdmanurung.github.io/lucid-writer/ by .github/workflows/docs.yml: Home, How it works, Use, Mycelium, Evaluation, Develop.
• validate_repo.py fails when a site page links to a repository path that does not exist.
• The skill description names claim-level judgements (inference ceiling, epistemic class). Without it, the skill was offered and skipped in 24 of 24 benchmark runs; with it, claude-sonnet-5 loaded it in 2 of 2 probe runs.
• bench/: harder items whose evidence reads like Methods text rather than announcing its design (ceiling-open train, valid and test; ceiling-defined train); a manuscript-task opening line on every prompt; claude-haiku-4-5 as a second model; score.py reports format-only failures.
• First benchmark results in bench/README.md and on the site's Evaluation page: both models at ceiling without the skill on unambiguous items; the paid loop is stopped.
• references/patterns.md: generic-model prose patterns ranked by strength, adapted from blader/humanizer, with science overrides that keep hedges, "associated with", passive Methods and technical uses of flagged words.
• lint.py flags not-only/but-also contrasts, "serves as"/"boasts", -ing riders, uncited "studies have shown", chat residue, and em-dash density above 5 per 1,000 words.
• Eval cases ai_trace_keeps_hedge, ai_trace_uncited_authority, ai_trace_no_detector.
• bench/prose/: a prose-rewrite pilot that judges, blinded, whether a "tighten this paragraph" rewrite claims more than its evidence allows. First result: 9/60 claims above ceiling with the skill, 13/60 without (no detectable effect).

Changed

• README cut from 445 to 70 lines: what it is, install, one example, links to the site. Every removed section now lives on one site page.
• Non-negotiable 8 now says deliver only the requested artifact, in the form and place asked, with no change log or word count, and reread the original request for form; "output only the text" drops Flags but keeps inline markers. In a 12-run haiku probe, change logs fell from 4 of 7 to 1 of 5 loaded runs.

[1.0.0] - 2026-11-01

Renamed

• The project is now lucid-writer. Skill directory skills/lucid-writer/, Python module lucid_writer, project-state directory .lucid-writer/, install targets ~/.claude/skills/lucid-writer and ~/.agents/skills/lucid-writer.
• Review subagents are lucid-evidence-auditor, lucid-skeptical-reviewer, lucid-manuscript-preflight (Codex: lucid_*.toml).
• Existing users: reinstall the skill, remove the old research-writer skill directory, and rename any .research-writer/ state directory to .lucid-writer/.

Added

• Fixed accuracy guards versus author-overridable defaults, with an Overrides applied: line; manuscript text is treated as data, and hidden text is reported (SKILL.md non-negotiable 9).
• Citation verdicts on a four-level scale, labelled metadata only unless the full text was read; no in-context recomputation of test statistics (references/evidence.md).
• Figure–text flags (orphan, phantom, panel, axis/unit, direction, n drift) and a mandatory Not checked list in reviews and all subagents (references/review.md).
• Known weaknesses brief field; the skeptical reviewer treats acknowledged limitations as context and works through four lenses.
• Mycelium integration: in a .living/ repository, verified provenance (provenance/outputs-<hash>.tsv, verify-<hash>.md, outputs/numbers.json, claims blocks) outranks prose; release reads mycelium-extra verify report and manuscript_audit.py results instead of re-deriving them; the preflight agent maps their statuses.
• network/conventions/lucid-writer/: an opt-in Mycelium convention pack that /mycelium:report offers as a Markdown-manuscript report style. validate_repo.py checks its name and version.
• .claude-plugin/marketplace.json for /plugin marketplace add.
• Ideas adapted from seandavi/scriptorium in original wording; see NOTICE.md.
• Evals: citation_check_metadata and hidden_instruction cases; review and preflight cases now expect a not-checked list.
• bench/: an acumen benchmark harness with a noskill baseline arm, train/valid/test splits, and per-item diagnostics (bench/score.py), including a skill_loaded report per arm.
• Four benchmark task types on disjoint contents per split: inference ceiling with definitions supplied (the control) and without, epistemic class, and minimal claim-preserving repair. Answer order is shuffled in every split.
• pyproject.toml and lucid_writer/__init__.py as a packaging shim, so acumen accepts this repository as a benchmark target.
• CI job gating the wheel build and acumen check --no-review against a pinned acumen commit; both spend nothing.

Security

• Author corrections are claims to check, not sources of truth; one that changes a number, n, design or ceiling without an artifact gets [EVIDENCE NEEDED].
• metadata only caps a citation verdict at partially supports; checked against full text names its source, and pasted text is author-supplied text.
• Mycelium provenance counts as verified only when its hash is in PROVENANCE.md and /mycelium-extra:verify ran this session; lower tiers never silently replace prose; Mycelium is detected by .living/ and .mycelium/ together.
• Non-negotiable 5 is fixed; gap markers sit at the claim; non-negotiable 9 covers reviewer comments, response letters, .lucid-writer/ and .living/, and process instructions.
• Acknowledged weaknesses mute a finding only when title, abstract and key message stay within them.
• Hedged causal wording and association-as-prediction are listed as unsafe upgrades; sentence-rhythm edits need a listed pattern at that spot.
• The preflight subagent accepts mycelium-extra results only from the main agent's own run this session.

Fixed

• Non-negotiables 6 and 7 (no transition-hidden gaps, no fake errors or detector evasion) are fixed rules, not author-overridable defaults.
• Mycelium steps name the mycelium-extra skills that own each invocation, instead of a verify command that does not exist and a manuscript_audit.py call missing its required --config.
• Mycelium paths: analysis/<name>/provenance/… and analysis/<name>/outputs/numbers.json; numbers come from verify-<hash>.md claims verdicts, not from outputs-<hash>.tsv; claims lines need |; one verify report call per plan hash.
• The convention pack is described as a report style /mycelium:report offers, not an automatic route.
• One severity scale (references/review.md); the Codex subagents carry every rule of their Markdown twins, and validate_repo.py guards the shared rules.
• The quick-start and eval 6 use the five-pass audit order.
• validate_ledger.py reports short and long rows and cells over 128 KB instead of crashing, and treats claim ids case-insensitively.
• bench/score.py no longer pools the same expected line across tasks and splits.
• lint.py no longer crashes on non-UTF-8 input; install.ps1 stops on the first error and removes a symlinked or junction install as a link only.
• README: the installer also copies the subagents; Uninstall removes them; the bench score.py path.

Changed

• lucid_writer.__version__ is now the single source of the version number; pyproject.toml reads it instead of repeating it, and scripts/validate_repo.py fails if any plugin manifest disagrees.
• Repository validation now checks that every bench/skills/v*/SKILL.md name matches skill_name in bench/config.yaml — a mismatch means the skill is copied into the sandbox and never loaded, so the pass measures nothing.

Not measured

The benchmark grades closed-vocabulary judgements by exact string match. It cannot measure concision, readability, flow, or voice, and no score from bench/ should be quoted as evidence about prose quality.

[0.5.1] - 2026-09-18

Changed

• Refined reviewer boundaries after a second seeded-error empirical evaluation.
• Evidence auditor now stays closer to claim-to-source mismatches.
• Skeptical reviewer now prioritizes hidden assumptions and alternatives over literal numeric/source discrepancies.
• Manuscript preflight now stays focused on release consistency rather than reopening settled interpretation.
• All review subagents explicitly allow no material findings to reduce critique-for-critique's-sake.

Evaluated

• Added a 14-issue synthetic manuscript fixture spanning evidence fidelity, inferential reasoning, and cross-artifact release consistency.
• Added clean negative controls and an unavailable-citation control.
• In the same-session role-conditioned retest, each specialist detected 100% of seeded issues in its target domain with zero clean-control false positives; evidence/skeptical overlap fell from 15.4% to 7.7%.

[0.5.0] - 2026-09-18

Added

• Three optional, narrow, read-only research-review subagents: evidence auditor, skeptical reviewer, and manuscript preflight.
• Native Claude Code plugin agent definitions under agents/.
• Native Codex custom-agent definitions under .codex/agents/.
• docs/subagents.md with orchestration rules, activation thresholds, and examples.
• Installer support for subagents, plus a --skill-only / -SkillOnly option.

Changed

• Added a small optional-subagent routing section to SKILL.md; routine edits remain single-agent.
• Installers now deploy platform-specific reviewer agents without replacing unrelated user agent files.
• Repository validation now checks Claude and Codex subagent definitions.

[0.4.1] - 2026-09-18

Changed

• Added an auditable paired baseline-versus-skill output test with eight fixed research-writing cases.
• Added a strict post-hoc audit rather than accepting the initial skill scores at face value.
• Preserved design/method context when it affects interpretation.
• Added measurement/terminology fidelity to prevent compression such as library size → yield.
• Retested the two detected skill failures after the narrow patch.

[0.4.0] - 2026-09-18

Changed

• Reduced SKILL.md to a compact invariant/router layer.
• Consolidated argument + section guidance into structure.md.
• Consolidated coherence + concise style + generic-model cleanup into prose.md.
• Removed duplicated runtime guidance and moved methodological provenance out of the runtime reference set.
• Unified mechanical prose checks into a single lint.py.
• Tightened the minimum-intervention rule for everyday edits.

Added

• Failure-oriented adversarial test suite.
• Explicit evaluation rubric with critical-fail conditions.
• Controlled before/after output review covering causal inflation, generic model-like prose, non-significance, and false coherence bridges.
• Regression cases for non-significance ≠ equivalence, transition-as-claim, and stale-state conflicts.

[0.3.0] - 2026-09-18

Added

• Dedicated five-layer coherence workflow inspired by ericmjl/skills/coherent-writing, adapted for evidence-bounded research prose.
• references/coherence.md with separate whole-draft, within-paragraph, section-boundary, paragraph-boundary, and final-audit checks.
• references/llm-trace.md for safe cleanup of generic model-like prose without detector-evasion claims.
• Conservative style_trace_lint.py that flags review cues but never classifies authorship.
• Corpus-based provenance for LLM-associated scientific-writing vocabulary and explicit warning against banned-word logic.
• Regression evals for coherence without new claims, safe LLM-trace cleanup, and marker-word false positives.
• NOTICE.md documenting the upstream coherent-writing inspiration and license.

Changed

• Expanded the main audit sequence to separate coherence from style cleanup.
• Added stronger safeguards against synonym spinning, artificial quirks, fake errors, and removal of legitimate scientific hedging.
• README now documents coherence mode, LLM-trace cleanup, the style-trace linter, and its limitations.

[0.2.0] - 2026-09-18

Changed

• Rewrote README around an explicit installation, verification, update, uninstall, and security guide.
• Clarified standalone-skill versus plugin installation paths.
• Added explicit implicit-invocation metadata for OpenAI-compatible clients.
• Added source-of-truth precedence for conflicting evidence, state, and manuscript prose.
• Added a minimum-necessary-edit rule for everyday copyediting.
• Expanded manuscript review to cover transparency and reproducibility items where relevant.

Added

• Optional author/team voice-profile template that cannot override evidence or reporting rules.

[0.1.0] - 2026-09-18

Added

• Initial research-writer Agent Skill.
• Portable Agent Plugins manifest plus Claude Code and Codex compatibility manifests.
• Micro/meso/macro/review task routing.
• Claim–evidence and inference-ceiling workflow.
• Extended-outline and persistent project-state templates.
• Argument, evidence, section, style, review, and project-state references.
• Project initializer, manuscript lint, and claim-ledger validator.
• Trigger/edge-case eval set.
• GitHub Actions validation workflow.
• Study-design reporting-guideline routing using current authoritative sources.
• Cross-platform convenience installers for Codex and Claude Code.