Add Register Awareness section + register-shifting checkpoint#106
Open
SipengXie2024 wants to merge 3 commits intoblader:mainfrom
Open
Add Register Awareness section + register-shifting checkpoint#106SipengXie2024 wants to merge 3 commits intoblader:mainfrom
SipengXie2024 wants to merge 3 commits intoblader:mainfrom
Conversation
Targets darwin rubric dim 3 (edge case coverage) and dim 8 (effectiveness). Baseline run on 2026-04-27 found that rule 26 (hyphenated word pair overuse) and the Personality and Soul section misfire on academic and technical prose: stripping hyphens from domain compound modifiers like 'cold-cache compile time' and 'per-contract dispatch identity' creates syntactic ambiguity, and injecting first-person voice into multi-author scholarly papers breaks the intended register. This patch inserts a Register Awareness section between Personality and Soul and CONTENT PATTERNS. It defines five registers, lists rules safe in every register, and specifies overrides for academic / technical inputs (preserve hyphens on domain compounds, keep person, keep scholarly verbs, apply rule 8/10/14 selectively). It also tells the skill to skip rules entirely on code blocks, quotes, and bibliography entries, and to handle mixed-language inputs without translation or diacritic stripping. No existing rule wording changed. File grows from 559 to 606 lines, well under the 150 percent budget.
Targets darwin rubric dim 4 (checkpoint design). The Round 1 Register Awareness section gives static rules for restraining specific edits in academic and technical registers, but the skill still has no runtime safety net for cases where the register triage misclassifies the input. This patch adds an explicit checkpoint in Process step 11 and surfaces it in Output Format. Before delivering the final rewrite, the skill must scan the diff and list any changes that fall into the high-risk categories: person switch, stripped hyphen on a compound modifier, informalized scholarly verb, removed em dash from academic prose, voice injection into scholarly prose, technical term replaced with a bland equivalent. Each flagged edit is shown as 'original phrase' to 'rewritten phrase' and the user is asked whether to keep it. In batch or automation contexts where no user is available, the skill defaults to reverting flagged edits and notes the reversion. This preserves the safety net without forcing a blocking question. Also tightened Process step 5 to remind the skill to keep scholarly verbs in scholarly prose, and added register classification as the first item in Output Format so the user can audit the triage call.
Targets darwin rubric dim 2 (workflow clarity). The top of the file had a six-step Your Task list that duplicated the canonical Process at the bottom of the file. Two parallel workflow descriptions force the reader (and the executing agent) to reconcile which is authoritative, and the top list omitted the Register Awareness triage and the register-shifting edits checkpoint added in earlier rounds. This patch replaces the Your Task block with a short pointer that names Process as the canonical sequence and lists the upstream sections to read first (Voice Calibration, Personality and Soul, Register Awareness, rules). Process and Output Format remain the only authoritative workflow definitions. No rule wording or step semantics changed.
This file contains hidden or bidirectional Unicode text that may be interpreted or compiled differently than what appears below. To review, open the file in an editor that reveals hidden Unicode characters.
Learn more about bidirectional Unicode characters
Sign up for free
to join this conversation on GitHub.
Already have an account?
Sign in to comment
1 participant
Add this suggestion to a batch that can be applied as a single commit.This suggestion is invalid because no changes were made to the code.Suggestions cannot be applied while the pull request is closed.Suggestions cannot be applied while viewing a subset of changes.Only one suggestion per line can be applied in a batch.Add this suggestion to a batch that can be applied as a single commit.Applying suggestions on deleted lines is not supported.You must change the existing code in this line in order to create a valid suggestion.Outdated suggestions cannot be applied.This suggestion has been applied or marked resolved.Suggestions cannot be applied from pending reviews.Suggestions cannot be applied on multi-line comments.Suggestions cannot be applied while the pull request is queued to merge.Suggestion cannot be applied right now. Please check back later.
Summary
Three commits adding register-aware restraint to
humanizerso it does not over-edit academic, technical, or code-adjacent prose. Discovered while running an autonomous skill-optimization loop (Karpathy autoresearch pattern, 8-dimension rubric scored from 70.9 baseline → 85.75 after these changes).Commits
074345a— add Register Awareness section (between Personality and Soul and Content Patterns). Lists 5 register types (academic / technical / quote / code / default), names which rules are SAFE in every register, and which need restraint in academic/technical (rule 26 hyphens, rule 8 copula, rule 14 em dashes, rule 10 rule-of-three, scholarly verbs, person preservation).f3c8c98— add Process step 11: register-shifting edits checkpoint. Six categories of edits to flag (person change, hyphen stripped, scholarly verb informalized, em dash removed from academic prose, voice injected, technical term replaced). Includes a batch-mode fallback that defaults to keeping the original wording when the user is unreachable. Updates Output Format withRegister classificationandFlagged register-shifting editsitems.f41073b— remove Your Task / Process duplication. The top-of-file 6-step "Your Task" list duplicated the Process section; replaced with a one-line pointer to Process as the canonical sequence.Net effect
git log -p | grep -E '^[+-]### [0-9]+\.')per-contract,cold-cache,native-artifact,micro-level), academic pluralwe, and scholarly verbs (shows,reports,preserves) now correctly survive editing of academic inputVerification
Tested on three prompts: AI-flavored blog (happy path), academic LaTeX paragraph from a real systems paper (restraint test), and product release note (formatting noise). The academic prompt's 4 hyphenated compounds and "We report that speedup as an observed systems effect" hedge are now preserved across the optimized humanizer pass.
Test artifacts and per-round scoring matrix live at https://github.com/alchaincyf/darwin-skill (see runs/humanizer-20260427/FINAL_REPORT.md).
🤖 Generated with Claude Code