Canvas Humanizer

canvas-humanizer

Transform a verified local draft without changing its assignment claims, citations, quoted material, required structure, or platform state. This skill is an optional drafting stage, never a correctness guarantee or submission step.

Inputs and boundaries

Require:

• draft_path: existing .docx or .md;
• output_path: new file with the same extension;
• voice_register;
• course_id, assignment_id, and the caller's stable work_dir;
• optional student_identity, hard_locks, and rubric anchors.

work_dir must end in course-<course_id>__assignment-<assignment_id>, as produced by src.course_artifacts.stable_work_dir. Both input and output must be inside the current assignment's private run tree. Refuse to overwrite the only source copy; keep the pre-pass draft.

This skill performs local file I/O only. Never call Canvas or another learning platform. It does not write assignment result.json; the calling course skill does that after re-running its own verification.

Write humanizer_log.json beside output_path. Diagnostic values in that file are internal pipeline outcomes, not Canvas Pilot result statuses.

Deterministic helpers

Use the existing symbols instead of reimplementing them:

from src.humanizer_segmentation import split_sentences, segment_paragraph
from src.humanizer_segment_extract import extract_locks
from src.humanizer_score import score_candidate, divergence

For DOCX extraction, the existing command is also available:

python -m src.humanizer_segment_extract --draft <draft.docx> --out-dir <state-dir> --voice <voice>

It writes _state.json with paragraph/sentence IDs and masked locks.

Stage 1: preserve before rewriting

Read the complete draft and rubric anchors. Mark metadata/title/name blocks and paragraphs shorter than ten words as passthrough unless the caller explicitly opts them in.

Extract and lock:

• direct quotations and quoted titles;
• dates, years, percentages, measurements, and required numbers;
• author/source names, citations, URLs, and bibliography entries;
• rubric-required headings and exact prompt language;
• caller-supplied hard locks.

Resolve overlaps longest-first and mask each sentence with paragraph-local [LOCK_N] placeholders. Split into sentences before masking. Record original paragraph index and intra-paragraph sentence index; those two values are the canonical segment identity.

Stage 2: generate candidates

For each humanizable sentence, create a small candidate pool using two strategies. Batch independent native Codex subagents to reduce wall time, but keep each subtask bounded to one masked sentence and one transformation.

Strategy R: round trip

Generate up to three candidates through distinct intermediate languages. Each subtask must:

1. translate masked English to the assigned intermediate language while preserving every [LOCK_N] byte-for-byte;
2. translate that result back to English in the requested voice;
3. return only the candidate sentence.

Treat this as controlled diversification, not proof of detector evasion.

Strategy P: structured paraphrase

Generate up to three candidates using different structural operations:

• clause reorder;
• active/passive voice flip when grammatically possible;
• verbal/nominal form flip.

Each subtask preserves all locks and meaning, stays within the requested voice, returns one complete sentence, and returns the original unchanged when the operation is genuinely inapplicable.

Never pass the full private assignment to a candidate subagent. Give only the masked sentence, voice, exact operation, and necessary rubric anchor.

Stage 3: deterministic gates

Substitute lock text back and call score_candidate(...) for every candidate. Reject a candidate if:

• any required lock is missing, duplicated, changed, or reordered improperly;
• its word-count ratio falls outside the helper's length-dependent band;
• it is identical to the original because the transformation was inapplicable;
• it introduces an unresolved placeholder.

Record lock, word-count, and divergence measurements in the log.