| name | honor-audit |
| description | Audit a draft KTH submission for honor-code risk — copied content, missing citations, undisclosed AI use, group-attribution gaps, attendance-record issues. |
honor-audit
A pre-submission read-through that flags concrete risks against Rules 1, 2, 4, and 5. Output is a punch list, not a verdict.
This skill does not detect plagiarism algorithmically. It surfaces patterns a careful human reader would notice and asks the author about them.
When to use
- After the submission is drafted but before it is submitted.
- After
honor-disclose is filled in — the audit cross-checks the disclosure against the artefact.
Checks
Rule 4 — no copying
- Passages with a stylistic shift from the surrounding text (different vocabulary range, sentence rhythm, formality).
- Code blocks whose style differs from the rest of the file (naming, error handling, comment density).
- Suspiciously polished sections that contradict the author's stated skill level or earlier drafts.
- Verbatim or near-verbatim matches to sources cited without quotation marks.
For each hit: ask the author where it came from. If it is borrowed, require quotation/citation or a rewrite in the author's own words plus understanding (honor-defense-prep).
Rule 2 — disclosure completeness
Cross-check the disclosure block against the artefact:
- Any AI or source contribution visible in the artefact but absent from the disclosure → flag.
- Any disclosure entry with
Understood? = no or partial → flag and route to defense prep.
- Conversation history that suggests help received but not disclosed → flag (with the specific evidence).
Rule 1 — group accountability
For group submissions:
- Is every member named?
- Can every member defend every part? (Not "we divided the work" — Rule 1 makes accountability collective.)
- Are there sections only one member touched and the others have not read?
Rule 5 — attendance integrity
If the submission interacts with attendance records (lab sign-offs, seminar tickets):
- Are all listed attendees actually attended?
- Are there signatures or sign-offs for people who were not present?
Cross-cutting
- "AI-flavour" markers (em-dashes, hedging phrases, plausible-but-wrong citations) in submissions where AI use is undisclosed → flag.
- Citations that the author cannot locate or summarise → flag (likely hallucinated).
- Numbers in the text that disagree with numbers in tables/figures → flag (often a sign of partial rewrites by mixed authors/tools).
Output
A bulleted list, ordered by severity:
## Audit findings — <submission>
### Must fix before submission
- <issue> — <location> — <recommended action>
### Should resolve
- ...
### Worth a second look
- ...
### Clean
- <areas checked and found in order>
Then: explicitly ask the author to address each must fix item. Do not produce a "cleaned" version of the submission — the author must do that themselves (Rule 3).