4.0.0 drops the no-fabrication rule, and the pattern 17 example brings back the invented NYT interview from #187
Maintainers usually reply within 6 days
Nobody has claimed this yet.
Assessment
- Difficulty
- 2/5
- Estimated time
- 1-3 hours
- Newbie friendliness
- 67/100
- Issue type
- Bug
- Clarity
- Mostly clear
- Activity status
- Active
- Tech stack
- python
- Domain
- ai, documentation
Research direction
Read the 4.0.0 changes in #331 alongside the earlier rule in #189 and the discussion in #187. Find the skill’s Step 2, Step 3, and pattern 17 example, then compare them with the three proposed remedies in this issue. Done means the default avoids unsupported additions and the example no longer presents an invented interview as a source.
Written by the indexing model from the issue text.
Description
4.0.0 (#331) removes the rule from 2.9.0 (#189) that a rewrite must not add a fact, name, number, date, quote or citation that is not in the source. Step 2 now says "Where the text is generic, add a specific detail you know to be true", and step 3 checks only for lost or changed facts, no longer for added ones.
I see two problems with this.
-
The only guard left is the model's own belief. A model cannot reliably check whether it "knows" something to be true. I use the skill on emails, PR descriptions and documentation written in my name. A plausible but wrong date, figure or source in those is worse than a generic sentence, and harder to catch in review. #331 says as much: "This reverses a rule, with no model eval. The skill may now add a detail that turns out to be wrong."
-
The pattern 17 example now demonstrates fabrication. Its new After ("In a 2024 New York Times interview, she argued …") is the example #189 removed after #187, and #331 describes it as "an unnamed, made-up person". The skill's own reference example shows the model inventing a quote and its source.
Any of these would help:
- Restore the source-or-user rule as the default, and allow added specifics only when the user supplies the facts or opts in.
- At least keep "an unsupported addition is an error" in step 3, so added claims get the same check as lost ones.
- Replace the pattern 17 After with one that keeps the source's claim or cuts it.
3.0.0 works as expected for me, and I'm staying on it because of this change.
- Dominant language
- Python
- Stars
- 53.9k
- Forks
- 4.3k
- Avg merge
- 6d 11h
- Merged PRs (30d)
- 9
Getting set up
This project ships no dev container, Dockerfile or contributing guide, so setting up is up to you: start from its README, and see our first-contribution guide for the general steps.
First steps
- Read the whole issue, then the project's contributing guide.
- Comment on the issue to say you are picking it up — it saves two people doing the same work.
- Fork the repository and make your change on a branch.
- Open a pull request that references the issue number.
More from blader/humanizer
-
validate-package.py: duplicated pattern section in SKILL.md passes validation (dict key collapse defeats the numbering check)Possibly taken @TanmayShah29 claimed this today. Open
Difficulty 2/5 1-3 hours Newbie friendliness 25/100
Maintainers usually reply within 6 days
-
Difficulty 4/5 3-5 days Newbie friendliness 57/100
Maintainers usually reply within 6 days
All issues in blader/humanizer
Similar issues
-
bug(mcp): handle malformed and partial reranker responses safelyPossibly taken @KamilEjaz890 claimed this today. Open
Difficulty 2/5 1-3 hours Newbie friendliness 85/100
kubeflow/docs-agent#301 ·
Maintainers usually reply within 4 days
-
Update the demo docsOpen
Difficulty 1/5 Under an hour Newbie friendliness 72/100
pulp/pulp_workflow#78 ·
-
Bug
Difficulty 2/5 1-3 hours Newbie friendliness 62/100
pgadmin-org/pgadmin4#10797 ·
Maintainers usually reply within 2 days
-
Difficulty 2/5 1-3 hours Newbie friendliness 66/100
gcivil-nyu-org/team5-mon-fall26#132 ·
Maintainers usually reply within 1 day
-
Difficulty 1/5 Under an hour Newbie friendliness 76/100
gcivil-nyu-org/team4-mon-fall26#291 ·
Maintainers usually reply within 1 day