Put the hallucination-hunter agent where it can actually be found

[AlveeeRahman]

The hallucination-hunter agent has shipped since 1.1.1 and was almost invisible — two prose sentences, and nothing else. Not in either flowchart, not on the Pages site, not in the repository description. A subagent nobody can see is a subagent nobody runs.

Why the map never showed it

skill_mapper.py builds its graph from resolved markdown links in SKILL.md. SKILL.md named the path in backticks — which is text, not a link. So the only route to the file was through README.md, and README.md is not on the path Claude follows under progressive disclosure.

The file was reachable by a human browsing GitHub, and not by the agent that was supposed to use it.

Making it a real link takes the map from 35 resolved links to 36 and puts SKILL.md one hop from the agent. Verified with the tool, before and after:

before:  31 files mapped · 35 resolved links   →  n_README_md --> n_agents_hallucination_hunter_md
after:   31 files mapped · 36 resolved links   →  n_SKILL_md   --> n_agents_hallucination_hunter_md
                                                  n_README_md  --> n_agents_hallucination_hunter_md

That is the same class of defect skill_mapper.py exists to surface, found in this repository by running this repository's own tool on itself.

The conceptual flowchart was also wrong

It said "six independent inspections" and drew five — claim_auditor.py was missing from the diagram entirely. Both it and the agent are now in it, with the agent placed where it actually sits: downstream of claim_auditor, taking its UNVERIFIED list as a worklist rather than starting from prose.

The "genuine output" block had drifted

Regenerated from the current tool rather than hand-patched — the node-id scheme changed at some point (SKILL_md → n_SKILL_md) and the committed block still used the old one. It stays trimmed to README size, and says so.

New: a features table

The README had no at-a-glance summary. Added one covering all seven scripts and the agent, with a column saying what each one answers rather than what it is called — and marking plainly which one is not a script.

Pages site

Gains a "Where the scripts stop" section explaining the boundary the agent exists to cross: a deterministic check can prove a documented flag does not exist, but not that a cited paper fails to say what it is cited for, that a version went stale, or that a benchmark number has no source. Those need judgement, and judgement is where a model hallucinates — so the agent is built to make guessing expensive: tools first, evidence per finding, UNVERIFIED rather than a guess, and a closing statement of what it did not check.

Also adds the agent to the resources list and to .github/repository-metadata.yml.

Checks

• claim_auditor.py → TRUTHFUL (0 contradictions)
• script_tester.py → exit 0
• pytest tests → 196 passed
• JSON-LD parses; meta description 163 chars

scripts/ and tests/ are untouched — this is documentation and one link.

🤖 Generated with Claude Code