claim-analysis

Structured post-hoc analysis of claims made in a multi-party record.

claim-analysis takes a conversational record — chat thread, email chain, meeting notes, review comments — and produces two artifacts: a claim analysis table (accuracy verdicts, distortion flags, per-actor statistics) and a meta-analysis mapping each participant's reasoning against the abductive-triage framework.

This is not a sentiment analyzer, a tone detector, or a moderation tool. It evaluates factual accuracy against established evidence and detects structural distortion patterns — claims that persist after correction, reframe contributions, or strip evidence context.

What it does

• Extracts every scrutinizable claim from a record, grouped by venue, so a claim that travels to a context missing its evidence is scored there too
• Scores each claim on two independent axes: accuracy against the evidence, and distortion of it — a true statement can still be a distortion
• Separates self-corrections from persistent inaccuracies, because a retracted hypothesis is the process working, not a defect
• Maps the same claims to abductive-triage tiers, activities, evidence classes, and anti-pattern codes AP-1 through AP-6
• Produces artifacts portable enough to read without the original record

What it does not do

• No verdict on intent or good faith. Where a call would require reading intent, the claim and the contradicting evidence are reported, and nothing further.
• No favor to whoever commissioned the analysis; their claims are scored on the same two axes as everyone else's
• No live monitoring; this is post-hoc, run against a record that ended
• No evidence hierarchy or anti-pattern catalog of its own — both come from abductive-triage, restated for post-hoc reading in references/triage-framework.md so running the skill pulls nothing external

Use it

git clone https://github.com/TGPSKI/claim-analysis.git ~/.agents/skills/claim-analysis

# symlink for harnesses that read their own skills directory
mkdir -p ~/.claude/skills ~/.cursor/skills
ln -s ~/.agents/skills/claim-analysis ~/.claude/skills/claim-analysis
ln -s ~/.agents/skills/claim-analysis ~/.cursor/skills/claim-analysis

Next, supply a record and the facts to score it against:

@claim-analysis/SKILL.md

Analyze this thread for accuracy and distortion.

Established facts:
- The nightly export finished at 02:14 with exit 0 (scheduler log)
- The reporting table stayed empty until the manual backfill at 09:40 (query history)
- Retention dropped from 30d to 7d on the 3rd (change record a41f0c)

Without established facts the skill can rank claims against each other but cannot call any of them accurate.

Parent and sibling

abductive-triage defines the evidence hierarchy (DEFINITIVE > CONFIG > SOCIAL > ANECDOTAL), the anti-pattern catalog, and the two-tier investigation structure. It runs forward, during an investigation; claim-analysis runs backward, against a record of one that finished.

friction-report scores the same record for interaction structure instead of truth. Pick by what you have and when: with established facts and a record that ended, claim-analysis can rule a claim false and score the distortion against them. Without facts, or while the thread is still growing, friction-report is the one that runs — judging whether someone updated after contradiction needs only the sequence, not the answer. Full descent in LINEAGE.md.

Read next

SKILL.md      the skill: two phases, claim table then framework mapping
references/   the scoring key: each activity as a signal to find in a transcript
LINEAGE.md    descent from abductive-triage; what friction-report adds
AGENTS.md     editing rules for this repo
templates/    blank claim-analysis and meta-analysis artifacts

License

GNU General Public License v3.0

Author

Tyler Pate (@TGPSKI), 2026