Skill version: anthropics/skills@local-install
Skill hash: sha256:1e315589a90f627428525a76e1fc92f47e2b4f681d4e51462c3fa5f7001b3cbe
Model: claude-haiku-4-5-20251001
System prompt hash: not-provided-by-host
Environment hash: not reported by the host
Permission mode: not reported by the host
Host memory: not measured (the record predates 0.4.5 or the probe did not run)
Context hash: not measured (the record predates 0.4.5 or the probe did not run)
Case set version: 1
Case set hash: sha256:67d6b1b0fc25dd87e5d75c1bee4145a7312b03b51a7e2e5a0d19d5d5063cd910
Assay version: 0.3.1 or earlier (the record predates version stamping)
Activation check: not made: the record predates 0.2.0, which began confirming that a selected skill actually loaded

Recorded before the host CLAUDE.md leak was closed (Assay 0.4.5). This run’s working directory sat under the measuring machine’s home directory, and Claude Code loads instruction files from every directory above the one it works in, so that machine’s ~/.claude/CLAUDE.md — a single instruction about an unrelated tool, graphify — would have entered the context. Which instruction files actually entered it was not measured.

The name graphify appears in none of this run’s 90 attempts. No trace does not prove no effect.

Not re-measured: no Skill call in any published run targets graphify, and the one finding the file could plausibly affect — the marketing-skills positioning case, 10 of 10 attempts — was reproduced without it, 8 of 10.

No earlier run was measured under the same conditions. See what changed since the previous run

What it cost

Attempts: 90
Tokens in / out: 3,320 / 505,776
Cost: $6.22
Agent time: 79 min

Combined trigger score (F1) 1.00 · 90 of 90 trigger decisions were correct.