| name | verify-inheritance |
| description | Treat memory files, summaries, handoff docs, and other agents' reports as unverified claims. Inheritance tells you where to look; only the current checkout tells you what's true. Apply whenever acting on knowledge you did not derive this session. |
Verify Inheritance
Documented failures from a single knowledge-mining exercise (multiple cheap investigator agents summarizing session archives):
- An investigator attributed benchmark Q&A sessions to human work and built a narrative on it.
- Another invented plausible file paths for a real domain (the domain existed; the paths never did).
- Another declared a merged-and-verified refactor "incomplete, 56% test pass rate" — contradicted by a direct run of the suite.
- A fabricated test filename from one report nearly propagated into the permanent playbook.
- A secret (a password in shell history) nearly propagated into a handoff document.
None of these were malice; all were confident summaries of partially-read evidence. Human-written docs age into the same state.
Rules
- Claims have grades. Verified-this-session (I ran it / read it) > primary record (the actual transcript/commit) > secondhand summary (memory, report, README). Only the first grade justifies action without checking.
- Before acting on any inherited claim about code: grep the symbol, run the command. File paths, line numbers, config names, "X is broken/done" — all decay. Line refs are locators, not truth; resolve by symbol name.
- Quarantine subagent reports. Findings from any delegated investigation are quarantined until ground-truth-checked. Require investigators to mark UNVERIFIED explicitly; treat unmarked confidence as unmarked risk.
- Never propagate secrets. When mining histories or writing handoffs, actively filter credentials, tokens, and passwords — summarizers copy them without noticing.
- Fix the document in the same session you find the divergence. A wrong playbook is worse than none; staleness discovered and left in place becomes the next agent's confident mistake.
- Write your own docs to be verifiable. Pair every fact with its re-verification command ("weights at path/file.py —
grep -n weight_vector"). A doc that can be re-checked cheaply gets re-checked; one that can't gets believed.
The one-line version
Inheritance tells you where to look. Only the repo tells you what's true.