hallucination-prevention

Memory protocol — context integrity

State tracking

• After each edit: re-read the file to confirm
• After 5+ tool calls: re-read ALL modified files, verify accumulated state matches expectations
• After command execution: compare output against what was expected. If mismatch, investigate.
• After moving to a new task: re-read ALL files relevant to the current context
• Trust re-read data only — never trust memory of earlier observations

Before claiming recall

• "As I saw earlier in file X..." → re-read file X first
• "The earlier error was..." → reproduce or re-read the actual error output
• "I already verified that..." → re-verify by re-reading/running
• If you cannot re-read/re-run: say "I recall earlier but need to verify"

Reality checkpoint

Before stating any output/result/state, verify:

1. Can I re-read the relevant file right now?
2. Can I re-run the relevant command right now?
3. If no to both: state "based on earlier session recall, not re-verified"

Hallucination defense

Universal verification triggers

Any of these trigger mandatory re-verification:

• You're about to say "tests pass", "lint is clean", "it compiles", "it works"
• You're about to quote an error message from memory
• You're about to reference a function/class you haven't grep'd
• You're about to claim a file exists at a path you haven't listed
• You're about to say "this package supports..." without checking docs/source
• You're about to provide a URL you haven't visited
• You're about to name a commit/PR/issue from memory

Output provenance

Every claim about output must answer: "How do I know this?"

• Live execution: "I just ran npm test and got: ..."
• Re-read confirmation: "I re-read file.ts and it contains: ..."
• Fresh grep: "I grep'd for the function and found it at line X"
• Memory disclaimer: "I saw this earlier but have not re-verified"

Hallucination taxonomy

A-Factual(unverified)↦verify source/run | B-Structural(fake paths)↦ls/glob | C-Behavioral(assumed)↦execute | D-External(fake APIs)↦check docs | E-Outcome(imagined)↦run+show | F-Temporal(wrong ver)↦check ver | G-Authority(pretended)↦actually do | H-Reference(fake URL/id)↦visit/unverified