fresh-context-verify

Fresh-Context Verify

Same-context self-checking ("am I sure?") is proven unreliable — a wrong anchor stays wrong because the reasoning that produced it is still in the window (intrinsic self-correction degrades reasoning; Huang et al., ICLR 2024). The only fix is an EXTERNAL oracle: re-derive the value in a clean context that never saw your assumption.

When to run

• A periodic FALSIFY-CHECK prompt told you to (you have made many edits on one assumption).
• You keep hitting inconsistencies / contradictions and the work is not converging.
• A single number / address / offset / field is load-bearing for a long chain of work.
• Before committing days more effort to a path built on one unverified value.

Protocol

1. Name the anchor. State the exact value or assumption and the ONE claim everything downstream depends on (e.g. "the date gate is checked at sub_X" or "field +Y holds the timestamp").
2. Spawn a clean re-derivation. Use the Agent tool (Explore or general-purpose) with a prompt containing ONLY the raw evidence (the binary / file / data + the question) and NOT your reasoning or your current answer: "Independently determine X from this evidence. Do not assume any prior value." The fresh agent cannot inherit your wrong anchor.
3. Compare. Match → the anchor survives; continue with higher confidence. Differs, or the fresh agent cannot reproduce it → STOP. The anchor is suspect — that is your 2-day-waste early-warning.
4. On mismatch: do not patch around it. Re-derive from the disagreement; if the two derivations still conflict, surface both to the user.

Why a subagent, not "re-check yourself"

The subagent runs in a separate context window — it is the independent sample / fresh reviewer the evidence demands. Re-reading your own chat is not independent; it re-confirms the bias. Keep the fresh agent's input minimal and bias-free, and never tell it your current answer.