context-durabilitylisted

# Context durability A long session gets **compacted**: earlier conversation is replaced by a summary, and recent context is kept. This is normal and usually harmless. It is not harmless in finance, for one specific reason: > **Re-deriving a number does not fail loudly. It silently produces a different number.** A lost variable in code throws an error. A lost running total in finance work gets recomputed with a slightly different scope — one exclusion forgotten, one threshold remembered wrong — and the second answer looks exactly as confident as the first. Nobody finds out until a tie-out fails three steps later, or until nobody finds out at all. So the rule is not "avoid compaction." Compaction is unavoidable on long work. The rule is: > **Anything you would have to re-derive must be on disk before you need it again.** --- ## What evaporates Compaction keeps a summary. Summaries keep conclusions and drop bookkeeping. These are the things that reliably disappear, and every one of them is load-bearing in finance work: | Lost | Why it hurts | |---|---| | **Thresholds you stated but did not write down** | You restate them after seeing results — which inverts the whole point of `tie-out` | | **Exclusions you applied** | "Excluded intercompany and voids" becomes "excluded intercompany". The total moves and nothing errors | | **Running totals held in reasoning** | Recomputed over a subtly different set | | **The already-checked list** | "I've done accounts 4000–4200" is e