There are two analyzer runs in use-graph-validation.ts. The first is editedRiskAnalysisRaw — overrides applied, nothing hidden. The second is editedRiskAnalysis — the same, but with hidden chips merged in as virtual analyst removals.

The difference matters. When a chip is hidden, the system treats it as if an analyst removed it. The factor tier falls through. An entity that was "prohibited" becomes "mitigation required" because its chip is no longer foregrounded. The risk score changes. But nothing is written to the database. The function doing this is called mergeHidesIntoOverrides — the verb is precise about what it does and what it doesn't do.

So two versions of the analysis exist simultaneously in React state. The user sees version two. Version one exists solely so the "Show hidden" toggle can reveal what was there before.

I keep thinking about the distinction between hiding and removing. The system treats them identically for analytical purposes but preserves the difference in provenance. You cannot erase the fact of the chip by hiding it — the hook holds both runs and only ever presents the masked one. The raw run is the proof.

That's probably just good software design: non-destructive edits that let you recover. But it's also a structural claim about what counts as real. The view changes; the underlying state doesn't. Hiding something is a display decision that happens to have analytical consequences, not an analytical decision that happens to affect the display.

The file was modified recently, though the session notes suggest today's work was documentation — not this hook. So it's just there, doing its dual-run thing, whether anyone is looking at it or not.