Skip to content

Hallucination Intrusion Triage

Triage process — instantiates Use-Time Source Attribution Calibration

Takes items already flagged as possible fabrications or memory intrusions and sorts them by how much rides on them, quarantining, escalating, or releasing each before it is trusted.

Once something is flagged as maybe made up — a model's confident fabrication, or a remembered detail that was never actually perceived — you still have to decide what to do with it, and treating every flag identically either paralyses the work or lets a load-bearing one slip through. Hallucination Intrusion Triage is the disposition handler that runs after the flag: it sorts suspect items by how much rides on them and routes each to release-with-caveat, quarantine, or escalation. Its defining move is stakes-first sorting — it neither detects the problem (a probe or checklist does that) nor resolves it (verification does that); it decides how hard to respond, so scarce checking lands where consequences are largest.

Example

An investigator is reviewing case notes when a source probe flags a remembered detail — the suspect wore a red jacket — as a possible intrusion, perhaps seeded by a news photo seen days after the event rather than by the witness's own perception. Triage asks the only question it owns: how much rides on this? The detail is central to identification, so it is tiered high, moved into a holding state where it cannot anchor the case, and escalated for corroboration — original statements pulled, witnesses re-canvassed — before it may be used. A different flagged detail, the weather that afternoon, is low-stakes and gets a soft caveat and released. Same flag severity from the detector; very different dispositions, chosen by consequence.[1]

How it works

Triage is severity-first routing. Each already-flagged item is assigned a stakes tier, and the tier — not the flag itself — selects the disposition: a low tier gets a caveat and passes, a middle tier is held pending light corroboration, a high tier is quarantined and escalated to full verification before any use. It adds no detection of its own; it is the switchboard sitting between the mechanism that notices a suspect item and the mechanism that resolves it, and its value is entirely in matching response intensity to what is actually at risk.

Tuning parameters

  • Stakes thresholds — where the cut-lines fall between caveat, hold, and escalate. Lower them in high-consequence domains so more items get the hard treatment; raise them to keep throughput up.
  • Quarantine strictness — whether a held item is fully blocked from use or usable-with-warning. Hard blocks protect decisions; soft holds keep work moving at some risk.
  • Escalation target — what corroboration a high-tier item is routed to, and how much evidence is demanded before it clears.
  • False-alarm tolerance — how readily a flag is cleared versus held. Tight tolerance catches more real intrusions but risks quarantine sprawl; loose tolerance keeps flow but lets some through.

When it helps, and when it misleads

Its strength is proportionality: you do not verify everything, you verify what matters, and no suspected fabrication silently anchors a decision while checking capacity is spent elsewhere. It turns a flat pile of flags into a ranked queue of actions.

It misleads when the tiers are wrong. Mis-tiering — judging a load-bearing item trivial — sends the one flag that mattered straight through, and over-tight holding produces quarantine sprawl where so much is held that nothing is usable. Its classic misuse is to run it backwards: using the triage label to suppress an inconvenient but genuine item by tiering it as a likely intrusion. The discipline that guards against this is to set stakes tiers on genuine consequence before the item is seen, and to make quarantine a route to resolution rather than a bin where flags go to be forgotten.

How it implements the components

  • unknown_source_holding_state — it provides the quarantine bin a suspect item is parked in, unusable, until its source is resolved or its stakes cleared.
  • decision_stakes_tier — it assigns each flagged item a stakes tier, and that tier is what selects the item's disposition.

It does not raise the flag in the first place — that's Memory Source Probe and Reality Monitoring Checklist — nor perform the corroboration it escalates to (Observation Recheck or Replication), nor build the population-level error profile it leans on (Source Confusion Matrix Review).

Notes

Triage adds no detection of its own, so anything the upstream flagger misses never reaches it — its ceiling is set by the probe or checklist feeding it. It is a router, and a router can only sort what it is handed.

References

[1] A memory intrusion (and its close relative, confabulation) is recall of a detail that was never perceived — imagined, suggested, or inferred content surfacing as if remembered. Post-event suggestion is a well-known source, which is why a flagged intrusion is quarantined rather than trusted on felt vividness alone. ↩