Skip to content

False-Positive Review Board

Review board — instantiates Signal Value Preservation

A standing body that reviews signals that fired wrongly, attributes each false alarm to its source, and feeds the pattern back so issuers bear the cost of crying wolf.

The False-Positive Review Board is a recurring human forum that examines confirmed false positives, assigns each one to the issuer or rule that produced it, and turns the pattern into consequences and fixes. The one idea that makes it this mechanism is that it supplies accountability — a place where crying wolf has a name attached — rather than a number. Where a dashboard counts false alarms, the board adjudicates individual cases and attaches an owner and a cost, closing the loop that metrics alone leave open. It exists because a signal devalues fastest when issuing a weak one is free; the board makes it not free.

Example

A large platform removes user posts through a mix of automated classifiers and human moderators, and some of those takedowns are wrong. Each wrongful removal quietly teaches users that a takedown no longer means a real violation — the signal erodes. A false-positive review board convenes on a cadence to sample reversed and appealed takedowns, adjudicate whether each was truly unjustified, attribute it to the specific rule, model version, or team that issued it, and require the worst-offending sources to fix or retire their trigger. External review bodies such as Meta's Oversight Board play a related role for high-profile cases. The effect is that issuers begin to internalise the cost of over-flagging, and "removed" regains its meaning because someone, somewhere, now has to answer for the removals that shouldn't have happened.

How it works

On a set cadence the board pulls a sample of firings later confirmed false — plus contested and appealed cases — and adjudicates each: was it genuinely unjustified, and if so, which source produced it? Each adjudicated case is recorded in the false-alarm ledger with an owner attached. Persistent offenders face graduated consequences — a mandatory rule fix, loss of issuing privileges, or referral onward for revocation. The board is the accountability loop that the raw metrics feed but cannot themselves close.

Tuning parameters

  • Sampling versus full review — review every false positive or a representative sample. Full review is thorough but doesn't scale; sampling scales but can miss a concentrated pocket of harm.
  • Attribution granularity — whether a false alarm is blamed on a team, a rule, or a specific model version. Finer attribution drives sharper fixes but invites more disputes.
  • Consequence strength — from advisory feedback to suspending an issuer's privileges. Stronger consequences deter over-flagging but can push issuers to under-report their own errors.
  • Independence — an internal panel versus an arms-length body. Independence resists capture but loses context; the trade-off sharpens as the stakes of the signal rise.

When it helps, and when it misleads

Its strength is that it puts a name and a cost on false alarms — the one thing metrics alone never do — and issuers reliably change behaviour once someone reviews their misses. Its failure modes are that a board that only hunts false positives pushes issuers to under-fire and miss real cases (the mirror error), and that after-the-fact review is slow and can curdle into a blame ritual. It is captured when the reviewed parties staff it, and misused to punish individuals for what are really systemic problems. The discipline that guards against this is to track misses alongside false alarms so the board doesn't optimise one error into the other, keep membership independent of the issuers under review, and aim consequences at rules and incentives before individuals.[1]

How it implements the components

  • issuer_accountability_hook — its core: it attributes each false alarm to a source and attaches a consequence, so issuing a weak signal is no longer cost-free.
  • false_alarm_and_miss_ledger — it adjudicates and owns the entries this ledger holds, deciding what truly counts as a false alarm and who bears it.

It does not compute precision or base-rate trends (signal_base_rate_monitor) — that's Base-Rate and Precision Dashboard, which it consumes — nor execute the actual credential removal (issuance_authority_boundary) — that's Credential Revocation or Downgrade.

Notes

The board reviews false alarms, but a board that watches only false alarms will quietly push its issuers into the opposite error — under-firing until real cases are missed. Reviewing misses alongside false positives is therefore not an optional refinement but a condition of the board doing its job; without it, the mechanism trades a visible error for an invisible one.

References

[1] Standing review bodies for wrongful signals are real and varied, from internal quality-assurance panels to arms-length bodies such as Meta's Oversight Board, which reviews contested content-moderation decisions. Named here only as a real instance of the pattern — the mechanism is the accountability loop itself, not any one body. ↩