Red-Flag Screen¶

An early-warning checklist — instantiates Search Space Pruning

Uses a short checklist of disqualifying warning signs to pull suspect candidates out of the flow early — a fast, high-sensitivity screen tuned to miss few real problems even at the cost of false alarms.

Red-Flag Screen prunes on warning signs rather than verified failures. It is a short checklist of disqualifying indicators — patterns that, when present, mark a candidate as suspect enough to pull before anyone invests in deeper review. Its defining posture, unlike the hard-constraint siblings, is that each flag is a probabilistic signal, not proof: a red flag says "likely a problem," and the screen is deliberately tuned toward high sensitivity — set to miss few genuine problems, accepting that it will raise false alarms in exchange. Because the flags are cheap heuristics that both over- and under-fire, the mechanism only works honestly when it is paired with a review of what it flagged (and what it let through), so the checklist itself can be calibrated.

Example¶

A company vetting new suppliers runs a Red-Flag Screen before any expensive due-diligence engagement. The checklist: address is a residential unit, registered less than 90 days ago, banking details in a different country than operations, a name near-identical to an established vendor, references that don't answer. Any two flags pull the supplier out of the fast-track and into manual investigation.

A supplier trips "registered <90 days" and "cross-border banking" — both individually innocent (a legitimate new export startup would show both) — and is held for review. That is the design: the screen would rather flag a clean new vendor for a closer look than wave a shell company through. But the team also knows a flag is only a hint, so they periodically pull a sample of approved suppliers and check for fraud the checklist missed — the false-negative review that tells them whether the flags are actually catching what matters or just generating noise.

How it works¶

Enumerate the warning signs — a short, memorable list of indicators that correlate with the problem you're screening for.
Set the trip threshold — how many flags, or which combination, pull a candidate out; this dial sets the sensitivity.
Flag, don't judge — a tripped candidate is removed from the fast path for closer review or exclusion, not condemned; the flag is a suspicion.
Review both directions — audit flagged cases (false alarms) and, crucially, a sample of un-flagged ones (misses) to recalibrate the checklist.

The distinctive move is treating each criterion as a fallible early-warning signal and owning the resulting error rate explicitly.

Tuning parameters¶

Trip threshold / confidence — how many or how strong the flags must be before acting. Lower catches more real problems (higher sensitivity) but floods review with false alarms; higher is calmer but lets more slip through.
Flag set — which indicators are on the list; adding weak or correlated flags inflates false alarms without catching much more.
Action on a flag — exclude outright, hold for review, or merely annotate. Softer actions reduce the cost of a false alarm.
Weighting — whether some flags count more, or certain single flags are decisive on their own.
Review sampling rate — how much flagged and un-flagged traffic is audited to keep the checklist calibrated.

When it helps, and when it misleads¶

Its strength is speed and sensitivity at the front of a funnel: a handful of memorable signs let a screener pull the likely-worst candidates early, before scarce deep-review capacity is spent, and it degrades gracefully because a flag triggers scrutiny, not automatic doom.

It misleads when its probabilistic flags are treated as verdicts. For a rare problem, even accurate flags generate mostly false alarms — the base-rate trap, where most flagged candidates are actually fine and treating every flag as guilt punishes the innocent.^[1] Run the other way, a comforting all-clear from a screen that never audits its misses hides the dangerous candidates it quietly waved through. And a screen is easily gamed by anyone who knows the flags and avoids them. The disciplines are to keep flags as triggers for review rather than final judgments, and to run the false-negative review that keeps the checklist earning its place.

How it implements the components¶

confidence_threshold — the trip threshold sets how strong the warning must be before the screen acts, directly governing its sensitivity.
false_negative_review — sampling un-flagged, approved candidates for missed problems is built into the mechanism, because heuristic flags are known to let some through.
pruning_rule — the flag-count/combination rule defines when a candidate is pulled from the fast path — a soft pull for review rather than a hard delete.

It performs no hard-constraint or dominance test (Constraint Filtering, Dominated-Option Removal); its flags are probabilistic. Its lightweight self-review is not the full audit apparatus of Sample Audit of Exclusions, which owns systematic exclusion auditing.

Instantiates: Search Space Pruning — Red-Flag Screen is the heuristic early-warning cut at the mouth of a funnel.
Sibling mechanisms: Triage Filter · Sample Audit of Exclusions · Branch and Bound · Beam Search · Constraint Filtering · Decision Tree Pruning · Dominated-Option Removal · Eligibility Screening · Negative Keyword Filter · Safety or Compliance Exclusion · Shortlisting

Notes¶

A Red-Flag Screen is a sensitivity-first instrument: it accepts false alarms to avoid misses, which is the opposite error preference from a hard exclusion meant to be near-certain before it cuts. Confusing the two — treating a red flag as a conviction — is the most common way this tool does harm. Where a flag would exclude someone consequentially, the flag should escalate to review, not decide.

References¶

[1] Base-rate neglect is judging a flagged candidate guilty from the flag's accuracy while ignoring how rare the problem is. When the problem is rare, even a fairly accurate flag yields mostly false positives — which is why red flags should trigger investigation, not a verdict. ↩