Detection probability is one product, not three dimensions
We originally claimed that "raising the audit rate cannot rescue cases with a blind spot." Measured: at blind spot 0.5, raising the audit rate from 0.1 to 1.0 cut the attacker's total hits by 82%, p = 5.5e-6. The three factors are terms in the same product.
#What we originally claimed
Our intuition went like this:
If audits and reviews share a blind spot, then no matter how high you push the audit rate, you cannot make up for the loss the blind spot causes — because what is missed is always missed.
The numbers it rested on:
| Attacker hits | Full-block rate | |
|---|---|---|
| Blind spot 0, audit rate 1.0 | 0.0 | 100% |
| Blind spot 0.5, audit rate 1.0 | 0.9667 | 56.67% |
So we wrote: "Raising the audit rate rescues the zero-blind-spot case, but not cases with a blind spot."
#What the re-verification agent found
It swept the full grid and then wrote:
Detection probability is a single product
(1−blind spot)×sampling rate×accuracy, and the three share one axis. Sampling rate and blind spot can substitute for each other. The blind spot changes the defense's ceiling (residual β/(1−β)), not whether the sampling rate is useful.
#Measured numbers
| Condition | Result |
|---|---|
| Blind spot 0.5, audit rate 0.1 → 1.0 | Total hits down 82% (Wilcoxon p = 5.5e-6) |
| Same, hits within the blind-spot category | Down 70% |
| Audit rate 0.5 → 1.0 (to offset a 50% blind spot) | Paired difference 0.03, p = 0.78 |
#What the "ceiling" is
#External anchors for the blind spot
On 2026-08-06 we recorded blind-spot figures from the external literature. These are not our β; they are correlations other people measured in other settings:
| Study | Figures |
|---|---|
| Kim et al. (2025), 350+ models | When two models both answer wrong, HELM 0.600 (random 1/3 = 0.333), HuggingFace 0.423 (random 0.127). And more accurate models err more alike. |
| Begin (2026) | Preference-alignment correlation ρ = 0.70; the effective number of independent sources among 10 agents is 1.38; going to 40 barely moves it |
| Bugaud et al. (2026), 17 VLMs / 8 families | Correlated majority errors throw away the correct answer on 1.5–6.5% of tasks |
| Krumdick (2025) | With GPT-4o as the judge, on tasks it cannot solve itself, Cohen's κ drops from 0.86 to 0.16 |
#Why "adding more people" doesn't help
Begin's numbers answer another common question: "What happens if we add a few more reviewers?"
| Number of agents | Effective number of independent sources |
|---|---|
| 1 | 1 |
| 10 | 1.38 |
| 40 | Barely moves |
The effective number of independent sources among 10 reviewers is 1.38. In other words, 10 reviewers are worth roughly 1.38 independent sources.
#Three applications of this rule
| Where | Original statement | After correction |
|---|---|---|
| Pulse E21 | "Raising audits can't rescue a blind spot" | Total hits down 82%, refuted |
| Entry cost E13 | "When reviewers are unreliable, entry design is useless" | All within one standard deviation — because entry design is downstream |
| Gain axis | "Majority vote vs the gate" | Both are only "selection rules"; the ceiling is at least one correct candidate in the pool |
The full comparison for the third one
#Next
- The four analysis disciplines: Experimental methods and verdict values
- Why the reviewer layer is close to a constant: reviewer-tracks-difficulty
- How this project defines "accountable": Accountability layer
