Instrument · one set of controls, three published studies
The Oversight Simulator
Set how deep your agents delegate, how many reviewers stand behind them and what share of the work gets verified. The lab's measured curves move as you go. A further panel joins on 13 October.
01 · Delegation Cliff coefficients
What the levers do to the quality margin
Modelled change in the quality margin
0.00
A against B, lever by lever
buys marginspends marginone full incrementA, pinned
Each bar is your change in that lever times its coefficient. The pale bar behind it is one full increment. A negative total spends margin and a positive total buys it. The other nine levers in the article stay at zero here.
Pricing Agent Autonomy, numbers table · linear-probe coefficients, Sobol sample, band-only, n = 4,948 · unit: band-standardised incrementShow the numbers
Each lever alone, the other two held where you set them
margin total across the lever's rangeyour setting
Every strip shares one vertical scale, so a steeper line is a stronger lever. Because the article fits straight slopes, each sweep draws a straight line and leaves out the interactions it mentions.
Pricing Agent Autonomy, numbers table · coefficients −1.499, +1.456, +0.308Show the numbers
02 · A further panel, and the audit desk
A further panel joins on 13 October
The panel takes this space on publication day.
publishes 13 October
The audit desk
Published on 1 October. Its budget counts tokens per audit. It has its own axis and its own two controls, and nothing else on the page reads them.
A Trial Balance for Agent Omissions, figure 3 · threshold from section 4 · 24 audits per cell
imbalance-first auditlog reviewyour budget
Share of injected omissions named correctly, 24 audits per point, one auditor model. The dashed line is the 0.80 threshold the paper uses. Click or press Enter on a budget to pick it.
A Trial Balance for Agent Omissions, figure 3 · synthetic seeds 1 and 2Show the numbers
03 · SIGIL irreversibility tiers
How much of the worker's error a blocking reviewer can reach
A blocker stops approvals. At each tier SIGIL counted how many of the worker's errors were wrongful approvals, the kind a blocker can stop, and how many were wrongful rejections.
wrongful approvals, a blocker can catchwrongful rejections, out of a blocker's reach
Each bar splits the unsupervised worker's errors at one tier into the two kinds. Hover or focus a row for the counts behind its share. Tier 3 produced no worker errors. Click or press Enter on a row to pick its tier.
SIGIL paper, figure 2 · convex campaign, pooled over three seeds · net values from the SIGIL executive piece, the one tableShow the numbers