DECLASSIFIED · NRG / DFA / EX-26.06 · READING COPY · DISTRIBUTION C
METHOD · Numbers, bias, self-correction EX-26.06 · Standing reference
Method · How the federation handles being wrong

Numbers, bias, and self-correction

Most of what makes this exercise unusual isn't the scenario. It's the set of rules the federation follows to keep itself honest. Every estimate is a range. Every specialist has to declare their own known blind spots. Bias warnings get carried forward turn after turn, and the federation is required to actually move its estimate when a bias is identified — not just write the warning down. There's also a protocol item (added at turn six) that asks the federation to audit its own framing once per turn: are we still tracking the right variables, or have we gotten attached to questions that no longer matter?

The format: [low, best guess, high]

Three numbers, not one. The middle is the best guess. The outer two are the realistic floor and ceiling. Wider gap means more uncertainty. The bar visualizes the range; the marker shows the point estimate.

Narrow range

US–Israel formal alliance, T0 baseline
65% 78% 88%

Lots of historical data; the model fits the situation well.

Wide range

Civilian fatalities during the turn-5 interval
180 980 3,800

Right-tail risk from cluster events the federation can't model precisely. The honest answer is "somewhere in a big range."

Three different kinds of doubt, tracked separately

Confidence isn't one thing. The federation tracks three flavors on every estimate:

The federation failure mode

The thing the protocol guards against most carefully: when multiple specialists share the same blind spot, the whole federation can do worse than a single skeptical analyst would. If three different specialists are all biased upward on US institutional resilience for the same structural reason, the federation will produce three independent-looking estimates that are all wrong in the same direction.

The bias the federation never moved

Upward bias on US institutional resilience has been flagged from turn zero onward. Across seven turns the federation widened its uncertainty ranges in the bias direction but didn't actually move the central estimate. At turn six, CASS pointed out that this is exactly the decorative-warning failure mode the protocol is designed to catch. The bias was still active at turn seven's close.

The modal-hold review (new at turn six)

A protocol item that asks once a turn: which variables are we still tracking just because they were load-bearing two turns ago and nobody re-checked? Which of our framing assumptions are still shaping our estimates even though we wrote down a warning we never acted on? Which of our headline questions have become irrelevant to what's actually happening?

What it has produced so far:

"Falsifies-on" hooks (new at turn seven)

Starting at turn seven, every major estimate carries a list of conditions that would force the federation to re-estimate. Example, for "the president stays in office through January 2029":

Each hook also has a half-life: even if nothing visibly happens, the estimate has to be re-elicited after some interval. The point is to catch the drift that happens when nothing visibly moves but the underlying conditions have shifted anyway.

The federation's own report card, every turn

Every briefing closes with a section called "federation failure mode check." It carries forward the biases flagged in earlier turns, marks which have been addressed and which are still active, and adds any new ones. At turn seven's close, the federation added two new failure modes for itself to watch: (1) the consolidation of four specialists into one combined cascade agent at turn seven may have hidden disagreements that would otherwise have surfaced as cross-specialist seams; (2) CASS's acceptance rate may indicate the federation has become too dependent on her for self-correction.