In development · January 2027

The metric moved.
Was it really the launch?

RL environments that train data-analysis agents to check the story, not repeat it.

observed
without the event
true effect, known by construction
01Task

A simulated company. A claim to check.

01

Brief

A KPI changed. A memo blames one event.

02

Investigate

A sandboxed company database. No network.

postgressqlpython
03

Answer

Structured, not prose.

causesclaimscounterfactuals
04

Graded

Against the simulated truth and the query log.

02Grading

Every world is run twice. No judge models.

Twin worlds

With and without each event. Decoys leave the same footprint and do nothing.

Asymmetric reward

Confident and wrong costs the most.

Justified conclusion reward
Honest abstention small cost
Unchecked guess partial credit
Endorsing a false claim large penalty
03Quality

A bank ships only if no shortcut pays.

0
LLM judges in the reward
100%
reference score on every task
<0
shortcut reward, upper 95% bound, every level
1:1
evidence pack per bank, tied by digest
Empty answerscores below zero
Scripted shortcutsmean ≤ −0.05
Learned attack on visible featuressame bar
Selection and information checkscapped priors
Leak and tamper checkspass
Determinismsame answer, same reward
04Status

Gated before every release.

NOV 2026

Grader, simulator and gates built. First gated bank.

DEC 2026

Frontier pass rates per level. Training-transfer test.

JAN 2027

First tasks available.

05Contact

Spec, samples and evidence pack on request.

contact@guananira.org
Harbor tasks · HUD tasksets · synthetic data only
Licensed for training and evaluation · held-out banks never sold