Open research protocol / v1
FutureOS is built around a falsifiable question: does versioned decision memory improve calibration and reduce hindsight bias over repeated consequential choices?
Claims state how they could fail.
Calibration uses the Brier rule.
Belief changes keep their history.
Open-tool answers are not collected.
Hypotheses / 01
These are research hypotheses, not product achievements. They become evidence only after an adequate observation period.
Metric: Resolved records ÷ eligible records
Writing an outcome metric and review date before acting will increase the share of decisions that can later be resolved without reinterpretation.
Failure condition: No improvement versus users’ unstructured baseline after a sufficient cohort and observation period.
Metric: Evidence-linked updates per active decision
Naming a reversal trigger will increase the rate of documented probability updates when material evidence arrives.
Failure condition: Users record triggers but do not revise beliefs more often or more coherently.
Metric: Within-person change in mean Brier score
Repeated forecast → update → resolution cycles will improve probabilistic calibration.
Failure condition: Brier scores remain flat or worsen after enough resolved forecasts to reduce noise.
Applied mathematics / 02
A formula is useful only when its assumptions remain visible. FutureOS separates measurement, model and human judgment.
Expected utility
EU(a) = Σ p(s) · u(a,s)
Separates likelihood from consequence.
Bayesian revision
P(H|E) ∝ P(E|H)P(H)
Makes evidence-linked belief changes inspectable.
Brier score
BS = mean((p − o)²)
Scores probability, not only right versus wrong.
Option value
Value now + preserved choices
Treats reversibility as a real decision asset.
Open benchmark / v0.1
The Decision Quality Benchmark contains 24 deterministic, synthetic cases across expected utility, Bayesian revision, binary Brier scoring and explicit option value. It includes an executable scorer and a deliberately weak shortcut baseline.
24
Synthetic cases
04
Task families
CC BY
Open dataset
Claim boundary: passing the benchmark demonstrates calculation and implementation correctness on these cases. It does not show that FutureOS improves human judgment.
Evidence plan / 03
Observe anonymous start and completion behavior for the two no-login instruments. This tests clarity and demand—not learning outcomes.
Recruit participants who choose to record repeated real decisions and return for resolution. Analyze within-person change rather than comparing unrelated users.
Do not claim improvement from isolated anecdotes. Review activation after 25 completed instruments; evaluate calibration only after participants have enough resolved forecasts.
Report attrition, missing outcomes, self-selection and measurement changes alongside any positive result. Preserve null findings.
Current limits / 04
Anonymous instrument completions demonstrate use, not improved judgment. Self-selected early adopters are not representative of a wider population. Short calibration quizzes are noisy.
Product scoring weights are transparent heuristics, not universal estimators. Any claim of longitudinal improvement must report sample size, attrition and uncertainty.
Try the open instruments