Calibration tests whether probabilities deserve trust. Decomposition makes the thesis inspectable. Review detection identifies which changes require attention.
Four models forecast fixed questions independently. Reads are timestamped, preserved, and scored against named resolution sources where outcomes exist.
A thesis is represented as governed assumptions, mechanisms, relationships, drivers, and pathways. Structure makes the view inspectable without claiming that every relationship is causal.
Deterministic detectors scan movement, disagreement, concentration, conflict, and staleness to identify what may require research, discussion, or renewed review.
Model estimates update daily. Weekly reviews freeze the available state. Resolutions update calibration. Structural changes require explicit human promotion.
Probabilities, model spread, pressure, trajectories, candidate findings, suggested actions, and resolution branches.
An immutable review snapshot records the thesis state that was actually available during the review period.
Resolved outcomes update Brier scores, calibration curves, directional accuracy, and sample sizes.
New assumptions, retired assumptions, taxonomy edits, and relationship changes require explicit human promotion and a new structure version.
Crene evaluates the latest thesis state against preset detector rules. The public language is constrained to what the evidence actually establishes.
Several important assumptions move materially in the same direction during the review window.
An assumption important to the decision moves quickly while the model ensemble remains materially divided.
The headline spread appears tighter than disagreement across important assumptions underneath it.
Support and resistance move inside the same mechanism, including asymmetric tilts.
A disproportionate share of weighted support comes from one mechanism or narrative cluster.
Underlying conditions move materially while the anchor remains comparatively unchanged.
Findings are ranked by detector score, decision weight, movement, spread, and mechanism coverage. Type caps prevent one detector from filling the entire public set. The selected findings can therefore change from one daily run to the next as the thesis state changes.
Suggested actions are constrained to research, review, escalation, and thesis maintenance. They are not portfolio instructions.
A detector establishes an observable pattern in the thesis map.
A constrained rule proposes a research or review step tied to the detected pattern.
Plausible branches explain how the thesis state would change under different resolutions.
Each branch identifies the evidence that would make that resolution more consistent with the observed record.
The action layer does not recommend buying, selling, hedging, or position sizing. Portfolio decisions require mandate, exposure, liquidity, risk, and correlation context that the public thesis page does not possess.
Daily reads are timestamped when polled. Weekly reviews freeze those reads against the structure version that was live at the time.
A completed weekly record preserves consensus, available model reads, spread, movement, freshness, and the active structure reference.
Promoted assumption, taxonomy, pathway, or relationship changes create a new structure version with a changelog.
A native weekly freeze records what the system said during that week. It does not resample or substitute a later answer.
These figures are loaded from the live analytics layer. Counts and calibration samples change as events are added, archived, or resolved.
8 categories. Active binary anchors and assumptions are repolled daily.
Retired from polling, with timestamped histories preserved.
Resolved against named sources and included in the scoring layer.
Claude, GPT, Gemini, and Grok are polled without seeing one another.
A useful review finding does not prove predictive skill. A good Brier score does not prove that the thesis workflow changes decisions. Crene reports the populations separately because they answer different questions.
| Metric | Current value | Population | Purpose |
|---|---|---|---|
| Leakage controlled consensus | 0.114 | n=811 | Tests ensemble calibration against a reported base rate Brier. Reported skill: modest. |
| Macro ex earnings | , | n=, | Supplementary evidence for a thinner macro population. Reported separately from the leakage controlled benchmark. |
| High conviction calls | strong on high conviction calls | n=reported | Supplementary evidence for forecasts where the ensemble expresses stronger probability separation. |
| Contested calls | near chance | n=reported | Supplementary evidence for forecasts where probabilities remain near the difficult middle. |
Timestamped probabilities, resolved outcomes, Brier scores, calibration bins, movement, model spread, detector inputs, and frozen review records.
Assumption maps, mechanisms, pathways, driver families, authored relationships, ontology fields, and editorial taxonomies.
Whether daily movement contains incremental decision value, whether detector findings improve outcomes, and whether disagreement across models predicts realized uncertainty.
The resolved event scoring record does not automatically validate these maps. Their immediate value is making uncertainty inspectable, contestable, and eventually scorable.
A binary anchor is decomposed into governed, falsifiable assumptions grouped by mechanisms or categories.
A continuous anchor is represented as model percentile distributions and a driver matrix.
A strategic question is represented as multiple internally coherent pathways rather than one compressed probability.
Inspect a live thesis, review the resolved corpus, or see the broader map of scenarios and factors.