Crene keeps an investment question, the evidence behind it and the human decisions it informs connected as the view evolves.
This page explains how the record is structured, how forecasts are evaluated and where the evidence stops. Public measurement results and private workspace track records have separate populations.
Each stage is explicit so the record can change without erasing what came before.
Define the investment question, current view, horizon, assumptions, drivers and observable conditions.
Bring new evidence and changed conditions back against the preserved view.
Make Maintain, Revise or Retire an explicit human decision, with its rationale attached.
Keep prior states, finalized decisions, provenance and version history available for later review.
A defined binary question, a continuous driver and a set of coherent futures need different representations. Their histories can be preserved without treating their outputs as interchangeable.
A binary investment thesis is decomposed into governed, falsifiable assumptions grouped by mechanisms or categories.
An estimate is a probability that the exact saved question resolves Yes. Its definition, target date, resolution criteria and source establish what is being assessed.
A continuous anchor is represented as model percentile distributions and a driver matrix.
Factors use p5, p25, p50, p75 and p95 outputs. Rates, prices, growth and other continuous outcomes require suitable continuous scoring rules.
A strategic question is represented as multiple internally coherent pathways rather than one compressed probability.
Joint state reasoning connects binary and continuous components. Pathway labels and distributions are editorial framings, not probability weights.
Observation, reassessment, human governance and resolution are recorded separately. Viewing saved work does not refresh models. A new assessment has its own review and confirmation; recurring monitoring is controlled separately.
The workflow follows the relevant question and enabled monitoring settings rather than one universal daily schedule. Pausing future updates does not stop requests already started.
The historical research workflow used preset detector rules to identify candidate findings in saved thesis states. These examples explain the method; they do not imply that retired public examples or every private component are being monitored today.
Several important assumptions move materially in the same direction during a review window.
An important assumption 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.
Research and review suggestions are useful when they identify the evidence that would distinguish competing interpretations.
Describe the observable pattern in the saved record.
Tie the pattern to a research, review, escalation or thesis maintenance step.
Explain how the thesis state could change under different resolutions.
Identify the evidence that would make each branch more consistent with the record.
Suggested review steps are not instructions to buy, sell, hedge or size a position. Portfolio decisions also depend on mandate, exposure, liquidity, risk and correlation context.
Timestamped inputs, versioned structure, finalized decisions, and frozen review states help a later reviewer reconstruct what was known, what changed and what the team chose at the time.
Saved review snapshots retain the available estimates, model reads, disagreement, freshness and definition references from that review. They describe the evidence available at that time.
Reviewed changes to definitions, assumptions, pathways and relationships preserve the earlier state and the reason for the change. Model views and the team decision remain distinct.
A later answer is not evidence of what a model said earlier. Historical cadence-specific freezes and saved decision history preserve the original read for comparison.
These figures describe the retained historical corpus. Automated public example updates have been retired. They do not measure the current private workspace forecasts.
Stored binary anchors and assumptions are retained for historical inspection.
Preserved scenario maps with API supplied component and pathway coverage.
Earlier events resolved against named sources and included in the scoring layer.
Historical families and versions. Shared evidence and related questions can introduce dependence.
Public corpus statistics response: Source time unavailable. This is the response time, not a forecast or outcome date. The benchmark and confidence chart below have separate source populations.
Crene's public scoring corpus comes from an earlier phase focused on short horizon events. The snapshot benchmark, broad corpus and confidence chart have different stored populations and selection rules. They do not establish current private workspace accuracy, investment returns or the reliability of a long horizon thesis.
For a binary event, the Brier score is the squared difference between the forecast probability and the outcome: (p − y)², where Yes is 1 and No is 0. Scores range from 0 to 1; lower is better.
Calibration asks a different question: across repeated events forecast at a given probability, how often did they happen? Small samples and related events limit what the comparison can establish.
A constant 50% forecast scores 0.25 for either outcome. That is a reference, not a universal best baseline.
| Metric | Reported Brier | Sample | Population and interpretation |
|---|---|---|---|
| Snapshot consensus · 24h resolution gap | — | n=— | Latest retained estimate per available model at least 24 hours before recorded resolution. Coverage can vary. Same-sample base-rate Brier: —; reported Brier improvement: —. The recorded-resolution gap does not establish when outcome evidence became public or independently rule out information leakage. |
| Macro ex earnings | 0.2305 | n=383 | A separately filtered macro population. Its sample and score cannot be combined with the benchmark or broad corpus. |
Until an eligible outcome exists, the workspace shows Not scored. Inactive or superseded events are excluded from current scores while their history remains preserved. Related events can overstate the apparent amount of independent evidence.
Outcome evidence is an editor attestation, not independent source verification. Numeric outcomes and scenario pathways need different scoring rules. Private workspace results do not enter this historical public summary.
The legacy leaderboard and Brier analytics use different filters and can cover different events. Joining them by a model-family name would imply a shared evaluation cohort that has not been established.
The Data page reports model-family Brier, directional accuracy and sample counts from the canonical historical statistics endpoint. Versions can be combined within a family; the reported identifier is not a current-model benchmark. These rows are not a controlled model ranking.
The legacy model-family row count is not currently available. Availability does not establish that models were independent or used the same event sample.
Inspect reported model-family samples →Timestamped probabilities, resolved outcomes, Brier scores, calibration bins, definition versions, saved model responses, explicit decisions and preserved review history.
Assumption maps, mechanisms, pathways, driver families, authored relationships, ontology fields and editorial taxonomies. Structure is not scored forecast evidence.
Whether recorded changes improve team decisions, whether disagreement identifies useful uncertainty and how performance holds up across future questions, model versions and market regimes.
Start with the view, the decision it affects and the evidence your team already uses. Keep decisions, version history, and preserved review states connected as the question evolves.
Private evaluation workspaces are offered by invitation. A review inquiry does not automatically create an account.