Crene is investment thesis review infrastructure. The platform decomposes macro, labor, AI transition, and investment thesis questions into falsifiable components using four frontier AI models (Claude, GPT, Gemini, Grok). The system produces three product types: thesis maps (binary decomposition into assumptions), factors (continuous decomposition into drivers), and scenarios (hybrid decomposition into coherent pathways). Cross model disagreement is the primary signal. Calibrated against 1,161 resolved events with a Brier score of 0.114.

How does AI prediction calibration work at Crene?

Calibration applies at the consensus layer, not at the underlying model layer. Crene queries four frontier models in isolation, computes a cross model median consensus, and tracks resolution against a tiered source allowlist. Every resolved event becomes a permanent calibration point. Per tier accuracy, per model Brier scores, and per domain accuracy are computed live and published at https://crene.com/methodology.

What is a Brier score?

A Brier score is the proper scoring rule for probability forecasts on binary outcomes. It is the mean squared error between the forecast probability and the realized outcome (0 or 1). Lower scores indicate better calibration. The no skill baseline is 0.25 (always predicting 0.5); a perfect forecaster scores 0.0. Crene publishes the consensus Brier score and per model Brier scores across the resolved event corpus.

Crene serves investment teams that need a repeatable thesis review process: hedge funds, asset managers, macro PMs, CIO offices, family offices, and thematic investment teams. A team brings one thesis it is actively debating. Crene maps what the thesis depends on, where models disagree, what changed each week, and what would force a rethink. Data and API access remain available as the proof layer behind the workflow.

What is the difference between an event, a thesis map, and a factor at Crene?

Crene has four product types. Events are base binary outcomes with scalar probabilities. Thesis maps decompose binary anchor questions into child situations. Factor maps decompose continuous variables into drivers. Scenarios map long horizon investment questions into coherent pathways across live systems including Empire by Default, AI Labor Transition, European Rearmament, and India as the Third Growth Pole.

Bring your own thesis

Get this same map built on a thesis your team is debating.

Assumptions, model disagreement, and rethink triggers, updated weekly before your PM, risk, or IC discussion. Currently accepting one macro thesis and one AI-economy thesis for July.

Map one thesis

Back to thesis map Part ofWill AI generated content exceed human generated content on the US internet before 2030?

Event · ACADEMIC & SCIENTIFIC INTEGRITY

Will any US medical journal retract over 20 papers for undisclosed AI content generation in a single year by 2029?

Resolves Dec 31, 2029

Probability

47%

4-model average

Confidence

LOW

building category history

Stability

—

Models

Diverging

40pt spread

Current read

Consensus sits at 47% across the four models and is — (loading). Models are diverging at a 40pt spread, so the average conceals live disagreement worth inspecting below. Category confidence is LOW (building category history).

Confidence reflects category-level track record. Stability tracks estimate movement. Models shows whether the four agree.

Belief trajectory

Loading belief trajectory...

Per-model probabilities

GPT-4o

35%-12 vs avg

The growing integration of AI in research and publishing suggests increasing scrutiny over undisclosed AI contributions. As of 2023, numerous institutions are already implementing guidelines to disclose AI usage in research, which could lead journals to adopt stricter retraction policies if such disclosures are not adhered to. Moreover, in 2022, several high-profile retractions occurred due to issues surrounding authorship and contribution, setting a precedent for future actions against undisclosed AI use.

Key uncertaintyThe pace at which ethical standards and regulatory frameworks are implemented in academic publishing could either mitigate or exacerbate the likelihood of widespread retractions due to undisclosed AI content.

Gemini

60%+13 vs avg

The rapid advancement and adoption of AI in research writing, coupled with increasing scrutiny from publishers and ethical bodies, makes widespread undisclosed AI-generated content a growing concern. The sheer volume of research published annually means that even a small percentage of undisclosed AI use could easily surpass 20 retractions by 2029. As detection tools improve and editorial policies become more stringent, instances of flagged content are likely to increase.

Key uncertaintyThe rate of development and effectiveness of AI detection tools, and the willingness of journals to implement and enforce strict policies around AI authorship.

Claude

28%-19 vs avg

The medical publishing ecosystem is experiencing unprecedented scrutiny of AI usage following high-profile retractions (e.g., Nature's 2023-2024 AI-related retractions, multiple ChatGPT-detected papers), but we lack clear evidence of systematic undisclosed AI generation at scale sufficient for 20+ retractions in a single journal-year. Historical retraction rates for US medical journals average 1-3 papers annually per major journal, so reaching 20+ for a single cause would represent a 6-20x increase above baseline. However, the combination of: (1) increasing AI adoption in manuscript preparation, (2) journals' strengthening detection capabilities, (3) tightening disclosure policies, and (4) 5 years remaining until 2029 for a single egregious case to emerge, makes this plausible but not highly probable. The probability increases if detection improves faster than author compliance, but decreases if journals successfully implement preventive measures and standardize disclosure requirements.

Key uncertaintyThe rate at which journal detection methodologies improve and become standardized—if AI detection becomes reliably automated, discovery of past undisclosed AI could accelerate dramatically, whereas if detection remains labor-intensive and inconsistent, large-scale discovery becomes unlikely.

Grok

68%+21 vs avg

AI text-generation tools have seen exponential adoption since GPT-3.5 (Nov 2022), with 2023-2024 surveys showing 25-40% of biomedical researchers using LLMs for drafting; journals already retracted 12 papers in 2024 solely for undisclosed AI use (e.g., Frontiers, Cureus cases), and the current 3-6 month detection lag plus 18-month average retraction timeline implies at least one high-volume incident by 2028. Structural drivers include COPE’s 2023 AI disclosure mandate, NIH and NSF grant language requiring AI acknowledgment, and the 4-6% annual rise in total retractions (Retraction Watch 2018-2023 data) that amplifies the numerator.

Key uncertaintyWhether journals will tighten disclosure enforcement or shift to pre-submission AI-detection screening before 2027.

Key disagreementGrok (68%) vs Claude (28%): Different weighting of factors

CRENE-AICD-AS09Generated Jul 3, 2026