When will coding agents solve private professional tasks? Will robots execute reliable multi-object work? Will frontier-AI rules be enforced? Five original research candidates connect ambitious AI forecasts to observable events. Two now have separately registered official numerical vintages.
Research-draft release, 27 October 2026: five original research candidates, with two separately approved numerical forecasts (CM-003 20% and CM-004 15%) now registered; zero scored outcomes. No claim of existing calibrated forecasting intelligence.
Evidence Pulse 2.0 checks the published Cortist News index when JavaScript is available. The source-linked, dated research articles remain accessible through Cortist News.
For the EU fine and Waymo Munich milestones, we have clarified the exact decision evidence, event window and reasons certain announcements or permits do not qualify. Both still require independent verification before a forecast can be locked.
Review the Article 101 rule · Review the Munich paid-ride ruleEvery candidate has been checked against its official sources. Five important definition gates remain open; none is ready for a formal probability or scoring.
Read all five source audits and the next verification stepsEach candidate links to an original benchmark, legal authority or operator statement. Every page spells out what would count as “yes”, what would not count, and what must be clarified before assigning a probability.
Resolution target: SWE-bench Pro V2 private benchmark: at least 90% audited pass@1
Strongest because withheld tasks reduce contamination risk; not a test of general intelligence.
Examine rules, source evidence and exclusions →Resolution target: Real-robot LADDER five-object sequences: at least 50% complete
Direct physical reliability test; definition not yet ready to score.
Examine rules, source evidence and exclusions →Resolution target: At least one formally adopted fine decision under Article 101
Enforcement is more verifiable than policy promises.
Examine rules, source evidence and exclusions →Official 20% forecast and frozen evidence rulesResolution target: First publicly available paid driverless ride within Munich
Commercial deployment, not a prototype announcement.
Examine rules, source evidence and exclusions →Official 15% forecast and frozen evidence rulesResolution target: MLPerf Endpoints: at least 2x GPT-OSS-120B standardized throughput per watt
Chips and inference efficiency need full-system like-for-like measurement.
Examine rules, source evidence and exclusions →New, dated accountability review: See all eight published provisional forecast drafts and the exact quality blockers for all five Milestones. No outcome or Brier score has been verified yet.
This is the strategic foundation for later calibration and accuracy measurement, not an artificial quota of predictions.
Exclude already achieved or ambiguous outcomes. Freeze benchmark splits, source versions, inclusions and exclusions.
Define the authoritative public result and what to do if a benchmark or official source changes.
Publish an externally verifiable timestamp and one explicit probability. Preserve every later revision as a new version.
We do not use a saturated public software leaderboard as an apparently difficult target. Nor will we pre-register a score on OSWorld 2.1 without accounting for its independently documented scoring problems. Reliable forecasting requires stable measurement before attractive numbers.
Short-horizon binary AI research questions, designed to be resolved against clearly specified real-world evidence. These first five are public candidate definitions and are not yet numerical predictions.
CM-003 and CM-004 now have separately approved 20% and 15% v1 original numeric forecasts. None has an outcome yet. The three other candidates remain unapproved.
Cortist must prospectively register exact outcome criteria, a deadline, a publisher-approved frozen probability vintage and then publish evidence-based event adjudication. Only then can a proper scoring rule such as the Brier score apply.
Machine-readable, dated candidate register (v1) · Questions last researched 27 October 2026. Event conditions and cutoffs are not prospectively locked forecasts yet.
