The date or period the evidence describes. Forecast and plan dates remain future source statements, not observations.
OBSERVATORY / SYNCHRONIZED EVIDENCE
Timeline
Five evidence tracks · independent units · historical knowledge states
Date definitions
The source publication date when the source establishes one; null otherwise.
The date Vardark first recorded the admitted source revision used by this view; exact UTC instants preserve multiple reviewed revisions on one day.
The date of a Vardark assessment or editorial review; null for unassessed measurements.
2022—2030
Dates align; values are comparable only within a row.
Compute capacity
H100e, estimated ownership
2022-03-31 = 100 within Compute onlyData-centre electricity
TWh/year, all data centres
100 = first compatible observation in each Energy source revisionAdvanced-node supply
thousand 300 mm wafers/month
2024-12-31 = 100 within Chip Supply onlyAgent task horizon
expert-minutes by reliability level
No cross-release indexAI research automation
benchmark-specific result
No cross-release indexMilestones, methods, benchmarks and facilities
Dated events with linked evidence.
What Vardark knew by 06 Sep 2026
Corrections append a replacement; they do not erase the statement that was previously published.
Agent Horizon shows 4 compatible P50/P80 estimates across 2 public model cohorts under the same TH1.1 contract. They are concurrent model and reliability observations, not a time series; suite coverage and the 16-hour boundary limit generalisability.
Show revision history
Agent Horizon shows two compatible reliability levels, P50 and P80, for the same public frontier cohort. They are concurrent choices of success requirement and do not constitute a time series; suite coverage and the 16-hour boundary limit generalisability.
Agent Horizon shows 4 compatible P50/P80 estimates across 2 public model cohorts under the same TH1.1 contract. They are concurrent model and reliability observations, not a time series; suite coverage and the 16-hour boundary limit generalisability.
RExBench v3 reports that the best result, even with human-written hints, is below 44% on realistic research extensions.
Show revision history
RExBench reports that the best result, even with human-written hints, is below 40% on realistic research extensions.
RExBench v3 reports that the best result, even with human-written hints, is below 44% on realistic research extensions.