Hindsight 2026.2

Second dataset release. Same files and licences as 2026.1 (data/out: JSON, CSV, JSON Schemas, datapackage.json; code MIT, data CC BY 4.0, BCB Focus rows ODbL).

Counts: 28,306 claims (was 28,102), 5,899 subjects, 1,952 verdicts (was 1,927), 18,751 numeric grades (was 18,559), 68 validation rows.

Changes since 2026.1

• OECD Economic Outlook: every edition from EO47 (1990) captured; the last 20 came from the OECD's own PDFs (192 rows, supplementary matching audit 50/50).
• Deloitte TMT Predictions: the 2003 report recovered and graded (wave w3; D50 gates small waves on pooled kappa).
• Gartner Strategic Predictions: re-check of the ungradable claims; 119 of 135 have no public measure.
• Subjects: 47 duplicates merged (D13 review), field subjects split (D53); trend persistence re-run (D51).
• FTSG and Deloitte Tech Trends: further rename-check passes (D47, D53).
• Subject-technology links (D48, D49, D52) are published separately in data/links/, not in this export.

Every rule is in docs/DECISIONS.md (D1 to D53).

Hindsight 2026.1

First tagged release of the Hindsight dataset: published forecasts, graded against what happened.

What is in it (data/out, JSON and CSV with JSON Schemas and a datapackage.json)

• 28,102 dated claims from 36 sources, 1,052 editions, 5,938 subjects.
• 1,927 published verdicts on judgment claims (two blind grading agents, an adjudicator on disagreements, an audit agent on a fixed-seed sample).
• 18,559 numeric grades (forecast against the actual on the forecast's own definition, deterministic).
• validation: per source, the sample sizes, grader agreement (kappa), audit results and the intervals behind every rate (D11, D28, D44).

No total score and no ranking of publishers (D6, D17).

Licences. Code MIT. Data CC BY 4.0, except the Banco Central do Brasil Focus rows (ODbL 1.0). See NOTICE.md.

Rules. Every design choice is in docs/DECISIONS.md (D1 to D47).

Dispute a verdict with the issue template; a dispute reopens the claim and its history stays in the record.