Point-in-time retrieval
Every observation resolves against the vintage available on the decision date, preventing revised data from leaking into historical judgment.
Independent build · Point-in-Time Decision Intelligence
Area10
A decision should be judged by what was knowable then—not by revised data available now.
39.1Meconomic observations
161,525catalogued series
3,415graded predictions in one audit
The constraint
Most research tools show the latest economic value. That quietly contaminates historical analysis, backtests, and narratives with information that had not yet arrived.
Non-obvious diagnosis
Point-in-time truth is not a chart option. Vintage data, source health, claims, hypotheses, decisions, and later evaluation have to remain connected.
The build starts here. The feature set follows from this diagnosis.What the system makes possible
Preserve the information set, then preserve the reasoning built on top of it.
A financial-research system with vintage-aware retrieval, scheduled ingestion, source drift probes, freshness and incident checks, cited briefs, hypothesis tracking, forecast ledgers, adversarial review, and durable decision records.
Source-health gates
Freshness, drift, revision status, and incidents sit in the decision path rather than in a separate monitoring console nobody checks.
Traceable reasoning
Sources connect to claims, hypotheses, briefs, decisions, and later scores, keeping the path from evidence to action inspectable.
Adversarial review
A dedicated critique layer tests time horizon, evidence quality, counterarguments, and unsupported certainty before conclusions harden.
Evaluation without hindsight
Predictions and decisions can be scored later without rewriting the original information set or rationale.
See it working
Inspect the mechanism, then inspect the evidence.
A conclusion can be traced back to its evidence and information set, then evaluated later without rewriting history. The audited warehouse held 39,103,981 observations across 161,525 catalogued series.
One real revision · three stored vintages
The same December 2023 housing-starts observation changed 7.4% after first publication.
A model using 1,568 to explain a decision made when only 1,460 was available is using future information—even if both values now appear historical.
Verify the source series at FRED ↗︎2024-02-011,460First stored vintage
2024-03-011,562Second stored vintage
2024-12-011,568Latest stored vintage
Actual reasoning dashboard
The operating surface keeps posteriors, market disagreement, extreme signals, discoveries, contradictions, and system health visible as one decision context.
Actual review screen
The adjudication surface exposes a time-horizon mismatch rather than burying disagreement in a summary.
Actual critique workflow
The critique viewer keeps evidence, objections, and revision decisions together instead of collapsing them into an opaque confidence score.
How the parts compound
A feature matters when it improves the next decision.
- 01Ingest
Store observations with release dates, vintages, source state, and revision history.
- 02Constrain
Resolve only the evidence that was available at the decision timestamp.
- 03Reason
Connect sources, claims, hypotheses, and briefs in one traceable chain.
- 04Challenge
Run adversarial review before the system accepts a conclusion.
- 05Evaluate
Score the decision later while preserving what was known and why it was chosen.
The decisions that mattered
Architecture follows the failure mode.
The implementation is the visible surface. These choices determined whether it could solve the underlying problem.
- 01
Freeze the information set
Resolve an observation against the vintage available on the decision date—not the value later revised into the series.
- 02
Gate on source health
Freshness, revision status, incidents, and source drift become prerequisites for confident analysis.
- 03
Preserve the chain
Keep source, claim, hypothesis, briefing, decision, and later score linked into one inspectable record.
Evidence classOperating product + warehouse audit
Source
Product captures dated Apr/May 2026, a live-SQL warehouse audit dated 30 Jul 2026, and vintage rows from the operating database.
What it establishes
The critique and adversarial-review workflows are real; the warehouse scale is measured; and the vintage layer preserves revision history.
Boundary
No investment-performance claim. Product captures establish behavior, not independent validation of every generated conclusion.