Automate investment research
Automate the research, keep the audit trail and the exceptions
Automation earns trust by showing what it could not do. Each run below lists its input, the sources it retrieved, the output it produced, and then the exceptions it raised: contradictions between sources, stale evidence, missing conditions, unresolved entities, and anything held for human review. A run with no exceptions is rare and is not treated as a better run.
Runs
Input
Gold edges lower around $4,370–4,380 an ounce as stronger dollar and profit taking offset geopolitical support from Middle East conflict
GEOPOLITICAL · event 2026-09-21 12:27Z
Generated output
The modest pullback in gold despite ongoing US–Iran risks suggests positioning was stretched, but conflict headlines keep the metal’s safe‑haven bid in play.
Instruments resolved
GC=F · DXY
Confidence 0 · dislocation 0
Human review state
Held for review - below the auto-publish confidence threshold.
Exceptions raised by the automation
- Missing evidence. No invalidation condition could be extracted from the available sources.
- Incomplete mechanism. Fewer than two explicit causal steps were recoverable; the chain stops early.
- Held for review. Model confidence is below the auto-publish threshold; a human must approve.
Methodology
Exception derivation. Exceptions are computed from the record, not authored. A run missing an invalidation condition raises one whether or not anyone looks at it, so this page cannot quietly become a highlight reel.
Contradiction handling. When sources disagree on a material fact, the run records both and refuses to pick, rather than averaging them into a confident sentence.
Confidence thresholds. Runs below the auto-publish threshold are held for human review. The threshold is a versioned operating value, not a claim about correctness.
Limits. Automation compresses the retrieval and structuring work. It does not replace judgement about whether the mechanism is the right one.