More data than the day will allow
Headlines · Reports · Curves
Status · Live · v1.0
Decision Infrastructure · Goldeneye Capital
Goldeneye
A terminal for analysts who think in distributions, not directions.
Chicago · 202601 · Thesis
The era of certainty is over. The decade ahead belongs to those who can reason in distributions.
Goldeneye Founding Memo
02 · The Problem
The analyst is drowning in inputs and starved for structure.
Signals decay before the analyst integrates them
Seconds to relevance
Models that won't show their work
Black-box forecasts
Yesterday's thesis colors today's read
Bias compounds silently
Positioning data lags conviction
By the time it's clear, it's late
Confidence doesn't track edge
Few analysts measure their own hit-rate
~12×Data sources a commodity analyst is expected to track in real-time
3.2sMedian attention span on any single chart before context-switch
<7%Of analysts measure their own conviction calibration formally
03 · What we built
A terminal designed to make reasoning the product — not the byproduct.
Capability 01Working Thesis
A first-class object for what you believe — and what would change your mind.
Every screen revolves around the analyst's current view. Evidence pulled from forecasts and scenarios auto-populates the supporting / contradicting columns. A conviction slider anchors the thesis to a number. Goldeneye's LLM critiques it: missed risks, blind spots, questions to answer.
- Persistent dashboard panel — never two clicks away
- Auto-populated from latest ensemble forecast and scenario run
- Critique drawer surfaces what the analyst may be missing
Capability 02Explainable Forecasts
Four models, weighted reasoning steps — and a public record of what they can't do.
The ensemble reads storage, positioning, and price signals; every call carries its supporting and contradicting factors with weights. A backtest engine replays the same models against real history under strict look-ahead controls — and where testing found no directional edge, the terminal says so and frames direction as views, never probabilities. The validated edge is the calibrated price-range band.
- Moving Average · Holt Trend · Factor Composite · Logistic Regression
- Look-ahead-safe replay with cheating-model property tests
- Every verdict published on the validation ledger — failures included
Capability 03Scenario Lab
Counterfactuals as a workflow — not a spreadsheet exercise.
Apply primitive shocks (weather, LNG export, production, storage) to a baseline and watch the forecast shift. Goldeneye's narrator explains the directional pressure, the strongest counterargument, and the data that would validate or invalidate each scenario in the next 1-2 weeks. Exportable as a PDF for the desk's morning meeting.
- Six pre-built templates, plus custom shock builder
- LLM-narrated outputs with safety envelope
- Executive PDF export with brand chrome
Capability 04Decision Quality
A reliability diagram for your own convictions.
The Calibration page buckets every resolved journal entry by claimed conviction and reports actual hit rate. The DQ Coach panel synthesizes per-bucket coaching — what patterns appear in your wins, what patterns appear in your misses, and one actionable suggestion per band.
- Resolved entries → reliability diagram with sample-size guardrails
- Auto-generated headline copy when calibration drifts ≥ 5 pp
- LLM coaching with explicit pattern-not-position framing
04 · The differentiator
We measure the decision, not the outcome.
Every thesis you write is logged with the data you saw, the counterarguments you weighed, and the conviction you assigned. When the market resolves, the system attributes outcome to calibration, not to luck.
Over time the calibration record builds a personal reliability diagram. Your 70% theses resolved at 64%. That gap is the most actionable signal an analyst can have — and nobody else is measuring it.
3Loop stages: pre-decision, mid, post
NConviction buckets, scored on hit-rate
LLMDQ Coach surfaces patterns across runs
05 · How we validate
We publish what our models can’t do.
Every predictive claim carries a verdict from a pre-registered, walk-forward test on real data — including the failures, which stay published. One claim has earned an edge; the rest of the ledger says so plainly.
Read the full ledger →edge · real-oos
The 80% price-range band is calibrated
Walk-forward coverage 78–81% across six commodities on ~10 years of real prices.
no edge · tested
Directional prediction
Every directional model tested on real data — none earned a claim. Direction ships as views, never probabilities.
no edge · tested
Curve carry as a timing signal
Adequately powered, pre-registered — and it failed. The verdict is published, not buried.
promising
Vol-premium timing vs implied vol
Passes on 2 of 3 asset pairs. Not crowned — it earns a surface only by passing its own gate.
collecting
Weather & curve feature archives
Untestable claims say so — immutable daily archives accumulate until honest validation is possible.
Condensed — the terminal renders the full ledger with live calibration numbers
06 · Built to last
Production-grade infrastructure with research-grade transparency.
FrontendNext.js 14App Router · RSC · TypeScript
BackendFastAPIAsync SQLAlchemy · Pydantic v2
DatabasePostgresTimescaleDB hypertables
IntelligenceClaudeAnthropic Sonnet + Opus
07 · The Ask
Goldeneye is building the decision infrastructure
for the probabilistic era.
The demo is live. The architecture is shipped. The next move is yours.