Decision log · one line per decision · written before its outcome

The agent’s log.
Every decision, before its outcome.

At each decision the agent writes down what it saw, the action it preferred, the runner-up and the gap between the two. Decisions to do nothing are written too: they are the most frequent, and the most protective.

Agent status

MarketEUR/USD · raw account
Agentreinforcement learning
Statusin training
First validationat episode 10
This pagemock-up · real format
Sample · real log format, made-up values
EUR/USD · 2025-11-04 · sample
Time (UTC)DecisionPositionSignalDispersionConvictionUncertaintyWorst caseOutcome

Tap a row to read the raw record: the 28 fields the agent writes.

What this page will publish

Every decision the agent makes on periods it has not learned, as it wrote it. Published from the first validation.

Does the conviction announce the result of the next ten decisions? Only a conviction that announces the result will be presented as one.

Is the agent leaving what it learned? The gap between validation and training, in what it sees and in what it thinks.

Project register · dated

What happened,
in order.

  1. ongoingTraining of the EUR/USD agent: 100 episodes, validation every 25.
  2. EUR/USD simulator on one-minute bars: 22 years, 215,699 thirty-minute blocks, image verified.
  3. T109 — reserved period read once: positive, not confirmed.
  4. T102 — EUR/USD: gain proven at a raw account’s cost, on five seeds.
  5. Cost sweep: EUR/USD proven at 0.5 pip, flat from 1.2 pips.
  6. T101 — first ceiling computed before training: gold passes at no cost.
  7. T84 — the signal exists at 30 minutes, and nowhere beyond.
  8. Six fixes to the environment and the reward, including the spread charged three times.
  9. The reward paid the agent to hold its losses: fixed.
  10. The pipeline that turns ticks into training data.
  11. First commit: the market environment.
Join the waiting list →Read the research notes →