Brújula Systems
Home The Model Editorial How it works ES
← How it works

Transparency

What Brújula is actually running on, in real numbers. Not marketing counters. Live queries against the same database that produces the daily editorial.

The system, today

Numbers refreshed live from the database on page load (cached 60 seconds).

Public sources connected
49
Prediction markets, scientific archives, macro registers, climate data, media signals. All public and verifiable.
Sub-theses tracked
32
Grouped under 8 master theses. Each sub-thesis is anchored to at least one measurable evidence stream.
Events ingested (24h)
Total ever: . Every one is stored raw and re-processable.
Thesis moves (7d)
Probability updates in the last seven days. Every move logs the event that triggered it and the model's reasoning.

Model spend, last 30 days

Every LLM call the system makes is logged with tokens and cost. We publish the aggregate because a forecasting product that hides its unit economics is a product that isn't sure they add up.

Total LLM cost · 30 days
Full-stack: ingestion signal extraction, thesis application, daily editorial generation, translation and image prompts.
Total LLM calls · 30 days
Includes every level: cheap extraction (Level 1), editorial synthesis (Level 2), multi-model arbitration on critical moves (Level 3).
Prompt cache hit rate
Share of input tokens served from Anthropic's 1-hour prompt cache. Higher is cheaper — the system aggressively caches shared prompts.

Cost by provider

loading…

Most active sources, last 7 days

Not the ones we say we watch — the ones actually producing signal. Some sources are anomaly-driven and stay silent for weeks until they trigger; the ones below are the workhorses.

The model, today

The single thesis that moved the most over the last 28 days. Same data card the Brújula X thread posts every day as the anchor image. Refreshed daily around 13:30 UTC.

Today's top thesis movement — probability curve over the last 28 days

Teal = probability going up · Amber = probability going down · No red/green (design choice, doc §32)

Registered predictions — verifiable by third parties

Every prediction is snapshotted to a GitHub Release the day it is registered AND anchored on the Bitcoin blockchain via OpenTimestamps. The SHA-256 of the source block in PREDICTIONS.md is embedded in the release. A third party can (1) recompute the hash locally, (2) compare with the release, and (3) verify the Bitcoin timestamp with ots verify on the .ots file — offline, without trusting GitHub or us.

Neither GitHub nor Brújula can forge all three simultaneously.

The full source is in PREDICTIONS.md. API endpoint: /api/predictions.

What isn't here yet

A calibration score. The first predictions were registered in May 2026, and none have resolved yet. When the first resolution lands, its Brier score will show up here alongside every other number.

Waiting for real resolutions instead of interpolating them is the whole point. Meanwhile the snapshots above make sure we cannot pretend today's probabilities were different once the events happen.

© 2026 · Building in public Privacy Terms
Back to how it works →