Calibration: how reliable is it?
Here we will measure, with data and without being able to rewrite history, whether the confidence Tlatolli declares matches what happened next. We will only publish results aggregated by confidence band, including the ones that turn out badly.
Draft methodology: it may be adjusted before the first measurement, and every change will be versioned.
Current status: insufficient data
Measurement starts when the measurement program (a MetaTrader 5 EA) records how the price moved after each tone change. Until there are at least 30 cases, each cell says "insufficient data". We do not show sample figures.
Confidence bands
0.5 to 0.6
Insufficient data
0 cases measured · published at 30 or more
0.6 to 0.7
Insufficient data
0 cases measured · published at 30 or more
0.7 to 0.8
Insufficient data
0 cases measured · published at 30 or more
0.8 or more
Insufficient data
0 cases measured · published at 30 or more
Is the reading of the news correct?
Each month, a panel of stronger models and a sample reviewed by the founder re-assess randomly chosen news. If Tlatolli declares 0.8 confidence, an independent reviewer should agree about 80 % of the time.
What did the market do next?
After each recorded tone change, the move at 1, 4 and 24 hours is measured in multiples of ATR(14), against control moments without news at the same hour and weekday. It is a historical description, not a forecast.
What counts as an event
- Sign change: the tone goes from positive to negative, or the other way, outside the ±0.05 neutral zone.
- Strong threshold: the tone reaches 0.3 or more in absolute value.
- Jump: the tone changes by 0.2 or more from the last event.
- For pairs, the same rules apply to the relative tone: (base − quote) / 2.
What we will publish
- By confidence band, market and horizon: the number of cases, the share in the direction of the tone with its 95 % Wilson interval, and the median move in ATR against the control.
- Calibration charts: declared confidence against observed agreement.
- The month-by-month history, with each version of the AI instructions measured separately.
What we will never publish
- Entry or exit prices, candles or individual trades.
- Only the good results: every band with enough data is published.
- A pair that does well as proof for the others: each pair is published separately.
The earlier studies we registered before seeing the data came out null. A result close to 50 % is to be expected, and it will be published anyway.
Sealed ledger
Every tone change goes into an insert-only ledger. Each UTC day is sealed with a SHA-256 hash chained to the previous day: changing a single past event would break the chain.
Currencies and markets
1 sealed days · 16 events · since October 28, 2026
Chain intact: the site recomputed 1 seals and they all link up.
Latest seals
| Day (UTC) | Events | Seal |
|---|---|---|
| 2026-10-28 | 16 | 2d2e969b08b3a896… |
Checked on Oct 29, 2026, 6:20 AM UTC
Pairs
1 sealed days · 6 events · since October 28, 2026
Chain intact: the site recomputed 1 seals and they all link up.
Latest seals
| Day (UTC) | Events | Seal |
|---|---|---|
| 2026-10-28 | 6 | 462185945912405b… |
Checked on Oct 29, 2026, 6:20 AM UTC
How to verify it: seal = SHA-256 of "previous seal|day|number of events|hash of the day's rows".
The full list is in the Pohuant API, with any read key: /v2/ledger/seals and /v2/ledger/pair-seals.