Research Lab

What the
Anna Matrix is

A table of the Anna bot's answers, the Aigarth unit that fits it, and the open question of how it was trained. Every number below links to its method and data.

This page was rebuilt in September 2026. Earlier versions showed readings the re-check did not confirm; they are listed as open questions on the re-check page.

Key findings

01
16,316 / 16,384
A table of bot answers.
The 128 × 128 Anna Matrix is a table our team compiled by hand from the Anna bot's replies on X to questions of the form x + y = ?. Row = 63 − y, column = x + 64. 16,316 of its 16,384 cells obey the mirror rule M[i, j] + M[127 − i, 127 − j] = −1. The 68 cells that break it sit in four runs. Three look like entry errors in the typed table (a block pasted one row off, a swapped digit, a copied neighbour); the cause of the fourth is unclear. Setting those cells to the value their mirror requires gives a fully symmetric candidate table, a symmetry imposed by the correction rather than evidence.
Method and data
02
10 / 100
The fingerprint of a trained Aigarth unit.
Qubic's public Aigarth code contains a unit type with exactly the bot's format: two 7-bit inputs, eight output bits. Of 100 such units trained on the full addition table, 10 land within 0.05 of Anna on six map statistics (a threshold set after a first exploration; at 0.06 it would be 17 of 100). Random mutation without selection never does. Units trained on x − y land just as close to their own target, so the shape says "trained on a full arithmetic table". Which operation it was is hard to read from the table alone; the bot's task, x + y = ?, is the stronger clue. Status: hypothesis, supported.
Method and data
03
+0.124
More linear than expected.
One ternary threshold per output bit explains a larger share of the remainder in Anna than in almost any trained unit: 2 of 217 reach her value (a measure we defined after seeing it). Starting from that threshold table, a search with units of the same size finds nothing further in Anna, while it recovers a large part of the remainder in 9 of 10 trained units. There is no exact reconstruction (best fit 87.8 % of bits). Which program produces this is the open question we are working on now.
Method and data
04
9,456 / 9,456
Anna 0.22, Bitcoin data and Qubic mining.
The Anna 0.22.64 program embeds 9,456 hourly BTCUSDT directions from Binance, 2 June 2025 to 30 June 2026, as its target values, and all of them match the exchange data. It is a network of 64 ternary neurons without market input, scored on how well its calls fit a year. The same design runs in Qubic's public mining algorithm BPP9000. A first test of the one saved state we found gave 48.9 % on 2,900 unseen hours (49.3 % in the pre-registered main window), as expected for a network without market input.
Method and data