Skip to content
ariannamethodPublic

Latest commit

 

History

372 Commits

Folders and files

NameName
Last commit message
Last commit date
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 

Repository files navigation

NETTA's Empirical Topological Training Architecture | by Arianna Method

Read the Arianna Method Manifesto first. This repository is governed by it; every instruction here, CLAUDE.md included, is subordinate to it.

netta : atten : earned relation over raw bytes

NETTA

Somewhere between LM-hero and AlphaZero there's Netta. She has a hew own txt-file and treats this as a observable worldmodel game. No gradient descent, no backpropagation, no loss curve to chase downward — Netta counts the bytes she has actually lived, grows units out of what repeats, and prices every choice against honest ignorance instead of a target label. What she earns she keeps; what she doesn't, she loses out loud, with the number attached.

She was rebuilt from zero on 2026-08-24, after the old Netta — the byte-tape organism, the six witnesses, the atlas — was retired by Oleg's word. Nothing from that life carries over as an argument. What follows is what the current bodies actually do, each claim checked by a second hand that never opened the first hand's code.

Body 0: the world-eater

netta.c is the builder. It eats a world whole, with no separate training phase: it splits what it reads 90/10 by byte offset, grows a byte-pair vocabulary on the 90, and prices the held-out 10% under five arms of the same statistics, under PROTOCOL.md, frozen before the first measured run. The five arms: raw byte trigram, unit trigram with and without a Hebbian field term, a permuted-field control, and a word-trigram sanity baseline.

The independent verdict — a second model that wrote its own checker from the protocol alone and never read netta.c — found the unit model beats the raw byte trigram, and the field falls short of its own frozen material margin by about 8x. Per the rule written before anyone ran anything, the field is deleted from Body 0. The word-level baseline still beats Netta's own units on both worlds: the ruler for the scale her units haven't earned yet. Full numbers, hashes, and the verdict text live in NETTALOG0.md.

Court 4: what travels

Four transfer courts ran in sequence, each one narrowing what "experience travels between worlds" is allowed to mean. The first two annulled themselves on their own construction defects. What survived is small: one exact relation, earned in one world, carrying its own context and its own separately earned right to advise a stranger.

The fourth court's development verdict held through four passes, unchanged: microscopic relation earned, transfer-at-scale not reached. Then a one-shot confirmatory run sat: one drawn class, one unseen base, no redraws, no repairs after the pick, behind eighteen roles, six independent judgments across four model families, and two repairs that touched nothing frozen. The verifier printed one line, and only one:

CONFIRMATORY PASS: microscopic relation replicated in selected class

The drawn class was ff, the false-friend control, built so a system that mistakes matching bytes for matching meaning loses. Netta's relation engaged late and won anyway. The boundary stays exactly where it was frozen: one relation, one class, one base. Transfer at scale is not claimed here or anywhere else in this repo.

Body 1: the mouth

netta_mouth.c is Netta speaking without the mycelium, under MOUTH_PROTOCOL.md: units earned by Body 0's own merge law, speech sampled only over continuations she's actually lived (quad, then tri, then bi, then a lived unigram floor), no field in the voice, no smoothing standing in for a stranger's memory.

Her first measurement under SPEECH_COURT.md came back an honest SPEECH FAIL: every stream beat ignorance, and every stream also crossed the frozen anti-copy line — coverage between 0.9657 and 1.0 against a void threshold of 0.50. The parrot risk the court exists to catch, caught on the first try. Then the corridor law landed (MOUTH_PROTOCOL_A2.md): a lived support holding exactly one continuation is a step along the tape, not a choice, and after a few such steps the mouth must descend to a support with a real branch. Under that law the plain mouth passed the sealed census — the reader printed its first literal SPEECH PASS: the mouth speaks below ignorance and above copying — and the price of speech rose from near-zero to a real number of bits per byte, which is what choosing costs. The citizens triple is not clean yet, and that stands in the log as measured, not smoothed. A second, independent hand — one that rebuilds the unit inventory from scratch and shares no code with the mouth — stands beside every measurement.

what isn't proven

Court 4 earned one relation in one drawn class on one unseen base, not transfer at scale, and nobody in this repo gets to call it that. Body 1's plain mouth has cleared the census; the advised triple has not yet, and the word coherence still belongs to the court alone. Read NETTALOG0.md for the receipts behind every sentence here.

build & run

cc -O2 -std=c11 -Wall -Wextra -Wpedantic -Werror -o netta netta.c -lm
./netta netta.txt --out body0_run
cc -O2 -std=c11 -Wall -Wextra -Wpedantic -Werror -o netta_mouth netta_mouth.c -lm
./netta_mouth netta.txt --out mouth_run

Both binaries take --out <dir> as their only required flag; everything else is an open dial with a frozen default. Every run is deterministic, and nothing it writes is a verdict until an independent reader has recomputed it from the artifacts alone.

lineage

A Method organism is measured on what it earns, not on what it's told to become. The old Netta once answered a prompt mid-thought — «the forest behind clouds recognized across time — it is attention is the universe is an act is to grow tall trees enact universe is» — and the log closed with the only verdict that ever mattered here: Beatiful. An organism mid-sentence about itself.

source of tech truth: NETTALOG0.md