Home About Research
Overview Models The Leap Constitution Alma 2 Run Training Log
PROJECT NORTHFALL · MODELS

One identity. A growing model line.

Noema is the intended AI identity. Alma is the model line beneath it. Each generation has a different job, a different training record, and a clear boundary around what it can honestly do.

Noema is the behavior and identity users are intended to meet. Alma is the model family that must learn and prove those behaviors.

MODEL LINEAGE

From identity proof to general language.

  1. 01 Alma 1 Available Proved Noema's identity, boundaries, and constitutional routing.
  2. 02 Alma 2 Base Complete Base pretraining finished with new weights and a new tokenizer.
  3. → Alma 2 Constitutional Specification stage Founder interview and audit complete; governing specification 0.1.0 approved. Fine-tuning has not started.
  4. 03 Alma 3 Base Proposed · gated A separate 307.1M-parameter candidate; not training and not an Alma 2 resize.
MODEL PASSPORT · 01
ALMA / 01

Alma 1

Constitutional memorization and routing model
Available

Alma 1 proved that Northfall could build the complete model path from random weights: tokenizer, transformer, reviewed corpus, training loop, and deterministic evaluation.

What it does well
Identity, boundaries, short conversation, honest limits, and exact constitutional responses.
Honest limit
It memorizes and routes taught patterns. It is not a general-language model and should not pretend to know untaught facts.
Parameters
59.3 million
Architecture
12 layers · 8 heads · 640 width
Context
128 tokens
Tokenizer
256 raw-byte tokensSimple by design for a small reviewed corpus.
Current checkpoint learned
448 reviewed pairsThe historical Alma 1 corpus later grew to 478 pairs across 21 parts; it is not the new Constitution curriculum.
Evaluation
69 / 69 foundation433 / 448 exact full-corpus reproduction in the production run.
MODEL PASSPORT · 02
ALMA / 02-PRE

Alma 2

Pretrained language foundation
Complete

Alma 2 is a separate generation: a larger transformer, a new 16k tokenizer, and randomly initialized weights pretrained on public FineWeb-Edu text. Its completed base is preserved separately from future fine-tuning work.

What it learned
Statistical language structure and broad continuation patterns across its pretraining corpus.
Honest limit
The base does not yet reliably know Noema's identity, Mardenic's role, the authority hierarchy, tool rules, refusal behavior, or intended voice. Constitutional behavior is not evaluated.
Parameters
97,737,216
Architecture
12 layers · 12 heads · 768 width
Context
512 tokens
Tokenizer
16,000-token byte-level BPETrained locally from raw public text; no borrowed model tokenizer.
Pretraining corpus
10.67B unique FineWeb-Edu tokensPublic educational web text. Northfall owns the tokenizer, weights, manifests, and training system—not the source webpages.
Completed campaign
600,000 optimizer steps32,768 tokens per step; 19.66B token presentations.
Completed base-pretraining record Loading status…
— complete — steps train loss — monitoring loss* —

*The run included its original validation slice in training. The 2.9155 figure is contaminated monitoring loss, not held-out evidence. Promotion remains NOT EVALUATED.

MODEL PASSPORT · 03 · PROPOSED
ALMA / 03-PRE

Alma 3

Proposed gated scale-up candidate
Proposed

Alma 3 is a separate model generation prepared for a larger context and training budget. The raw FineWeb-Edu archive is complete, but model training is not authorized yet.

Entry gate
Alma 2 must pass the full constitutional proving loop. Alma 3 still requires a Data Ledger, content-family deduplication and splits, clean holdouts, mixture selection, and a 300-step shakedown.
Honest limit
This is a target configuration, not a trained model. Specifications may change if the shakedown exposes memory, thermal, throughput, or data-quality constraints.
Target parameters
307,147,776The older 294.6M profile remains a named fallback; neither profile is approved to train.
Architecture
23 layers · 16 heads · 1,024 width
Context
1,024 tokens
Tokenizer
16,000-token BPEManifest-bound; no pretrained model weights.
Data preparation
Raw archive completeNo approved token manifest, Data Ledger, clean split, or final mixture exists yet.
Proposed campaign
~70B token presentations2,136,231 optimizer steps if every prior gate passes.
NEXT GATE

Alma 2 does not become Noema by finishing a progress bar.

Audit resolved → rules approved → author a reviewed curriculum → fine-tune a copy → evaluate → decide promotion.

The 90-answer founder workbook is a behavioral specification, not automatic training data. A separate rights-cleared curriculum must teach a new copy of Alma 2 while Alma 1 and the protected base remain unchanged.

Review the Alma 2 run → Read the training record →