Research line: the geometric stack

The "geometric stack" was a layered float model whose layers carried state with quaternion rotations. It was trained offline and then converted to 4-bit integers for serving. It matched an ordinary transformer on loss, served bit-exactly, and still wrote useless prose. It also used softmax, SiLU and matrix multiplication in training, so the owner's later bans made it a historical form. Some of its pieces live on as READY mechanisms that have never been trained.

Counts: archive/catalog.jsonl. Terms follow the glossary; ACTIVE, READY and PARKED are defined on the mechanism board.

flowchart LR
  A["24 to 26 Sep<br/>Float recurrent learner<br/>vs count baselines"] --> B["25 Sep<br/>Integer bridge and<br/>learned rounding"]
  B --> C["26 to 28 Sep<br/>Lorentz reads, codecs,<br/>exact relation store"]
  C --> D["29 Sep<br/>Quantization-aware<br/>training closes the gap"]
  D --> E["6 to 11 Oct<br/>D20 contest: no winner;<br/>claims withdrawn"]

The stack's arc: a good loss number, then repeated attempts to turn it into a model that talks.

What it set out to do

Train a deeper model with recurrence, geometric reads and exact memory, then compress it to integers and serve it without multiplications. The steps followed the owner's learning ladder (decision D8: reference baselines, then joint learning, then conversion to the discrete form): reference baselines, a joint recurrent learner, then conversion to the discrete form.

What happened, in order

1. Baselines and the recurrent learner (24 to 25 Sep). A neural reference reached 1.574 nats against 2.392 for the best 5-gram-plus-cache baseline (baselines). Both float recurrent arms reached about 2.11 nats and passed the engineering gate, but the quaternion arm trailed the ordinary control, so no geometric advantage was shown (result).
2. Integer forms (25 to 27 Sep). Four-bit quantization-aware training missed its +0.05-nat retention gate (+0.059 and +0.080) (result). Learned rounding passed (2.114 and 2.111 nats) and an integer bridge kept every source answer, about 28 times slower than float (learned rounding, bridge). A standalone integer command reproduced all 4,096 saved distribution hashes and ran 4.5 to 6.6 times faster after one optimization (serving). Continued language training lowered loss but produced 0 of 5 acceptable stories (continuation).
3. Geometric reads (26 to 28 Sep). A Lorentz (hyperbolic) read beat a dot-product read by 0.018 nats on code in 3 of 4 seeds, then lost to the matched dot reset when adapted (+0.046 and +0.013 nats) (cycle 3, adaptation). The stack matched a 7M-parameter transformer on repository code (1.998 against 2.011 nats), but generated code was useless (cycle 4). An exact relation store was perfect in distribution but its learned read trigger fired 0 of 826 times on held-out phrasing (exact relation store probe, D2).
4. Quantization (28 to 29 Sep). Round-to-nearest 4-bit export cost +0.0362 nats against a 0.02 gate (representation measurements). Quantization-aware training brought the integer engine within +0.0106 nats of its float model, bit-identical to a NEON reference (QAT).
5. Contest and withdrawals (6 to 11 Oct). In the D20 contest, a three-seed test of the geometric read against a parameter-matched control found no winner: the geometric arm scored 0.0073, 0.9995 and 1.0000 in class against about 0.25 for the control (D20 run). A claimed 12-times faster fused kernel was withdrawn because the pod never ran it (record).

Key results

What it taught

Loss and prose are different things: a good number coexisted with 0 of 5 stories (lesson 3 in what the project knows). Fixed geometry usually tied or lost to a matched control (lesson 6), and the quaternion-transport win is a hypothesis until a parameter-matched, lawful control exists. Compute went to the readout, not the geometry (lesson 14).

Current state

PARKED as a whole model. Its lawful pieces are tracked separately: the Hamilton gate and recurrence gates are READY but never trained. The next step in the dossier is a rebuild from lawful parts under D37 (offline float training, no matmul or softmax) with an exporter, which does not yet exist.

Go deeper: the generated catalog page lists all 84 records.