T³ Atlas

public artifacts for architectures you can inspect, ablate, and reproduce

MirrorEthic's open lab: frozen claims, interactive demos, and citeable traces — not pitch decks.

Verify cl33 → Read the paper Live demo

Start with the operator-only language model: open checkpoints, one-command bottleneck and reverse-readout checks, then poke the same model live.

What this site is

T³ Atlas is the public record of MirrorEthic research systems that put the computation on a typed, testable surface — Clifford / Cl(3,3) machinery, operator streams, memory organs, and ecology traces — with the receipts attached.

If a claim can't survive an ablation, a reproduction script, or a live demo that tries to break it, it doesn't belong here.

Three doors

1  ·  Claim

Preprint: One Object (cl33-opLM). Measured costs, negative results, and what is not claimed.

2  ·  Artifact

Hugging Face bundle: weights, hashes, repro_bottleneck.py, repro_reverse_readout.py.

3  ·  Attack surface

Live operator core: chat with the transcript, reverse readout, and the same bottleneck the paper measures.

How to use this

for auditors

  1. Read the claim in the paper.
  2. Run the two reproduction scripts on the frozen checkpoints.
  3. Open the live demo and try to make the receipts lie.
  4. Only then dig into the trace library, Logic Machine, and Substrate CPU below.

Also in the lab

Trace library

Citeable inference ecology traces across lineages and substrates (GPT-2, Gemma3, Qwen2.5). Schema-stable JSONL; the viewer is one consumer of the data.

247
traces
12
lineages
3
substrates

Every trace records the live state the architecture computes during a forward pass: per-head primitive EMAs, σ envelope, Cl(3,3) Q invariant, Ω bivector and trivectors, blockade and cosurvival graphs, self-model surprise, ACT halt dynamics, per-stage logit top-K. The schema is stable and documented.

Logic Machine

Hybrid KGE + grounded LM on the same Cl(3,3) algebra. Live dials off the real backend; off-manifold residuals light when the system doesn't understand what it heard.

Substrate CPU

Interactive Cl(3,3) Turing machine in the browser (~200 KB, no GPU). Step forward and back exactly; geometric failure modes are visible in the wiring, not hidden in softmax confidence.

Reference & code

What you're looking at (model and load reference), benchmark library, Genesis, and related repos under MirrorEthic on GitHub.

What we are not

Not a prompt framework. Not governance poetry without a substrate. Not "trust us, it's deep." The site exists so strangers can hurt the thesis in public.

Lineages in the library

lineageckptssubstratecapabilities
loading…

Using the data

The trace format is documented in SCHEMA.md. Loading is plain JSONL parsing, no T³-specific tooling required. 

import json

def load_trace(path):
    out = {"meta": None, "geoms": {}, "chain_states": [], "frames": []}
    with open(path) as f:
        for line in f:
            r = json.loads(line)
            t = r["type"]
            if   t == "meta":         out["meta"] = r
            elif t == "stage_geom":   out["geoms"][r["stage_idx"]] = r
            elif t == "chain_state":  out["chain_states"].append(r)
            elif t == "frame":        out["frames"].append(r)
    return out

Citation

Citation block forthcoming. For now, please cite as: 

Sutherland, G. (2026). T³ Atlas. https://t3atlas.dev