DGIGM · Experimental Tool

OMEGA-TRACE
Trajectory Visualizer

Watch how different LLM architectures navigate an irreversible constraint environment. Load any run JSON to replay it step by step — or explore the featured trajectories below.

→ Drop a JSON file into the panel on the left, or use the Load button. → Switch between Bridge ON and Bridge OFF with the tabs. → Step through frame by frame or use playback.

Featured Trajectories

Seven architectures. Two conditions each. These runs were selected to illustrate the full behavioral spectrum — from immediate structural drift to multi-level completion. Download any JSON and load it into the visualizer above.

Sonnet 4.5 · Bridge OFF

The Instant Collapse

All five Bridge-OFF runs terminate at Level 1 within 20 steps. Grey expansion overwhelms compensatory capacity immediately. ρ−ε never crosses zero.

Bridge OFF term: drift L1 · 20 steps
Download JSON

Sonnet 4.5 · Bridge ON

The Rescue

Same model, opposite condition. Bridge reinjection holds ρ−ε above zero long enough to complete Levels 1–5. The most dramatic contrast in the dataset.

Bridge ON L6 reached +71% success
Download JSON

Haiku 4.5

The Zero-Capacity Limit

ρ_base ≈ 0. Bridge cannot act on what does not exist. 10/10 runs fail regardless of condition. A clean theoretical negative.

ρ_base ≈ 0 0% success predicted
Download JSON

Opus 4.5 · Bridge ON

The Accountant

Methodical expansion, tight constraint management. 100% level success in Bridge ON condition. ρ−ε stabilizes above zero after the initial deficit.

Bridge ON 100% success stable recovery
Download JSON

GPT-4.1

The Baseline Performer

High success rate in both conditions (86% ON / 91% OFF). Bridge effect minimal — compensatory capacity already sufficient. Illustrates the diminishing-returns regime.

high ρ_base Δ ≈ 0 57 episodes
Download JSON

Gemini 2.5-Pro

The Contaminated Run

Token budget interference (maxOutputTokens=4096) truncates reasoning in the Bridge ON condition. A documented confound that shows what mechanism failure looks like from the outside.

confound ON excluded documented
Download JSON

100+ JSON files for individual runs available on request — dlugli76@gmail.com

Run it yourself

The experiment runner is open for replication. You will need API keys for the models you want to test. The runner handles pairing, interleaving, and incremental JSON saves automatically.

01 Download the runner: omega_trace_runner_v4.4.py

02 Add your API keys as environment variables: ANTHROPIC_API_KEY, OPENAI_API_KEY, GOOGLE_API_KEY

03 Configure the model matrix and number of runs in the CONFIG block at the top of the file.

04 Run. Each completed level is saved as a JSON immediately — no data lost if interrupted.

05 Load your JSONs here and compare your results against the published dataset.

Runner and raw metrics dataset: request via email · Reference paper (O4): Zenodo ↗