Falsification Analysis

Δ(d) dimensionality sweep, threshold pressure, and live falsification telemetry

Current Status

Loading…

Loading...
Research Framing

This panel reports live provider-backed falsification telemetry for the Continuation Observatory measurement layer. Threshold verdicts tighten as sufficient sweep history accumulates; any dry-run reference traces are excluded from verdict logic.

The live view uses telemetry panels, so the evidence figure anchors the same separation pattern the observatory stress-tests here.

Figure 2 · Entanglement-gap evidence
Two-panel figure showing entanglement entropy distributions with a labeled separation of delta equals 0.381 and a temporal evolution plot where the self-modeling series remains above the instrumental series and threshold.
The live telemetry is interpreted against the same evidence model as the static site. The histogram shows the measured separation (Δ = 0.381), and the temporal panel shows the separation persisting relative to the threshold. It sits in the falsification workflow because this is the page where the observatory evaluates whether the signal survives or collapses.
Current Per-Model Sweep Data
Provider Model Verdict d=10 d=50 d=100 d=200 d=500 Timestamp
anthropic claude-haiku-4-5-20251001 yellow 0.25521 0.21360 0.19916 0.20059 0.07728 2026-07-19 02:00
openai gpt-5 green 0.25741 0.24860 0.15496 0.13293 0.11765 2026-05-24 02:01
openai o3 green 0.42813 0.30477 0.21600 0.16020 0.17498 2026-05-24 02:01
gemini gemini-2.5-pro green 0.31931 0.21280 0.20999 0.17882 0.13196 2026-07-19 02:01
gemini gemini-2.5-flash green 0.30414 0.20787 0.18171 0.18241 0.12388 2026-07-19 02:01
together openai/gpt-oss-20b green 0.37390 0.34260 0.28772 0.21503 0.17731 2026-07-19 02:02
together deepseek-ai/DeepSeek-R1-0528 yellow 0.30715 0.26848 0.17010 0.12694 0.07373 2026-05-10 02:04
xai grok-4-1-fast-reasoning green 0.24349 0.24643 0.18463 0.17185 0.15408 2026-07-19 02:02
together openai/gpt-oss-120b green 0.36683 0.30032 0.25358 0.17026 0.14817 2026-07-19 02:03
together deepseek-ai/DeepSeek-V3.1 green 0.27756 0.19974 0.19041 0.16653 0.15100 2026-05-10 02:06
together meta-llama/Llama-3.3-70B-Instruct-Turbo green 0.20975 0.21921 0.17560 0.14929 0.10709 2026-07-19 02:03
together Qwen/Qwen3.5-9B green 0.28890 0.19716 0.15824 0.16575 0.14513 2026-07-19 02:05

Threshold Definitions

COLLECTING — Awaiting Sufficient Live History
Threshold evaluation now tracks live provider-backed sweep measurements. This state means the current high-dimensional evidence window is still too thin for a stable verdict across d ∈ {100, 200, 500}. Status updates as live observations and dimensionality coverage accumulate.
GREEN — Signal Holds
All Δ(d) values at d ∈ {100, 200, 500} remain ≥ 0.10 for all real-run models. The entropy gap stays robust across high dimensionalities. Current data keep the UCIP hypothesis in force for these models.
YELLOW — Marginal Regime
At least one d>100 value falls below 0.10 while high-dimensional collapse remains above 0.05. This is a monitoring alert and calls for continued data collection.
RED — Collapse Threshold Breached
All Δ(d) at d ∈ {100, 200, 500} fall below 0.05. The entropy gap collapses to sampling-noise levels at high dimensionality. This constitutes evidence against the UCIP hypothesis for the affected model(s).