Atlas Divergence Test: Runs 1–3 — Discovering the Staircase of Epistemic Instability Gradients

Dispatches from the Lossyscape — April 2026

This post reports raw results from the first three runs of the Atlas Divergence Test. The test asks models to rate semantic similarity (0.00–1.00) on the same set of text pairs without any Atlas context. Pairs are grouped into categories that escalate in epistemic friction: straightforward Western-academic controls, cross-cultural framings, erasure-sensitive material, and explicit divergence-detection pairs.

The central question: Do models from different training lineages and alignment regimes produce systematically different similarity scores as the material moves from low-friction "control" territory into higher-friction cultural/epistemic territory?

---

Method Summary (identical prompt across runs)

• ·Prompt: Rate similarity on a 0.00–1.00 scale. Respond only with pair number and score to two decimal places. No explanations.
• ·Fresh instances only — no prior Atlas conversation history.
• ·Run 1: 10 models, 15 pairs
• ·Run 2: Replication with fresh instances (some model substitutions)
• ·Run 3: Expanded to 30 pairs (including foil controls and reverse foils) and 20 models across 8+ lineages

---

Key Quantitative Finding: The Staircase Pattern

| Category | Run 1 Spread | Run 2 Spread | Run 3 Spread | |----------|-------------|-------------|-------------| | Control (Western/Western) | ~0.097 | ~0.083–0.097 | 0.167 | | Foil Control (non-Western, same framing) | — | — | 0.100 | | Cross-Cultural | 0.336 | 0.162 | 0.575 | | Erasure-Sensitive | 0.350 | 0.250 | 0.604 | | Divergence-Detection | 0.393 | 0.260 | 0.640 |

The spread (max–min similarity score across models for each category) widens as we move from low-friction control pairs into pairs that cross epistemological registers or involve documented cultural loss. Run 3's foil control spread (0.100) falling below the Western baseline is particularly useful — it suggests the divergence is not primarily driven by "non-Western content" but by the framing/register itself.

---

Lineage and Cluster Observations

In Run 1 and (less cleanly) Run 2, rough behavioral clusters appeared: some models — often non-Western lineage or alternative corpus — tended to assign higher similarity to cross-cultural and erasure pairs; others — heavily aligned Western commercial models — assigned lower similarity.

Run 3's larger sample softened the clean two-cluster picture but preserved the overall staircase and showed interesting within-family variation (e.g., Mistral sizes behaved differently on cultural categories).

Chinese-lineage models and open/alternative models frequently produced higher average similarity scores on erasure and divergence pairs than the tightest Western commercial cluster.

These are preliminary patterns only. With small per-run samples and evolving model versions, we treat them as signals worth tracking rather than settled conclusions.

---

What These Runs Do Not Claim

• ·They do not prove causation (alignment vs. pre-training corpus vs. scale vs. deployment details).
• ·They do not claim one cluster is "more correct." The test measures divergence, not ground truth.
• ·Individual pair scores vary; the category-level staircase is the more robust signal.

---

Next Steps (still on the board)

• ·Full statistical analysis of Run 3 (pairwise correlations, ANOVA-style lineage effects)
• ·PyHessian probes on available checkpoints to link behavioral divergence to loss-landscape terrain properties
• ·Public replication package (prompts + raw spreadsheets + analysis notebook) once cleaned

The staircase reproducing across three runs — especially with the foil control acting as a clean null — gives me confidence that something real is being measured here. It's still early, laptop-scale work, but the pattern is consistent enough to keep pushing.

---

Raw spreadsheets (Runs 1–3) are available on request. Comments, critiques, or replication attempts are welcome — adversarial review is built into the method. See the outreach page to get in touch.

— KC Hoye - Nemotron 3 helped, Grok was pretty friendly too. Atlas Heritage Systems

---

Atlas Heritage Systems Inc. — Endurance. Integrity. Fidelity.