First Dispatch from the Lossyscape
The site is live. The framework is at v11. The PyHessian experiment is Priority 1. Here is where things stand.
The site is live. The framework is at v11. The adversarial review log has eleven models across multiple lineages. The Google Sheet is updating automatically. The lossyscape is open for business.
Here is where things actually stand.
What exists
The Loss Landscape Vocabulary Framework has been through eleven models and twenty-seven major revision cycles. The core vocabulary is stable: terrain properties, navigator properties, structural integrity, potential and tension. The Skywork qualifier collapse hierarchy — three independent variables plus four derived readouts — is the most formally precise finding to date.
The Behavioral Signal Assessment protocol is complete at v0.2. Seven models, thirty stimulus pairs, three tiers. The pilot has not been run yet.
The GPT-2 small baseline measurements are in. One un-replicated first-pass run. Perplexity ranges from 19.1 on technical documentation to 102.5 on academic abstracts. Inter-head coupling peaks at layers 9-10 (0.936-0.945) with an unexplained dip at layer 4 (0.675). Both findings are directional only.
What is missing
PyHessian on GPT-2 small. This is Priority 1. The Skywork coupling → viscosity collapse hierarchy is a mathematical argument. The empirical test requires computing actual Hessian eigenvalue spectrum and measuring whether the attention correlation proxy correlates with it. Until that is done, the framework's central structural revision is unconfirmed empirically.
The Pythia checkpoint series. Priority 4 but the most important single experiment for the framework's theoretical claims. Memory and viscosity are defined as distinct navigator properties. In a frozen deployed model they are not independently measurable. Pythia is the only setup that can attack this problem — multiple checkpoints from the same training run on the same data.
The BSA pilot run. The protocol is designed. The workbook is built. The run has not happened.
What is next
PyHessian first. Then OPT-125M perplexity comparison. Then Mistral BASE versus INSTRUCT. Then Pythia.
We'll see what slithers across the lossyscape.