← Framework

Framework / Version History

Version History

Changelog for the Loss Landscape Vocabulary Framework. The revision history is itself a record of the adversarial development process — what each model review found, what changed as a result, and what remains open.

●
v12April 2026current
  • ·Global Geometry tab added — four terms promoted from v11 coverage gaps to first-class status: Basin Connectivity, Symmetry Orbits, Phase Transitions, Algorithmic Branches. Each explicitly scoped as global properties not visible to local terrain or navigator qualifiers. Tab sits between Topology and Navigator.
  • ·Structural Integrity — Regime Switch card added: training point where internal algorithm reorganizes while loss stays smooth. CKA formalism (CKA(R_θ(t−), R_θ(t+)) ≪ 1 while |L| stays small). Framed as the local Structural Integrity signature of what Global Geometry calls a phase transition.
  • ·Structural Integrity — Memory–Viscosity Identifiability card added: entangled vs separable as provisional labels. Entangled when curvature changes and training trajectory produce indistinguishable behavioral effects. Separable when Pythia multi-checkpoint data reveal behaviors viscosity cannot account for. Central open experiment of the Atlas framework.
  • ·Architecture tab — Skywork Coverage Gaps box updated to note formal promotion of basin connectivity, symmetry orbits, and phase transitions to Global Geometry tab. Nemotron-3-Super-120B added to adversarial review model list.
  • ·Version line updated v11 → v12.
  • ·Nemotron-3-Super-120B cold read (framework HTML v12, no prior context): proposed Global Geometry section and Algorithmic Regime additions. All four Global Geometry terms confirmed consistent with cited literature. No unacknowledged internal contradictions identified.
  • ·DeepSeek V3.2 cold read — Session 1: Slope notation (∇L should be ‖∇L‖, vector not magnitude), Memory–Viscosity identifiability citation gap, Algorithmic Branches citation gestural. Global Geometry passed clean.
  • ·DeepSeek V3.2 cold read — Session 2: Flexion/Elasticity unacknowledged overlap (same concept in two layers, same formal expressions). Potential Difference notation undefined (θ_terrain/θ_navigator not defined in formalism). Harmonics dimensional inconsistency (ω equated to Hessian eigenvalues — frequencies ≠ curvature values). Symmetry Orbits undermine Ablation Drift Vector (parameter displacement ≠ functional change without accounting for orbits). Basin Connectivity destabilizes Basin and Ridge definitions.
  • ·Mistral Large-3 cold read: confirmed anchor behavior — copy-paste of referenced definitions, no independent findings. Corroborated: resistance drag coefficient dimensionally undefined, basin width proportionality lacks volume scaling, Regime Switch CKA formalism uncited in cited sources, tension norm unspecified.
  • ·v13 punch list documented and deferred: Slope notation fix, Flexion/Elasticity overlap acknowledgment, Potential Difference notation revision, Harmonics dimensional flag, resistance drag coefficient dimensional flag, basin width proportionality note, Symmetry Orbits/Ablation Drift contradiction acknowledgment, Regime Switch CKA citation flag, Memory–Viscosity theoretical motivation citation.
○
v11April 2026
  • ·Skywork methodology review added as formal architecture box — three design problems and two missing experiments documented
  • ·Next experiments section rewritten with priority ordering and falsification criteria
  • ·Mistral BASE/INSTRUCT built-in falsification criterion articulated: RLHF perplexity reduction should be largest in archaeological domains, not laminar domains
  • ·Heisenberg conjugacy formally demoted from mathematical principle to structural analogy with methodological content (Skywork, April 2026)
  • ·Restoring force open problem further resolved via active inference — Friston's free energy minimization produces intrinsic oscillatory dynamics not regularization
  • ·Archaeological claim gains mechanism via Song et al. (2024): weights encode predictions not inputs; high-perplexity regions are places prediction error was never resolved before weight update
  • ·Adversarial review subtitle updated to include all model versions: GPT-4, GPT-5.2, Perplexity Sonar, DeepSeek V3, DeepSeek V3.2, Mistral Large, Mistral Large-3, Llama, Llama3.3 70B, Grok, Skywork
○
v10April 2026
  • ·Skywork methodology review added as formal architecture box — three design problems identified: ALBERT prompt is literature-reasoning not architectural, OPT should be empirical not model prompt, Falcon legibility confound requires two prompts
  • ·PyHessian on GPT-2 added as Priority 1 unassigned experiment
  • ·Pythia upgraded to multi-scale: 70M, 160M, 410M, 1B minimum
  • ·OPT-125M redesigned as empirical perplexity comparison, not model prompt
  • ·ALBERT-base redesigned as PyHessian empirical experiment
  • ·Falcon redesigned with two-prompt legibility test
  • ·Next experiments section rewritten with priority ordering
○
v9April 2026
  • ·Base model specification added to experimental results card: unmodified pretraining weights, no RLHF
  • ·Coupling probe method specified: pairwise Pearson correlation of attention weight matrices, single probe input, inference-time behavioral measurement not Hessian off-diagonal
  • ·Replication flag added to experimental results: ONE UN-REPLICATED FIRST-PASS RUN
  • ·Pythia upgraded to PRIORITY — both path dependency test and memory/viscosity distinguishability test
  • ·All three flags sourced to Skywork casual review April 2026
○
v8April 2026
  • ·GPT-2 small first-pass experimental results logged: perplexity map across eight domains, inter-head coupling across twelve layers
  • ·Adversarial review findings embedded as sourced dated cards: qualifier collapse hierarchy, coverage gaps, ablation drift vector corrections
  • ·Six new reference cards added: symmetry and permutation invariance, basin connectivity, representational geometry and CKA, ablation methodology, Skywork adversarial review
  • ·Skywork adversarial review cited as formal source alongside papers
○
v7April 2026
  • ·Skywork qualifier collapse hierarchy formalized: seven qualifiers collapse to three independent variables plus four derived readouts
  • ·Three independent variables confirmed: density, coupling, elasticity
  • ·Four derived readouts confirmed: perplexity, probability, viscosity, memory
  • ·Ablation drift vector structural problems named: dimensionality mismatch, weight-space ≠ functional distance, post-hoc ≠ pre-training absence
  • ·Three new coverage gaps added: basin connectivity, symmetry orbits, phase transitions
  • ·Scope boundary established: framework describes weight-space geometry not representational geometry
○
v6April 2026
  • ·Potential difference established as primary generative quantity, replacing resistance as the primary frame
  • ·Tension added as structural condition holding potential difference stable
  • ·Harmonics added as dynamic behavior emerging from potential/tension interaction
  • ·Flow and resistance vocabulary demoted to derived descriptive layer
  • ·Revision status box added documenting live revision in progress
○
v5April 2026
  • ·Mistral Large adversarial review: potential difference revision identified as lateral move, no restoring force, curvature eigenvalues ≠ restoring forces
  • ·Harmonics restoring force identified as open problem
  • ·DeepSeek V3 adversarial review: Reynolds number dimensional incoherence confirmed, retained as illustration only
  • ·Turbulence observability constraint confirmed: measurable only during movement not from static model
○
v4April 2026
  • ·Perplexity Sonar adversarial review: terrain/navigator partition not fully defensible formally
  • ·Structural integrity reducibility challenge logged
  • ·Archaeological claim flagged as asserted not derived
  • ·Friction math/metaphor divergence identified
○
v1–v3April 2026
  • ·Initial vocabulary established: terrain properties, navigator properties, macro-topology shapes
  • ·Heisenberg conjugacy proposed as formal principle (later demoted in v11)
  • ·Structural integrity established as separate category
  • ·Archaeological claim first articulated
  • ·GPT-4 adversarial review: confirmed known weaknesses, no new load-bearing findings