Model reconciliation

Model reconciliation

Tests of the original Ultra idea that each trie node makes a local model proposal and parents reconcile the child models leaf-to-root, weighting them by Fisher or precision. Four minimal prototypes on Tiny Shakespeare (legacy train and held-out PPL; depth 8, stride 1 where stated):

• Transition model e_child = e_parent + r[token], local state gradient and diagonal Fisher summed upward: unguarded, train PPL exploded 64.85 -> 645.06 in epoch 1; with a backtracking line search it was stable, held-out 19.95 at epoch 9 (train 19.67), weak but with train and held-out close.
• Distribution-only reconciliation (pseudo-counts passed upward): a full child merge hurt (held-out 35.88 -> 40.38); small child weights helped slightly, best 27.44 -> 26.75.
• Progressive logit-model reconciliation: best with child weight 0.0 (held-out 21.37, unigram baseline 28.43); every nonzero child weight was worse.
• Shared-basis head adapters (LoRA-like, fixed random basis): rank 8 reached 26.90 with no child benefit, and epoch persistence was flat.

Conclusion: child-to-parent messages never beat the zero-child ablation. Local model objects are not automatically composable. The design notes conclude that messages must live in a shared coordinate system in information form (precision plus information vector, pulled through a transition Jacobian) with a trust rule, and that pseudo-count merging is smoothing, not model merging.

Code and records (under research/ultra/):

• scripts/run_reconciled_transition_model.py, run_distribution_reconcile.py, run_progressive_logit_reconcile.py, run_logit_model_reconcile.py (no recorded results), run_shared_basis_adapter_reconcile.py
• agpt_ultra/reconcile.py (Fisher-weighted merge theta_p = (sum F_i)^-1 sum F_i theta_i)
• notebook/archive/state_fisher_results.md: "Reconciliation Through A Transition Model" through "Shared-Basis Head Adapter Reconciliation"
• notebook/archive/reconciliation_design_notes.md