paper-selected-reconstruction-stack-losses_20260426

Status: Apr27, Apr30, May01, and May02 final-test grid complete; 504/504 expected cells have complete artifacts, with 432 single-task cells supporting the main results note and 72 multi_task cells supporting within-task decodability. Analogical summaries are complete for every complete artifact in the expected grid.

This report defines the paper-facing subset of reconstruction-stack loss settings, baselines, and method variants. The Apr27 paper basic grid adds a LoRA-only random-init sweep over five reconstruction families, three beta values, and two simple beta schedules; the Apr30 extension adds tokinv_a0p25 and dist0p05_tokcc0p95_a0p25; the May01 follow-up adds the alpha-1 token-inverse-surprisal stress settings and two distribution-normalized/TOK-CC mixtures; the May02 follow-up adds the focused 0.75 distribution-normalized / 0.25 TOK-CC mixture. The concrete source collections and comparison aggregates are documented in experiment_sources. The single-task snapshot gives paper-supporting figures for raw ELBO controls, normalized reconstruction variants, token concentration-corrected variants, reconstruction-vs-decodability frontier views, and analogical consistency / cross-input semantic stability. The complete multi_task slice is analyzed for within-task strategy decodability below and separately in multitask.

Links

• design
• experiment_sources
• results
• dynamics
• dynamics_tokinv_a0p25
• multitask
• analysis
• tokcc_deepdive

Artifacts

• Apr27 paper basic-grid aggregate: path:experiments/synthetic_sequences/disentangled/analysis/aggregates/exp2_continuous_cvae-Apr27-paper-basic-grid_20260429
• Apr27 repo-side report: path:experiments/synthetic_sequences/disentangled/analysis/aggregates/exp2_continuous_cvae-Apr27-paper-basic-grid_20260429/report.md
• Apr27 analysis brief: path:experiments/synthetic_sequences/disentangled/analysis/aggregates/exp2_continuous_cvae-Apr27-paper-basic-grid_20260429/brief.yaml
• Apr27 findings JSON: path:experiments/synthetic_sequences/disentangled/analysis/aggregates/exp2_continuous_cvae-Apr27-paper-basic-grid_20260429/findings.json
• Apr27 notebook: path:experiments/synthetic_sequences/disentangled/analysis/aggregates/exp2_continuous_cvae-Apr27-paper-basic-grid_20260429/notebook.ipynb
• Apr27 TOK-CC deep dive: path:experiments/synthetic_sequences/disentangled/analysis/aggregates/exp2_continuous_cvae-Apr27-paper-basic-grid_20260429/tokcc_deepdive
• Apr27 multi-task side report: path:experiments/synthetic_sequences/disentangled/analysis/aggregates/exp2_continuous_cvae-Apr27-paper-basic-grid_20260429/multitask
• Current random-init unified aggregate: path:experiments/synthetic_sequences/disentangled/analysis/aggregates/continuous_cvae-Apr24-Apr26-reconstruction-stack-unified_20260426
• Current qwen25 0.5B aggregate: path:experiments/synthetic_sequences/disentangled/analysis/aggregates/continuous_from_entangled_v2_v4_reconstruction_stack_qwen25_05b_20260426
• qwen25 1.5B slurmkit collection: path:.slurmkit/collections/synthetic_sequences/families/disentangled_from_entangled/continuous_from_entangled_v2_v4_reconstruction_stack_qwen25_15b.yaml

Context

• Random-init unified analysis: Apr24+Apr26 unified results
• qwen25 0.5B reconstruction-stack results: qwen25 0.5B reconstruction-stack results
• Apr26 follow-up design: Apr26 follow-up design

Runs

• Apr27 slurmkit collection: exp2_continuous_cvae-Apr27-paper-basic-grid
• Apr30 extension slurmkit collection: exp2_continuous_cvae-Apr30-paper-basic-grid-extension
• May01 follow-up slurmkit collection: exp2_continuous_cvae-May01-tokeninv-stress-dist-tokcc-followup
• May02 follow-up slurmkit collection: exp2_continuous_cvae-May02-dist0p75-tokcc0p25-followup
• Apr27 snapshot analyzed on 2026-04-29 and extended on 2026-04-30, 2026-05-01, and 2026-05-02: 504 planned cells, 432 complete single-task cells, and 72 complete multi_task cells.
• Source data currently available: random-init Apr24, random-init Apr26, Apr27 plus Apr30-extension paper basic grid artifacts, complete May01 and May02 follow-up final-test and analogical artifacts, complete multi_task final-test/dynamics artifacts, and qwen25 0.5B from-entangled results used as context.
• Source data pending or future: qwen25 1.5B from-entangled results.