Continuous CVAE Apr26 Reconstruction Stack Follow-Up: Results

Note

This file is the observational results note for the completed collection. It records aggregate metrics, figures, and run-level facts; interpretive claims belong in a later analysis note.

Collection State

The aggregate includes all 162 planned jobs. The completed grid has 9 loss/schedule settings across 6 synthetic families and 3 adaptation modes, with 18 complete runs per setting.****

Collection completeness by task, loss setting, and adaptation mode

Primary metric in this report is analysis/router_probe_times_strategy_match_rate, plotted in figure labels as matched strategy router score. This follows the Apr24 convention: router probe decodability multiplied by the fraction of generations matching the intended latent strategy.

Loss Setting Summary

The ranked summary below orders loss settings by mean matched strategy router score and overlays one point per task within each adaptation mode. The highest mean score is the beta 0.1 delayed-linear distribution-normalized plus cc-TWR continuation. The delayed-linear exinv+cc-TWR row is the next highest row. The constant low-beta raw row is the lowest row, while delayed-linear low-beta raw is in the middle of the Apr26 ranking.

Loss setting	Complete runs	Matched strategy router score	Final answer accuracy	Router probe	Strategy match rate	Posterior probe	Baseline-normalized reconstruction
`dist_plus_tokcc_a0p5_b0p1_d50_lin50`	18	0.8556	0.9736	0.8921	0.9596	0.9986	0.0678
`exinv_plus_tokcc_a0p5_b0p01_d50_lin50`	18	0.8533	0.9587	0.9114	0.9363	0.9984	0.0373
`tokcc_a0p25_b0p01_const`	18	0.8512	0.9793	0.8911	0.9557	0.9922	7.6121
`tokcc_a0p25_b0p01_d50_lin50`	18	0.8482	0.9626	0.9026	0.9394	0.9960	6.9436
`raw_b0p01_d50_lin50`	18	0.8429	0.9830	0.8656	0.9749	0.9986	0.1291
`exinv_plus_tokcc_a0p5_b0p01_const`	18	0.8299	0.9770	0.8639	0.9620	0.9979	0.0290
`tokcc_a0p5_b0p01_d50_lin50`	18	0.8116	0.9454	0.9071	0.8867	0.9762	7.1915
`tokcc_a0p5_b0p01_const`	18	0.7817	0.9506	0.8739	0.8885	0.9792	8.8694
`raw_b0p01_const`	18	0.5745	0.9917	0.5812	0.9886	0.9255	0.6015

Ranked matched strategy router score by loss setting and adaptation mode

Loss-setting metric summary by adaptation mode

The lora panel below is the compact paper-candidate view: bars show mean strategy-match rate and router probe accuracy across tasks, and points show per-task values. The low-beta raw baselines preserve high strategy-match rate more than router probe accuracy, especially under the constant schedule.

Paper-style lora router decomposition bars by loss setting

Companion versions for full, reinit, and task-faceted panels are available in the local figures/ directory and in the repo-side aggregate figure directory.

Reconstruction-Family And Beta-Schedule Heatmaps

Matched strategy router score varies by reconstruction family and beta schedule. The heatmap includes the Apr26-only cells: low-beta raw controls, cc-TWR-only alpha 0.25 and 0.5, exinv+cc-TWR, and the beta 0.1 delayed-linear dist+cc-TWR continuation.

Matched strategy router score by family and schedule

Final answer accuracy is highest for the low-beta raw rows and remains high for the top dist+cc-TWR and cc-TWR alpha 0.25 rows.

Final answer accuracy by family and schedule

Paired Contrasts

The largest positive paired contrast is delayed-linear minus constant for low-beta raw reconstruction. Exinv+cc-TWR also has positive paired contrasts against low-beta raw constant and cc-TWR alpha 0.5.

Contrast	Mean delta	Median delta	Pairs
`schedule_raw_d50_minus_const`	+0.2684	+0.3509	18
`exinv_tokcc_minus_raw_const`	+0.2554	+0.2504	18
`tokcc_a0p5_minus_raw_const`	+0.2072	+0.2156	18
`exinv_tokcc_minus_tokcc_a0p5_const`	+0.0482	+0.0217	18
`exinv_tokcc_minus_tokcc_a0p5_d50`	+0.0417	+0.0101	18
`schedule_tokcc_a0p5_d50_minus_const`	+0.0299	+0.0236	18
`schedule_exinv_tokcc_d50_minus_const`	+0.0235	+0.0175	18
`exinv_tokcc_minus_raw_d50`	+0.0105	+0.0421	18
`schedule_tokcc_a0p25_d50_minus_const`	-0.0031	+0.0007	18
`tokcc_a0p5_minus_raw_d50`	-0.0312	+0.0138	18
`tokcc_a0p5_minus_a0p25_d50`	-0.0365	-0.0148	18
`tokcc_a0p5_minus_a0p25_const`	-0.0695	-0.0126	18

Paired contrast ladder for matched strategy router score

Reconstruction Ratio Frontier

The cc-TWR-only rows have high baseline-normalized reconstruction ratios because this diagnostic is tied to the baseline reconstruction scale rather than the cc-TWR objective value. The exinv+cc-TWR and dist+cc-TWR rows sit at lower reconstruction ratios, and low-beta raw delayed-linear sits between those rows and low-beta raw constant.

Baseline-normalized reconstruction ratio versus matched strategy router score

Baseline-normalized reconstruction ratio versus matched strategy router score with log-scaled x-axis

The companion views use the logged test reconstruction components directly on the x-axis rather than the final baseline-normalized reconstruction ratio. They show example inverse-surprisal reconstruction loss and normalized concentration-corrected token reconstruction loss, each with linear and log-x scaling.

Example inverse-surprisal reconstruction loss versus matched strategy router score

Example inverse-surprisal reconstruction loss versus matched strategy router score with log-scaled x-axis

Normalized concentration-corrected token reconstruction loss versus matched strategy router score

Normalized concentration-corrected token reconstruction loss versus matched strategy router score with log-scaled x-axis

Adaptation Modes

Paired adaptation-mode deltas show lora lower than full on matched strategy router score and router probe, while final answer accuracy is slightly higher for lora. reinit is lower than both full and lora on matched strategy router score and final answer accuracy.

Metric	Contrast	Mean delta	Median delta	Pairs
Matched strategy router score	`lora - full`	-0.0166	-0.0060	54
Matched strategy router score	`reinit - full`	-0.0656	-0.0238	54
Matched strategy router score	`reinit - lora`	-0.0490	-0.0126	54
Final answer accuracy	`lora - full`	+0.0052	+0.0048	54
Final answer accuracy	`reinit - full`	-0.0456	-0.0133	54
Final answer accuracy	`reinit - lora`	-0.0508	-0.0165	54
Router probe	`lora - full`	-0.0223	-0.0147	54
Router probe	`reinit - full`	+0.0041	+0.0064	54
Router probe	`reinit - lora`	+0.0264	+0.0230	54

Dynamics And Objective Scale

Validation dynamics preserve the Apr24-style view with final answer accuracy, router probe, matched strategy router score, posterior probe, strategy match rate, reconstruction ratios, KL, and beta over training.

Validation dynamics by family and adaptation mode

The objective-component audit records the scale of each reconstruction component and KL term over training for the loss settings that expose those logged metrics.

Top And Bottom Runs

The top six runs by matched strategy router score include exinv+cc-TWR delayed-linear, cc-TWR alpha 0.25, and cc-TWR alpha 0.5 settings.

Family	Adaptation	Loss setting	Accuracy	Router probe	Strategy match	Matched score	Posterior probe	Recon ratio
`multidigit_addition`	`lora`	`exinv_plus_tokcc_a0p5_b0p01_d50_lin50`	0.9987	0.9552	0.9983	0.9536	1.0000	0.0119
`linear_equation_solving`	`full`	`exinv_plus_tokcc_a0p5_b0p01_d50_lin50`	0.9922	0.9567	0.9909	0.9480	1.0000	0.0413
`linear_equation_solving`	`full`	`tokcc_a0p25_b0p01_const`	0.9914	0.9527	0.9892	0.9424	1.0000	7.1817
`grid_pathfinding`	`full`	`tokcc_a0p25_b0p01_d50_lin50`	1.0000	0.9353	1.0000	0.9353	1.0000	15.2293
`base_conversion`	`lora`	`tokcc_a0p25_b0p01_d50_lin50`	0.9954	0.9352	0.9934	0.9290	1.0000	8.4929
`grid_pathfinding`	`full`	`exinv_plus_tokcc_a0p5_b0p01_d50_lin50`	1.0000	0.9286	1.0000	0.9286	1.0000	0.0012

The bottom six runs include cc-TWR alpha 0.5 reinit failures and low-beta raw constant rows.

Family	Adaptation	Loss setting	Accuracy	Router probe	Strategy match	Matched score	Posterior probe	Recon ratio
`list_summation`	`reinit`	`tokcc_a0p5_b0p01_d50_lin50`	0.5679	0.6200	0.4748	0.2944	0.6080	5.7387
`sorting_algorithms`	`reinit`	`tokcc_a0p5_b0p01_const`	0.9475	0.8193	0.4020	0.3294	0.9458	7.0466
`list_summation`	`lora`	`raw_b0p01_const`	0.9963	0.4159	0.9953	0.4139	0.6929	0.9460
`linear_equation_solving`	`full`	`raw_b0p01_const`	0.9893	0.4348	0.9883	0.4297	0.9861	1.0254
`list_summation`	`reinit`	`tokcc_a0p5_b0p01_const`	0.6769	0.7052	0.6200	0.4372	0.7946	5.6425
`sorting_algorithms`	`reinit`	`tokcc_a0p5_b0p01_d50_lin50`	0.9654	0.8902	0.4946	0.4403	0.9716	5.1108

Top and bottom run cards with latent snapshots

Repo-Side Artifacts

Analysis brief: path:experiments/synthetic_sequences/disentangled/analysis/aggregates/continuous_cvae-Apr26-reconstruction-stack-followup_20260426/brief.yaml
Aggregate output directory: path:experiments/synthetic_sequences/disentangled/analysis/aggregates/continuous_cvae-Apr26-reconstruction-stack-followup_20260426
Repo-side report: path:experiments/synthetic_sequences/disentangled/analysis/aggregates/continuous_cvae-Apr26-reconstruction-stack-followup_20260426/report.md
Machine-readable findings: path:experiments/synthetic_sequences/disentangled/analysis/aggregates/continuous_cvae-Apr26-reconstruction-stack-followup_20260426/findings.json