Unified Apr24+Apr26 Reconstruction Stack: Results

Note

This file is the observational results note for the unified Apr24+Apr26 analysis. It records aggregate metrics, figures, and matched contrasts; interpretive claims belong in analysis.

Source State

The unified aggregate combines two completed source collections:

Collection	Complete cells	Loss settings	Complete runs per loss setting
Apr24 reconstruction stack	234 / 234	13	18
Apr26 reconstruction stack follow-up	162 / 162	9	18

The combined loss-setting table has 22 rows. The raw-baseline subset has 5 rows: Apr24 raw_b1_const, raw_b1_lin100, raw_b1_d50_lin50, Apr26 raw_b0p01_const, and Apr26 raw_b0p01_d50_lin50.

All Loss Settings Ranked

The ranked table orders every Apr24 and Apr26 loss setting by mean analysis/router_probe_times_strategy_match_rate.

Rank	Collection	Loss setting	Complete runs	Matched strategy router score	Final answer accuracy	Router probe	Strategy match rate	Posterior probe	Baseline-normalized reconstruction
1	Apr26	`dist_plus_tokcc_a0p5_b0p1_d50_lin50`	18	0.8556	0.9736	0.8921	0.9596	0.9986	0.0678
2	Apr26	`exinv_plus_tokcc_a0p5_b0p01_d50_lin50`	18	0.8533	0.9587	0.9114	0.9363	0.9984	0.0373
3	Apr24	`dist_plus_tokcc_a0p5_b0p01_d50_lin50`	18	0.8532	0.9543	0.9145	0.9330	0.9980	0.0379
4	Apr26	`tokcc_a0p25_b0p01_const`	18	0.8512	0.9793	0.8911	0.9557	0.9922	7.6121
5	Apr26	`tokcc_a0p25_b0p01_d50_lin50`	18	0.8482	0.9626	0.9026	0.9394	0.9960	6.9436
6	Apr24	`dist_plus_tokcc_a1_b0p01_d50_lin50`	18	0.8465	0.9546	0.9053	0.9355	0.9975	0.0356
7	Apr24	`dist_plus_tokcc_a0p5_b0p01_const`	18	0.8432	0.9765	0.8776	0.9624	0.9963	0.0354
8	Apr26	`raw_b0p01_d50_lin50`	18	0.8429	0.9830	0.8656	0.9749	0.9986	0.1291
9	Apr24	`three_way_equal_b0p01_d50_lin50`	18	0.8418	0.9484	0.9070	0.9276	0.9980	0.0406
10	Apr24	`example_inv_surp_b0p01_d50_lin50`	18	0.8319	0.9309	0.9128	0.9096	0.9980	0.0407
11	Apr26	`exinv_plus_tokcc_a0p5_b0p01_const`	18	0.8299	0.9770	0.8639	0.9620	0.9979	0.0290
12	Apr24	`dist_plus_tokcc_a1_b0p01_const`	18	0.8273	0.9715	0.8700	0.9528	0.9915	0.0890
13	Apr24	`dist_norm_b0p01_d50_lin50`	18	0.8185	0.9295	0.9015	0.9075	0.9984	0.0331
14	Apr24	`example_inv_surp_b0p01_const`	18	0.8166	0.9666	0.8615	0.9514	0.9936	0.0437
15	Apr24	`dist_norm_b0p01_const`	18	0.8149	0.9668	0.8601	0.9499	0.9962	0.0642
16	Apr26	`tokcc_a0p5_b0p01_d50_lin50`	18	0.8116	0.9454	0.9071	0.8867	0.9762	7.1915
17	Apr24	`three_way_equal_b0p01_const`	18	0.8087	0.9707	0.8491	0.9549	0.9969	0.0357
18	Apr26	`tokcc_a0p5_b0p01_const`	18	0.7817	0.9506	0.8739	0.8885	0.9792	8.8694
19	Apr26	`raw_b0p01_const`	18	0.5745	0.9917	0.5812	0.9886	0.9255	0.6015
20	Apr24	`raw_b1_lin100`	18	0.4432	0.9916	0.4485	0.9882	0.5817	1.0690
21	Apr24	`raw_b1_const`	18	0.4419	0.9913	0.4470	0.9884	0.6044	1.0718
22	Apr24	`raw_b1_d50_lin50`	18	0.4396	0.9914	0.4448	0.9879	0.7294	1.0712

All Apr24 and Apr26 loss settings ranked by matched strategy router score

Core Metric Bars

The selected metric-bar views show the top 10 combined settings plus all raw baselines, separately for each adaptation mode. Each figure shows matched score, router probe, strategy match rate, and final answer accuracy.

Selected loss-setting metric bars for full adaptation

Selected loss-setting metric bars for LoRA adaptation

Selected loss-setting metric bars for reinit adaptation

Raw Baselines

The raw-baseline table isolates standard or near-standard raw reconstruction controls from the normalized and cc-TWR variants.

Raw rank	Collection	Loss setting	Matched strategy router score	Final answer accuracy	Router probe	Strategy match rate
1	Apr26	`raw_b0p01_d50_lin50`	0.8429	0.9830	0.8656	0.9749
2	Apr26	`raw_b0p01_const`	0.5745	0.9917	0.5812	0.9886
3	Apr24	`raw_b1_lin100`	0.4432	0.9916	0.4485	0.9882
4	Apr24	`raw_b1_const`	0.4419	0.9913	0.4470	0.9884
5	Apr24	`raw_b1_d50_lin50`	0.4396	0.9914	0.4448	0.9879

Raw reconstruction baselines ranked by matched strategy router score

The raw-baseline delta plot shows task/adaptation matched deltas for Apr26 low-beta raw rows against the closest Apr24 beta-1 raw rows. Both contrasts have 18 matched task/adaptation/seed pairs.

Matched raw-baseline deltas by task and adaptation

Adaptation Modes

The adaptation-mode comparison shows the best loss setting within each collection and adaptation mode, alongside raw-baseline trajectories across adaptation modes.

Best loss settings and raw baselines by adaptation mode

Frontier And Family Views

The scatter plot places each loss-setting mean by final answer accuracy and matched strategy router score. Raw beta-1 rows cluster near high final answer accuracy and low matched score; the Apr26 delayed-linear low-beta raw row moves into the main high-score cluster.

Final answer accuracy versus matched strategy router score

The family/schedule heatmap groups loss settings by broad reconstruction family and beta schedule. The heatmap uses a single-color Blues scale.

Matched strategy router score by broad family and beta/schedule

The reconstruction frontier places run-level matched score against baseline-normalized reconstruction loss on a log x-axis. A separate component-loss view plots raw, distribution-normalized, example-inverse-surprisal, and cc-TWR reconstruction losses against matched score when those component metrics are logged.

Baseline-normalized reconstruction frontier

Component reconstruction losses versus matched strategy router score

Matched Contrasts

Matched contrasts pair runs by task, adaptation mode, and seed. Positive deltas mean the Apr26 setting is higher than the Apr24 comparator.

Contrast	Pairs	Mean delta	Median delta
`apr26_raw_b0p01_d50_minus_apr24_raw_b1_d50`	18	+0.4033	+0.3919
`apr26_raw_b0p01_const_minus_apr24_raw_b1_const`	18	+0.1326	+0.0416
`apr26_exinv_tok_d50_minus_apr24_example_inv_d50`	18	+0.0215	+0.0162
`apr26_tokcc_a0p25_const_minus_apr24_dist_tok_a0p5_const`	18	+0.0081	-0.0158
`apr26_dist_tok_b0p1_d50_minus_apr24_dist_tok_b0p01_d50`	18	+0.0024	-0.0112
`apr26_exinv_tok_d50_minus_apr24_dist_tok_a0p5_d50`	18	+0.0002	+0.0039
`apr26_tokcc_a0p25_d50_minus_apr24_dist_tok_a0p5_d50`	18	-0.0050	-0.0049
`apr26_tokcc_a0p5_const_minus_apr24_three_way_const`	18	-0.0270	-0.0049
`apr26_tokcc_a0p5_d50_minus_apr24_three_way_d50`	18	-0.0302	-0.0080

Selected Apr26 minus Apr24 contrast ladder

Repo-Side Artifacts

Aggregate output directory: path:experiments/synthetic_sequences/disentangled/analysis/aggregates/continuous_cvae-Apr24-Apr26-reconstruction-stack-unified_20260426
Combined loss-setting summary: path:experiments/synthetic_sequences/disentangled/analysis/aggregates/continuous_cvae-Apr24-Apr26-reconstruction-stack-unified_20260426/data/combined_loss_setting_summary.csv
Raw-baseline summary: path:experiments/synthetic_sequences/disentangled/analysis/aggregates/continuous_cvae-Apr24-Apr26-reconstruction-stack-unified_20260426/data/raw_baseline_summary.csv
Selected contrasts: path:experiments/synthetic_sequences/disentangled/analysis/aggregates/continuous_cvae-Apr24-Apr26-reconstruction-stack-unified_20260426/data/selected_contrasts.csv
Paper metric bar summary: path:experiments/synthetic_sequences/disentangled/analysis/aggregates/continuous_cvae-Apr24-Apr26-reconstruction-stack-unified_20260426/data/paper_metric_bar_summary.csv
Raw-baseline delta points: path:experiments/synthetic_sequences/disentangled/analysis/aggregates/continuous_cvae-Apr24-Apr26-reconstruction-stack-unified_20260426/data/raw_baseline_delta_points.csv