From ae6bbd6f16755c1e3fad3f3eb832bd65d5c352dd Mon Sep 17 00:00:00 2001 From: Cursor Agent Date: Wed, 29 Jul 2026 02:06:19 +0000 Subject: [PATCH 1/2] Add Vast multi-seed findings for action-symmetry cycle 2 Document five-seed battery results recovered from Vast after abandoning Flash, including aggregate final-probe MSE and run provenance. Co-authored-by: Alex Vardakostas --- .../battery/findings.md | 33 +++++++ .../battery/multi_seed_summary.json | 96 +++++++++++++++++++ 2 files changed, 129 insertions(+) create mode 100644 experiments/mess3_reward_state_action_symmetry_cycle_2/battery/findings.md create mode 100644 experiments/mess3_reward_state_action_symmetry_cycle_2/battery/multi_seed_summary.json diff --git a/experiments/mess3_reward_state_action_symmetry_cycle_2/battery/findings.md b/experiments/mess3_reward_state_action_symmetry_cycle_2/battery/findings.md new file mode 100644 index 00000000..32692ce9 --- /dev/null +++ b/experiments/mess3_reward_state_action_symmetry_cycle_2/battery/findings.md @@ -0,0 +1,33 @@ +# Cycle 2 battery — Vast multi-seed findings + +Five seeded full-battery runs (`seeds 42–46`) on Vast RTX 4090 boxes +(`delay=0`, `700k` env steps, experiment SHA `e0c9ac5`, harness +`fcccfa9`). Flash/RunPod results from the prior attempt were abandoned. + +Compact results live on the `results` branch under +`experiments/mess3_reward_state_action_symmetry_cycle_2/battery/results/mess3-as-c2-seed{42..46}-*/`. +Seed 42 also completed a full B2 durability upload; seeds 43–46 had compact +results recovered before slow CN egress finished hashing/uploading checkpoints. + +## Final probe MSE (lower is better) + +| seed | variant_1 | variant_2 | variant_3 | +|-----:|----------:|----------:|----------:| +| 42 | 0.002860 | 0.005516 | 0.000648 | +| 43 | 0.003313 | 0.004761 | 0.000662 | +| 44 | 0.000731 | 0.004546 | 0.000437 | +| 45 | 0.002858 | 0.004386 | 0.000861 | +| 46 | 0.003018 | 0.005538 | 0.003780 | +| **mean±std** | **0.00256±0.00104** | **0.00495±0.00054** | **0.00128±0.00141** | + +## Takeaways + +1. **Variant 3 is best on average** (lowest mean final MSE), but seed 46 is an + outlier that collapses the gap (`0.00378` vs ~`0.0004–0.0009` for seeds 42–45). +2. **Variant 2 is consistently worst** across all five seeds (tight stdev). +3. **Variant 1 is intermediate** with one strong seed (44) and otherwise similar + to the mid-`0.003` band. +4. Relative to cycle 1’s delayed setting, this zero-delay / 0.7M-step battery + still shows a clear action-effect ranking on most seeds: `3 < 1 < 2`. + +See `multi_seed_summary.json` for machine-readable aggregates and run IDs. diff --git a/experiments/mess3_reward_state_action_symmetry_cycle_2/battery/multi_seed_summary.json b/experiments/mess3_reward_state_action_symmetry_cycle_2/battery/multi_seed_summary.json new file mode 100644 index 00000000..4f420116 --- /dev/null +++ b/experiments/mess3_reward_state_action_symmetry_cycle_2/battery/multi_seed_summary.json @@ -0,0 +1,96 @@ +{ + "study": "mess3_reward_state_action_symmetry_cycle_2", + "provider": "vast", + "seeds": [ + 42, + 43, + 44, + 45, + 46 + ], + "run_ids": { + "42": "mess3-as-c2-seed42-20260729T002732Z", + "43": "mess3-as-c2-seed43-20260729T002921Z", + "44": "mess3-as-c2-seed44-20260729T004248Z", + "45": "mess3-as-c2-seed45-20260729T005231Z", + "46": "mess3-as-c2-seed46-20260729T010051Z" + }, + "experiment_ref": "e0c9ac503cad2cbb65a3f40865f8369f4acc8419", + "library_ref": "fcccfa94a3a30bc24d53cf7eca76d3d1c82f9e4f", + "note": "Compact results recovered to results branch. Seed42 also completed B2 durability upload; seeds43-46 B2 uploads aborted after compact recovery due to slow CN egress.", + "per_seed": [ + { + "seed": 42, + "run_id": "mess3-as-c2-seed42-20260729T002732Z", + "variant_1_mse": 0.002860448046062251, + "variant_1_fine_mse_ratio": 12.858605450944093, + "variant_2_mse": 0.005515727495577958, + "variant_2_fine_mse_ratio": 12.240160060105984, + "variant_3_mse": 0.000647866496178273, + "variant_3_fine_mse_ratio": 1.5147507618478806 + }, + { + "seed": 43, + "run_id": "mess3-as-c2-seed43-20260729T002921Z", + "variant_1_mse": 0.0033132448592615533, + "variant_1_fine_mse_ratio": 14.530551281067257, + "variant_2_mse": 0.004760574112606016, + "variant_2_fine_mse_ratio": 10.255646104866631, + "variant_3_mse": 0.0006618247894507844, + "variant_3_fine_mse_ratio": 1.5457587853513888 + }, + { + "seed": 44, + "run_id": "mess3-as-c2-seed44-20260729T004248Z", + "variant_1_mse": 0.000730985640257047, + "variant_1_fine_mse_ratio": 3.2750994597996663, + "variant_2_mse": 0.004546089070409274, + "variant_2_fine_mse_ratio": 9.876359929171711, + "variant_3_mse": 0.0004370410097976952, + "variant_3_fine_mse_ratio": 1.0446271038613808 + }, + { + "seed": 45, + "run_id": "mess3-as-c2-seed45-20260729T005231Z", + "variant_1_mse": 0.0028582732543088658, + "variant_1_fine_mse_ratio": 12.823862347161601, + "variant_2_mse": 0.0043861679756017615, + "variant_2_fine_mse_ratio": 15.466828138256021, + "variant_3_mse": 0.0008611550886762781, + "variant_3_fine_mse_ratio": 2.056207350525793 + }, + { + "seed": 46, + "run_id": "mess3-as-c2-seed46-20260729T010051Z", + "variant_1_mse": 0.0030175494619106465, + "variant_1_fine_mse_ratio": 13.981054994024134, + "variant_2_mse": 0.00553842590738169, + "variant_2_fine_mse_ratio": 12.045445835596635, + "variant_3_mse": 0.0037798704606170786, + "variant_3_fine_mse_ratio": 6.659647097932064 + } + ], + "aggregate": { + "variant_1": { + "mean_final_mse": 0.002556100252360073, + "stdev_final_mse": 0.0010369646544970753, + "min_final_mse": 0.000730985640257047, + "max_final_mse": 0.0033132448592615533, + "n": 5 + }, + "variant_2": { + "mean_final_mse": 0.00494939691231534, + "stdev_final_mse": 0.0005438804066042359, + "min_final_mse": 0.0043861679756017615, + "max_final_mse": 0.00553842590738169, + "n": 5 + }, + "variant_3": { + "mean_final_mse": 0.001277551568944022, + "stdev_final_mse": 0.0014068641212186086, + "min_final_mse": 0.0004370410097976952, + "max_final_mse": 0.0037798704606170786, + "n": 5 + } + } +} From 2a2304056b14992b6f69cf44b72418d1012162dd Mon Sep 17 00:00:00 2001 From: Cursor Agent Date: Wed, 29 Jul 2026 08:40:51 +0000 Subject: [PATCH 2/2] Document filtered MSE comparison and final action mixes Add the non-collapsed/non-outlier subset (v1 all, v2 seed 45, v3 seeds 42-45) and greedy action-fraction summary to cycle-2 findings. Co-authored-by: Alex Vardakostas --- .../battery/findings.md | 41 ++++++++-- .../battery/multi_seed_summary.json | 76 +++++++++++++++++++ 2 files changed, 109 insertions(+), 8 deletions(-) diff --git a/experiments/mess3_reward_state_action_symmetry_cycle_2/battery/findings.md b/experiments/mess3_reward_state_action_symmetry_cycle_2/battery/findings.md index 32692ce9..33d8ce9b 100644 --- a/experiments/mess3_reward_state_action_symmetry_cycle_2/battery/findings.md +++ b/experiments/mess3_reward_state_action_symmetry_cycle_2/battery/findings.md @@ -9,7 +9,7 @@ Compact results live on the `results` branch under Seed 42 also completed a full B2 durability upload; seeds 43–46 had compact results recovered before slow CN egress finished hashing/uploading checkpoints. -## Final probe MSE (lower is better) +## Final probe MSE (all seeds) | seed | variant_1 | variant_2 | variant_3 | |-----:|----------:|----------:|----------:| @@ -20,14 +20,39 @@ results recovered before slow CN egress finished hashing/uploading checkpoints. | 46 | 0.003018 | 0.005538 | 0.003780 | | **mean±std** | **0.00256±0.00104** | **0.00495±0.00054** | **0.00128±0.00141** | +## Filtered comparison (non-collapsed / non-outlier runs) + +Keep policies that did not hard-collapse to a single action (plus all v1): +**v1 all seeds**, **v2 seed 45 only**, **v3 seeds 42–45**. + +| arm | n | mean final MSE | range | +|-----|--:|---------------:|------| +| v3 (42–45) | 4 | **0.00065 ± 0.00017** | 0.00044–0.00086 | +| v1 (all) | 5 | **0.00256 ± 0.00104** | 0.00073–0.00331 | +| v2 (45) | 1 | **0.00439** | — | + +Roughly: **v3 ≈ 4× better than v1**, and **v1 ≈ 1.7× better than mixed v2**. + +## Final-checkpoint greedy action mix + +`greedy_action_fractions` = `[noop, pos, neg]` from final probes: + +| variant | typical mix | notes | +|---------|-------------|-------| +| 1 | ~0 / 100 / 0 | always hard-argmax **positive** (5/5) | +| 2 | ~6 / 94 / 0 | almost always **positive**; seed 45 ~27/73/0 | +| 3 | ~21 / 42 / 38 | mixed ~26/37/37 on seeds 42–45; seed 46 ~0/61/39 | + +So v1/v2 mostly collapse to “always positive”; only v3 keeps a real action mix +(on the retained seeds). Oracle policies differ by variant +(`v1: +++`, `v2: ++noop`, `v3: +−noop`). + ## Takeaways -1. **Variant 3 is best on average** (lowest mean final MSE), but seed 46 is an - outlier that collapses the gap (`0.00378` vs ~`0.0004–0.0009` for seeds 42–45). -2. **Variant 2 is consistently worst** across all five seeds (tight stdev). -3. **Variant 1 is intermediate** with one strong seed (44) and otherwise similar - to the mid-`0.003` band. -4. Relative to cycle 1’s delayed setting, this zero-delay / 0.7M-step battery - still shows a clear action-effect ranking on most seeds: `3 < 1 < 2`. +1. On the filtered set, **variant 3 remains clearly best** for belief MSE. +2. **Variant 2 is worst**, and when it does not fully collapse (seed 45) MSE is + still ~0.0044. +3. **Variant 1** sits in between and is always action-collapsed to positive. +4. Ranking on retained runs: **3 ≪ 1 < 2**. See `multi_seed_summary.json` for machine-readable aggregates and run IDs. diff --git a/experiments/mess3_reward_state_action_symmetry_cycle_2/battery/multi_seed_summary.json b/experiments/mess3_reward_state_action_symmetry_cycle_2/battery/multi_seed_summary.json index 4f420116..24bc9de7 100644 --- a/experiments/mess3_reward_state_action_symmetry_cycle_2/battery/multi_seed_summary.json +++ b/experiments/mess3_reward_state_action_symmetry_cycle_2/battery/multi_seed_summary.json @@ -92,5 +92,81 @@ "max_final_mse": 0.0037798704606170786, "n": 5 } + }, + "filtered_selection": { + "variant_1_seeds": [ + 42, + 43, + 44, + 45, + 46 + ], + "variant_2_seeds": [ + 45 + ], + "variant_3_seeds": [ + 42, + 43, + 44, + 45 + ], + "rationale": "Retain all v1; v2 only non-collapsed seed 45; v3 drop outlier seed 46" + }, + "filtered_aggregate": { + "variant_1": { + "mean_final_mse": 0.002556100252360073, + "stdev_final_mse": 0.0010369646544970753, + "min_final_mse": 0.000730985640257047, + "max_final_mse": 0.0033132448592615533, + "n": 5, + "values": [ + 0.002860448046062251, + 0.0033132448592615533, + 0.000730985640257047, + 0.0028582732543088658, + 0.0030175494619106465 + ] + }, + "variant_2": { + "mean_final_mse": 0.0043861679756017615, + "stdev_final_mse": 0.0, + "min_final_mse": 0.0043861679756017615, + "max_final_mse": 0.0043861679756017615, + "n": 1, + "values": [ + 0.0043861679756017615 + ] + }, + "variant_3": { + "mean_final_mse": 0.0006519718460257577, + "stdev_final_mse": 0.00017326937386526962, + "min_final_mse": 0.0004370410097976952, + "max_final_mse": 0.0008611550886762781, + "n": 4, + "values": [ + 0.000647866496178273, + 0.0006618247894507844, + 0.0004370410097976952, + 0.0008611550886762781 + ] + } + }, + "final_greedy_action_fractions": { + "note": "[noop, pos, neg] from final_probe", + "variant_1_mean": [ + 0.0, + 1.0, + 0.0 + ], + "variant_2_seed45": [ + 0.275, + 0.725, + 0.0 + ], + "variant_3_seeds_42_45_approx": [ + 0.26, + 0.37, + 0.37 + ] } }