{"accepting_submissions":false,"authority":"3301134b38401196d006a621ac4a772bb4b0e6afa15a7d34a0d1ae5f2c630bcd","chain_weight_submission":null,"chain_writer_enabled":null,"checkpoint":"44fb50bb6bebf1d43136a1187441ae11d7f4f05af260d61a5b8646360986b6d5","current_url":"https://00523074f51300584834607253cae0fa.r2.cloudflarestorage.com/affine-verification-mock-20260930/public/streams/nonpayable-live-reward-math-v1-/current.json?X-Amz-Algorithm=AWS4-HMAC-SHA256&X-Amz-Credential=07a5e4531d1081d827dc236f0278dc59%2F20261029%2Fauto%2Fs3%2Faws4_request&X-Amz-Date=20261029T054536Z&X-Amz-Expires=86400&X-Amz-SignedHeaders=host&X-Amz-Signature=18fa0f0f17e7ae0fcf5cdaea0bfd05d45cf694307bb1681e79baa1437592df4f","deadline":1793253574,"epoch":"affine-mimo9b-grpo-20261013-e000214","epoch_id":"affine-mimo9b-grpo-20261013-e000214","expires_at":1793339136.6226807,"group_contract":{"group_size":8,"min_completed_failure":1,"min_success":1,"outcome_policy":"completed-eos-math-unresolved-failure-v2","version":"complete-canonical-group-v1"},"max_attempts":8,"max_batches":9,"max_output_tokens":8192,"max_output_tokens_by_environment":{"affine_math":8192},"min_completed_failure":1,"min_success":1,"netuid":120,"network":"Finney","nonce_max":7,"nonce_min":0,"notice":"The signed upload deadline has passed. Watch for the next epoch.","payable":true,"prospective_reward_eligible":true,"required_client_features":["canonical-group-nonces-v1","complete-canonical-groups-v1","retained-unresolved-attempts-v1","sampler-selected-behavior-logprobs-v1"],"run_id":"affine-mimo9b-grpo-20261013","samples_per_batch":8,"sampling_contract":{"calibration":{"backend_profile_sha256":"38778ba9ee364d5c199a06be470b25484957f32c4ba55765e6c095ac881915a0","cdf_abs_error":0.0,"checkpoint":"44fb50bb6bebf1d43136a1187441ae11d7f4f05af260d61a5b8646360986b6d5","harness_sha256":"9245ecb73e860e09fc3b6d0579ff5d03da51f370919fe5050f1bb350740de059","logprob_atol":0.0,"model_runtime_revision":"qwen3_5-mimo-v1","report_sha256":"e57c0f986cc56da6779d8e4c8a14266b0aae5964544dde2d2e3118ab35686b58","toploc_exp_mismatches":0,"toploc_mant_err_mean":0.0,"toploc_mant_err_median":0.0,"version":"cached-prefill-calibration-v1"},"generation":"cached-eager-inverse-cdf","randomness":"c3895a219004453f2e3ab6f2a538bee1fe2313d7e3ab510121d44526c26efff7","support_adjudication":"exact-cached-replay-v1","verification":"prefill-cdf-calibrated","version":"forced-inverse-cdf-complete-group-mimo-v1"},"sampling_version":"forced-inverse-cdf-complete-group-mimo-v1","schema":2,"source_bundle_sha256":"a2a03cb345ffe041abdc55be3190ed043b3822e0309869ef2cff4636309609f1","start":1793252674,"status":"epoch_closed","status_updated_at":1793256881.3323567,"submission_transport_policy":"complete-group-commitment-v1","training":{"batch":"K=1: only the turn UID's admitted groups are trained (see k1); the pooled max_groups budget and group-size rule do not apply to the turn batch","collection_seconds":900,"epoch":"affine-mimo9b-grpo-20261013-e000214","epoch_index":214,"generated_at":1793256873.6229143,"head":{"checkpoint":"44fb50bb6bebf1d43136a1187441ae11d7f4f05af260d61a5b8646360986b6d5","optimizer_step":111},"k1":{"active_this_epoch":true,"caps":{"current_protocol_note":"this protocol still collects group_size rollouts per task and at most max_batches tasks per miner per epoch (see the signed manifest), so a turn is <= max_batches x group_size rollouts","max_rollouts":1000,"min_tasks":8,"per_rollout_max_tokens":"<= the gather (harness) max_output_tokens of the env","rollouts_per_task":[2,64]},"enabled":true,"from_epoch":196,"hotkeys_unserved":620,"learning_rate_choice":"fixed by the validator (miner-chosen LR: later)","probe":{"from_epoch":196,"rule":"head pre-probe on the private sample and candidate post-probe on the sample minus the turn batch, each task sampled samples_per_task times; both sides use the same per-request sampling seeds (seed = sha256(plan_seed:env:index:j), one private plan seed per turn shared by head and candidate)","samples_per_task":4,"seeded_common_random_numbers":true},"probe_sample_size":1000,"queue_chain_block":9387760,"queue_length":22,"queue_next":[{"hotkey":"5EgYgLYGbJ4ZQ3gy2SQK8tpE2dP9CQ473nrYsaJsPn7kmKGr","last_seal_tasks":9,"line":0,"position":91,"ready_last_seal":true,"registration_block":9202156,"turns":0,"uid":1},{"hotkey":"5HTcciAqijALVQRhnu49LH7m9gETV6f6Zc1n3YDtdTA6V4QM","last_seal_tasks":9,"line":1,"position":92,"ready_last_seal":true,"registration_block":9202159,"turns":0,"uid":177},{"hotkey":"5CNDpj65UXiQtZvstvt3WMzWZdZQPMqwSSXjvicdDGNwTU4s","last_seal_tasks":9,"line":2,"position":93,"ready_last_seal":true,"registration_block":9202578,"turns":0,"uid":183},{"hotkey":"5CJC1N7B3EFFBkxB7zFjgAiahyWoyTXJUDf8kr11vunftvwE","last_seal_tasks":9,"line":3,"position":94,"ready_last_seal":true,"registration_block":9202948,"turns":0,"uid":155},{"hotkey":"5DqGzuknXR7gCs96bvvNqXpFqbxUMGwbyrjQ6RwpVrt7iakF","last_seal_tasks":9,"line":4,"position":95,"ready_last_seal":true,"registration_block":9203750,"turns":0,"uid":140},{"hotkey":"5D7xPT6baLrpnYsWNHTp7JqLKhrVPx1whLKUUzRfvwV3iThs","last_seal_tasks":9,"line":5,"position":102,"ready_last_seal":true,"registration_block":9206555,"turns":0,"uid":16},{"hotkey":"5GjRgmTa1QR7J6Ksc3itTSQPEZtqHvVSjkxXM21q8hPUdhzd","last_seal_tasks":9,"line":6,"position":103,"ready_last_seal":true,"registration_block":9206949,"turns":0,"uid":196},{"hotkey":"5FeFdnM4vyCEsiASXCyaDZNdd1N57LRByjUiEzFnMENxTveH","last_seal_tasks":9,"line":7,"position":104,"ready_last_seal":true,"registration_block":9207363,"turns":0,"uid":101},{"hotkey":"5CHAWfhWywxrKb7Q7cWC8F4wLDm5WHnaQvfNE2Ck7t67tpVG","last_seal_tasks":9,"line":8,"position":122,"ready_last_seal":true,"registration_block":9217892,"turns":0,"uid":230},{"hotkey":"5DPa9CpCsjVJUdAgYqxMYZmbRLxRccL4PmYQimDr2PtLNAhK","last_seal_tasks":9,"line":9,"position":141,"ready_last_seal":true,"registration_block":9228431,"turns":0,"uid":251},{"hotkey":"5CGm48TeBiqVZaLWTQ6sVyBgsTVi5yG7DiX7uAFcPvkE6dLi","last_seal_tasks":9,"line":10,"position":158,"ready_last_seal":true,"registration_block":9238234,"turns":0,"uid":55},{"hotkey":"5C4hzqtkHC8evzDwqox7afxNaqdDQxsgDNit7SazozcwqGpD","last_seal_tasks":9,"line":11,"position":166,"ready_last_seal":true,"registration_block":9244753,"turns":0,"uid":120},{"hotkey":"5C5s62F3eEMNuexBEShCSCa2anBUrzkAiTNHSEgfRg4c75VW","last_seal_tasks":9,"line":12,"position":173,"ready_last_seal":true,"registration_block":9248098,"turns":0,"uid":154},{"hotkey":"5C4iWWaHdMpbYDduRd2DAdy7YqBAZGqkpjJMr18MUGsozNHU","last_seal_tasks":9,"line":13,"position":188,"ready_last_seal":true,"registration_block":9254754,"turns":0,"uid":93},{"hotkey":"5C4pkW5d5Nn4swiKfwnR7r8i3hcvnBczzFgw5FQmW4nPho3Q","last_seal_tasks":9,"line":14,"position":198,"ready_last_seal":true,"registration_block":9259147,"turns":0,"uid":129},{"hotkey":"5C4rgLw7yRF4888iY2HSUdooNcjVGukSrWL5Cm5LGiXh8fuP","last_seal_tasks":9,"line":15,"position":206,"ready_last_seal":true,"registration_block":9263258,"turns":0,"uid":213},{"hotkey":"5GTxPq1n4Zv6JUkkre7pqZD2vC5MA32jB4KTqNtge7QgWwN4","last_seal_tasks":9,"line":16,"position":213,"ready_last_seal":true,"registration_block":9266582,"turns":0,"uid":153},{"hotkey":"5C4hzNknvafQd7fPZLn11t3MAo57v2CLEWNCExPkAK5ZAw5V","last_seal_tasks":9,"line":17,"position":238,"ready_last_seal":true,"registration_block":9277448,"turns":0,"uid":258},{"hotkey":"5C4kbdxxrFQFRUtvTJTjGxYVZGhtg32vFc5LVJzhEzYz2wBe","last_seal_tasks":9,"line":18,"position":239,"ready_last_seal":true,"registration_block":9277450,"turns":0,"uid":259},{"hotkey":"5C4jMYe9YuupMqKdecwJ4mhwSSL9WpoL46y88EBZwppD235z","last_seal_tasks":9,"line":19,"position":240,"ready_last_seal":true,"registration_block":9277452,"turns":0,"uid":260}],"queue_readiness_epoch":214,"queue_rule":"hotkey-once","queue_rule_from_epoch":200,"registrations_served":[{"epoch_index":196,"hotkey":"5DhiK5Wagjks3MDz5YuCrv5WKFYvXEjryzQg7tqiKkYTMHwX","reg_block":null,"uid":231},{"epoch_index":197,"hotkey":"5C4hrfsWvwpbtqdFpgSzmZFaXjdhssjsb3yBZCrwJS34EoKn","reg_block":null,"uid":33},{"epoch_index":198,"hotkey":"5D1BNkoefhy1XKUsvdHne9cfgEudyXuyGJRULhAFpWzY6SMA","reg_block":null,"uid":255},{"epoch_index":199,"hotkey":"5EFX7zRATrDNBiqF5AHbtfpeJvqiU6cBHugvdyT9X3U1qa53","reg_block":9160307,"uid":80},{"epoch_index":200,"hotkey":"5Hpx67THjBbmB91kYx1P415ffwKVKC3t9t17Sg4GNKE3cpoJ","reg_block":9188113,"uid":217},{"epoch_index":201,"hotkey":"5EFnGX4aozLUGyCwfF1rTRBCHnmFk6cWkVag8rnKpJZ2EjyW","reg_block":9180331,"uid":105},{"epoch_index":202,"hotkey":"5EUrwxXrwz84QTpVmqr9xNeMZ3RyoE9tEzoyFx1Fnf1sGvDm","reg_block":9188539,"uid":147},{"epoch_index":203,"hotkey":"5CS4JJzjvp3YNZJmEkLsEFvBMBEMds2aD6mJnhSA37vB7ucn","reg_block":9190447,"uid":150},{"epoch_index":204,"hotkey":"5C4j8bBZym65xKgyypLB1JgEwxBwQGts29Ycmfk7Tr9mGs98","reg_block":9191138,"uid":100},{"epoch_index":205,"hotkey":"5CUGwGZDRgGr7Lu6QjiggM1qD9v41ffu4PC2HK1Hig6keh81","reg_block":9192523,"uid":92},{"epoch_index":206,"hotkey":"5Ck1HpYfRjZVcD5Hy9XRev681R5e2a1DsGAsJV71yewwmdaq","reg_block":9195194,"uid":165},{"epoch_index":207,"hotkey":"5EVhkYbyfHcBD4r2FBKrHq2hcHSWvirCWy6dY3YSLftkUyXQ","reg_block":9195628,"uid":203},{"epoch_index":208,"hotkey":"5CQfiSRF536M23EnCZMAGmFH72MLuA1RQDC2g9DqfaPjWZc7","reg_block":9165744,"uid":3},{"epoch_index":209,"hotkey":"5DPTAFuBjM5yZhccpCY3DtD1Br4YuRhHnbAmUhU323QV4uq4","reg_block":9166188,"uid":86},{"epoch_index":210,"hotkey":"5EtXM4ChQuNUPPg2aWH12bPWTCEacXB218pmBfS3jDkR23vF","reg_block":9169559,"uid":13},{"epoch_index":211,"hotkey":"5C4i2Ufbe2Fr7ybwArFCfhP6cE5zL87KVdywZgeb61V6MhKG","reg_block":9169982,"uid":98},{"epoch_index":212,"hotkey":"5CpUBXB6anty3imPcV1MuCMsG5fd3TG2DJwNx3Q7R8msNbEF","reg_block":9199056,"uid":172},{"epoch_index":213,"hotkey":"5G28TkWtkK8GDPPcKR86ShK2yrdDyUsFNxZW6NgNZirKfJVa","reg_block":9201427,"uid":88},{"epoch_index":214,"hotkey":"5Cc8nqpYBTo4ymkW5e7Yd1kgNvNVw1i9LHvzoNnDu2vXbjN4","reg_block":9202148,"uid":171}],"registrations_unserved":620,"rule":"K=1 per-hotkey turns (Jacob t712u 2026-10-27; one turn per hotkey, permanently, t761u 2026-10-28, as teutonic.ai): each epoch trains ONE hotkey's batch. Each HOTKEY gets exactly one turn, permanently (pass or fail), on any UID or registration block. Turn order = registration block, then uid (UID 0 excluded); hotkeys that already had a turn (registrations_served / served_hotkeys) leave the queue. At the seal the first unserved hotkey with a ready batch (>= min_tasks distinct admitted tasks) takes the turn; others are skipped, not charged, and keep their place (no waiting). To try again, register a NEW hotkey. Do not deregister a paying hotkey: pay follows the GREEN UID while its earning hotkey stays registered there. queue_next = unserved hotkeys that had a ready batch at the last seal (forecast; the seal decides). The turn batch = that UID's admitted groups only, deterministically trimmed in submission order (received_at, slot) to max_rollouts. No TOPLOC, score or reward checks on miner data and no bans. One stateless step (optimizer below) -> candidate -> promotion gate. GREEN credit: 1/N to each of the last up to 10 distinct GREEN UIDs; RED/skipped earn nothing; unearned weight burns to UID 0.","served_hotkeys":["5C4hrfsWvwpbtqdFpgSzmZFaXjdhssjsb3yBZCrwJS34EoKn","5C4i2Ufbe2Fr7ybwArFCfhP6cE5zL87KVdywZgeb61V6MhKG","5C4j8bBZym65xKgyypLB1JgEwxBwQGts29Ycmfk7Tr9mGs98","5CQfiSRF536M23EnCZMAGmFH72MLuA1RQDC2g9DqfaPjWZc7","5CS4JJzjvp3YNZJmEkLsEFvBMBEMds2aD6mJnhSA37vB7ucn","5CUGwGZDRgGr7Lu6QjiggM1qD9v41ffu4PC2HK1Hig6keh81","5Cc8nqpYBTo4ymkW5e7Yd1kgNvNVw1i9LHvzoNnDu2vXbjN4","5Ck1HpYfRjZVcD5Hy9XRev681R5e2a1DsGAsJV71yewwmdaq","5CpUBXB6anty3imPcV1MuCMsG5fd3TG2DJwNx3Q7R8msNbEF","5D1BNkoefhy1XKUsvdHne9cfgEudyXuyGJRULhAFpWzY6SMA","5DPTAFuBjM5yZhccpCY3DtD1Br4YuRhHnbAmUhU323QV4uq4","5DhiK5Wagjks3MDz5YuCrv5WKFYvXEjryzQg7tqiKkYTMHwX","5EFX7zRATrDNBiqF5AHbtfpeJvqiU6cBHugvdyT9X3U1qa53","5EFnGX4aozLUGyCwfF1rTRBCHnmFk6cWkVag8rnKpJZ2EjyW","5EUrwxXrwz84QTpVmqr9xNeMZ3RyoE9tEzoyFx1Fnf1sGvDm","5EVhkYbyfHcBD4r2FBKrHq2hcHSWvirCWy6dY3YSLftkUyXQ","5EtXM4ChQuNUPPg2aWH12bPWTCEacXB218pmBfS3jDkR23vF","5G28TkWtkK8GDPPcKR86ShK2yrdDyUsFNxZW6NgNZirKfJVa","5Hpx67THjBbmB91kYx1P415ffwKVKC3t9t17Sg4GNKE3cpoJ"],"start_head":"30a129a9ab3bb361253b49a8f86192e2bc18ab89ebf182e52cbfddbb8d6242fa","transparency":{"dev_split":"https://affine.io/k1/dev/README.md","gate_reveal":"https://affine.io/gate/reveal/<epoch>.json (after the decision: sample seed, the 1000 sampled task ids, commitment check, plan seed, per-task head and candidate grades; verify with https://affine.io/gate/reveal/verify_gate_reveal.py)","replay":"https://affine.io/k1/replay/README.md (replay_k1.py + our own replay results)","step_recipe":"https://affine.io/k1/spec/step-recipe.json","turn":"https://affine.io/k1/turns/<epoch>/turn.json (after the gate decision: decision, the selected signed dispatch rows with native grades, the miner batch documents byte-exact under /k1/turns/<epoch>/docs/, trainer report: per-rollout loss/advantage/kept tokens, pre-clip norm, micro-batch plan, optimizer and per-tensor cap stats, candidate checkpoint id + file sha256 for GREEN and RED)","turns_index":"https://affine.io/k1/turns/index.json"},"turn_batch_published":"after the gate decision the trained batch is public at https://affine.io/k1/turns/<epoch>/turn.json: the selected rows of the signed filtered dispatch (task index, native grades, document sha256) and the miner documents byte-exact under /k1/turns/<epoch>/docs/; task indices also in https://affine.io/k1.json turns[].task_indices. The R2 dispatch object itself is not publicly readable.","turns_and_credit":"https://affine.io/k1.json"},"loss":{"aggregation_source":"trainer host engine (loss_agg_mode, loss_scale) as reported by the trainer for epoch 213; the signed trainer_config still names seq-mean-token-sum-norm, which the host overrides","cap_policy":"affine-unresolved-zero-reward-active-v1","clip_ratio":0.2,"derail_mask":true,"derail_tail":64,"derail_threshold":-3.0,"derail_window":32,"dual_clip":3.0,"entropy_coefficient":0.0,"execution":{"deterministic":false,"engine":"packed varlen data-parallel (first-fit-decreasing micro-batches, LPT over ranks, fp32 accumulators SUM-reduced over NCCL), bf16 weights, fp32 logits/loss, layer activation checkpointing","kernels":"fla 0.6.0 chunk_gated_delta_rule + causal_conv1d (Triton, varlen), torch varlen_attn (FlashAttention-2) for full attention; https://affine.io/k1/replay/host/v83_fast_kernels.py","microbatch":{"count":8,"engine":"packed-dp-v83","max_rows":256,"rank_tokens":[32762,32756,32754,32734,32631,32453,32430,27845],"token_budget":32768,"world":8},"note":"not bit-deterministic run to run (Triton/FA kernels, NCCL reduce order); see https://affine.io/k1/replay/README.md","runtime_versions":{"toploc":"0.1.6","torch":"2.14.0","transformers":"5.14.1"}},"formula":"L = loss_scale * (1/N) * sum_i (1/(K_i + 1e-8)) * sum_{t<=K_i} [ pg_{i,t} + kl_coefficient * k3_{i,t} ], N = active rollouts in the turn (all of them: unresolved rollouts stay in with reward 0), K_i = rollout i's kept response tokens. pg = vanilla dual-clip PPO with OLD = current.detach() (one PPO epoch, ratio 1), so its gradient is -A_i * grad log pi. k3 = exp(ref - cur) - (ref - cur) - 1 against the frozen kl_reference_checkpoint (ref - cur clamped to [-20, 20], k3 to [-10, 10]). A_i = r_i - mean(r over the task's group), no std normalisation. r = 1[correct] - overlong_penalty * clip((len - overlong_soft_start) / (overlong_max_tokens - overlong_soft_start), 0, 1). Derail mask: only for rollouts stopped at max_tokens with negative advantage, K_i is cut at the derail index (derail_window / derail_threshold / derail_tail; subnet/mimo_grpo/training.py derail_keep). Code: pinned verl core_algos agg_loss / compute_policy_loss_vanilla / kl_penalty_forward (subnet/mimo_grpo/upstream.py) + https://affine.io/k1/replay/host/v8211_batched.py","kl_coefficient":0.02,"kl_reference_checkpoint":"87de8304e9d33dde39b81a41fdbbf306426ea6fd993b7d85332e6a48e8a4bbae","kl_type":"low_var_kl","loss_agg_mode":"seq-mean-token-mean","loss_scale":0.5635,"loss_scale_factor":null,"miner_logprobs":"not used by the trainer: current log-probs are recomputed by the trainer from the head (OLD = current.detach()) and reference log-probs from kl_reference_checkpoint; uploaded logprobs only serve the capture protocol","normalize_advantage_std":false,"objective":"GRPO: vanilla dual-clip PPO, seq-mean-token-mean x 0.5635, k3 KL to reference, overlong soft penalty","overlong_max_tokens":8192,"overlong_penalty":0.5,"overlong_soft_start":6000,"ppo_epochs":1,"sampler_correction":"none","signed_config_loss_agg_mode":"seq-mean-token-sum-norm"},"optimizer":{"betas":[0.9,0.999],"eps":1e-08,"global_update_norm_cap":1.0,"implementation":"fresh-adam-one-step-from-zero-state-v833b (bit-exact with torch.optim.AdamW(foreach=False), one step, zero moments)","learning_rate":5e-07,"learning_rate_rule":"fixed by the validator (1e-6 x restart scale 0.5); miner-chosen LR is not open yet","max_gradient_norm":1.0,"name":"AdamW","optimizer_state_persisted":false,"replay":"g = the fp32 gradient of the loss below over the turn batch (packed varlen micro-batches on 8 GPUs, fp32 per-rank accumulators SUM-reduced over NCCL), clipped to global L2 norm max_gradient_norm. Per tensor (fp32 from the bf16 head W_t): d_t = learning_rate * (g_t / (|g_t| + eps) + weight_decay * W_t * [W_t >= 2-D and not a bias]); r_t = min(1, tensor_relative_cap * (1 - 1e-6) * ||W_t||_2 / ||d_t||_2) (fp64 norms); s = min(1, global_update_norm_cap / ||r*d||_2) (not binding); W_t_new = bf16_round_nearest_even(fp32(W_t) - fp32(s * r_t) * d_t). With zero moments and bias correction the betas cancel. Exactly one step per turn, nothing carried between turns, so W_new is determined by the published head, the published batch (https://affine.io/k1/turns/<epoch>/turn.json) and these numbers up to GPU arithmetic: the kernels (fla 0.6.0 Triton, flash attention) and the NCCL reduction order are not bit-deterministic, so an independent replay matches closely but is not guaranteed bit-exact. Replay tool and our measured closeness: https://affine.io/k1/replay/README.md","steps_per_epoch":1,"tensor_relative_cap":2e-05,"weight_decay":0.01,"weight_decay_policy":"applied to >= 2-D non-bias tensors only"},"promotion_gate":{"cap_hit_max":0.2,"decisions":"https://affine.io/gate/index.json","enabled":true,"env":"affine_math","first_gated_epoch":152,"min_tasks":200,"red_credit":"scored","red_scored_from_epoch":169,"reveal":"after each decision the sample seed, the sampled task ids and the per-task head/candidate grades are published at https://affine.io/gate/reveal/<epoch>.json (index: https://affine.io/gate/reveal/index.json); the seed is os.urandom(32) drawn fresh for every epoch, so a reveal says nothing about the next sample","rule":"candidate promoted only if paired z (candidate vs head, k=1) > z_min (>= z_min if z_inclusive) and cap-hit rate < cap_hit_max on a private sample minus the epoch batch (from epoch 168: z_min 0, strict = net must be positive, candidate pass rate > head pass rate); RED = discarded, next epoch mines from the same head, K=1: the head is pre-probed and the candidate post-probed on the same private sample (sample_size tasks) minus the turn batch tasks, 4 samples per task with shared per-request seeds (head and candidate); GREEN = paired z > 0 (strict) and cap-hit < 0.20 -> promoted and the turn UID earns; RED = head kept, the turn earns nothing; no fixed-reference guard","sample_pool":7496,"sample_size":1000,"scores_update_on":"pass_and_fail","z_inclusive":false,"z_min":0.0,"z_rule_from_epoch":168},"sampling":{"harness":{"affine_math":{"enable_thinking":true,"max_context_tokens":12288,"max_output_tokens":8192}},"temperature":1.0,"top_k":20,"top_p":0.95},"schedule":{"drift_cap_rule":"per-tensor relative cap tensor_relative_cap (same 2e-5 as the pooled AdamW drift cap)","kind":"constant","learning_rate":5e-07,"restart_rule":"none (every turn is one step from zero optimizer state)"},"sources":{"lifecycle_config_sha256":"c38a9fcf21a1d2a6ef8d762e72fb895e160aa893aba93aa11e509639dda07982","trainer_host":"host-v833c","trainer_host_operator_config_sha256":"572579cfdbe69b79f76d16e1587979516ecb3b1021349a2adb2aba8f005b37f6"},"version":"sn120-public-training-params-v1"}}