{ "artifact": "Xyntetik-Kvist-14B", "artifact_files": { "safetensors": { "model-00000.safetensors": "6e7602deb8371e0b1b9e2608c2e4414fe4a4bbd77f19bd835294a5dedf9c3e1a", "model-00001.safetensors": "ca69b78b6f833cd887b428f5394894f79e4102afc332f91bed683c03972f0ce2", "model-00002.safetensors": "2e07b5fb9609dddd7cc4ee3354f10d05695274c06575c63eebbfd8b1c955e8d9", "model-00003.safetensors": "6f108a8e55e00dd9cda4400829217173c3696cc337c8d4cea252d3ce9b1ca30e" }, "gguf": { "Xyntetik-Kvist-14B-BF16.gguf": { "sha256": "daf422db9d0853d53ec8abb8fe9cd3f1e33ea53dffe7cca476225fbe1bc5d5c6", "bytes": 28903141184 }, "Xyntetik-Kvist-14B-Q8_0.gguf": { "sha256": "cdb73543a6fc22f7faa5bbab0176ab2d5b1b590b46870d2d22f2a2510caaacb6", "bytes": 15363224384 }, "Xyntetik-Kvist-14B-Q5_0-mix.gguf": { "sha256": "37af54b4cef4445bbbf23e8fbf0ceec8dc9b01b38dab0dc7f9e2bd2fd79f3a56", "bytes": 10295023424, "note": "requested as llama.cpp Q4_K_M; hidden width 5760 is not a multiple of 256, so K-quant tensors of that row length fall back; named for its majority type" }, "Xyntetik-Kvist-14B-IQ4_NL-mix.gguf": { "sha256": "c74b27a5a39c894c4d1a1a2ae140de0cee125fc5cc9a7c96db9b7355bf8709a3", "bytes": 7612384384, "note": "added 2026-09-28: IQ3_S on ffn_down and attn_output (rows divisible by 256), IQ4_NL on every 5,760-wide tensor, the output head and token embeddings (explicit per-tensor types, no Q4_0); named for its majority type", "types": { "IQ4_NL": 314, "F32": 313, "IQ3_S": 104 }, "share_by_type": { "IQ4_NL": 0.7739, "F32": 0.0006, "IQ3_S": 0.2254 }, "imatrix": { "tool": "llama.cpp b10353 llama-imatrix (static CPU build)", "model": "Xyntetik-Kvist-14B-Q8_0.gguf (sha256 cdb73543a6fc22f7faa5bbab0176ab2d5b1b590b46870d2d22f2a2510caaacb6)", "dataset": "Salesforce/wikitext", "config": "wikitext-103-raw-v1", "split": "validation", "snapshot": "b08601e04326c79dfdd32d625aee71d232d685c3", "parquet_sha256": "204929b7ff9d6184953f867dedb860e40aa69c078fc1e54b3baaa8fb28511c4c", "text_sha256": "8ef749789ca0693435d20b3f81d5638c19edcebc5a68586dcf09bdf47ef9542f", "chunks": 100, "ctx": 512, "flags": "-c 512 --chunks 100 -t 32 --process-output", "imatrix_sha256": "9f1ae3674ca5446e30781c68f3ac89320d1408af62da19a6c11ced72290d9e5b" }, "fidelity_vs_bf16": { "positions_scored": 500, "positions_failed": 0, "mean_kld": 0.003741668625101878, "top1_agreement_pct": 98.6, "top1_margin_qualified_pct": 100.0, "mean_top8_overlap": 0.954, "tie_band_nats": 0.5, "verdict": "PASS" } } }, "gguf_converter": "llama.cpp b10353 convert_hf_to_gguf.py --outtype bf16 (Old-archive/Gridcore/llama.cpp-b10353 on the box)", "quantiser": "llama.cpp b10353 llama-quantize (static CPU build, no imatrix) Q8_0 and Q4_K_M from the block-ordered BF16, then gguf_blockorder.py --order-from the BF16", "layout": "all three files rewritten by gguf_blockorder.py v2: token_embd, blk.0..51, output_norm, output; per-tensor sha256, every KV field and the raw header bytes verified identical (release-a5/layout-rewrite.json)" }, "method": "width pruning of Muse-Glimmer-30B's language model by activation importance (no training), general knowledge distillation from the frozen BF16 parent (top-64 forward KL), then an envelope phase interleaving general windows with agentic examples rendered in Runner's muse template, with a factored teacher at decision positions (Ornith-1.0-9B policy mass over the prefix trie, the parent's top-64 for the remainder) and the parent's top-64 elsewhere", "parent": { "repo": "meta-models/Muse-Glimmer-30B", "revision": "a4e59da52a7bc87ae7251dd5545c0dd437c44b68", "license": "apache-2.0", "files_sha256": { "model-00001-of-00002.safetensors": "8eef61530e1283642c77ce2e6721feb5c6f348fa055c00e90f2844a136372694", "model-00002-of-00002.safetensors": "b58cc2144ba1ba1af4420f67f4ca3ced7f09298510b80464cc75018a0be14381", "tokenizer.json": "c9dbee66967b58f31a7c27f723c3760da3526ccd0427578e8905b0abb0031c4d" }, "sha256_source": "Hugging Face LFS etags recorded by huggingface_hub at download (models/Muse-Glimmer-30B-safetensors/.cache/huggingface/download/*.metadata)", "parameters_total": 29776626688, "parameters_language_model": 27854780928, "parameters_vision": 1921845760 }, "gate_reference_model": { "file": "parent_q4_k.gguf (models/muse-recovered/)", "bytes": 15686266304, "sha256": "0956fcae4fda1e8f93f8784ace121accde2b8db4b00be7d3aa8dc926bf726075", "what": "the parent quantised to Q4_K by Runner's own quantiser; study scorer vs the BF16 parent: KLD 0.01930, margin-q 99.12% (RECOVERY.json of the SR1 release, parent_q4_k_control)", "used_for": [ "E1", "E2", "E3", "E5" ] }, "policy_teacher": { "model": "Ornith-1.0-9B", "repo": "deepreinforce-ai/Ornith-1.0-9B", "tokenizer_snapshot": "83dc1f5e24ef8527af019a6b3bf66ac0f1c2c999", "served_file": "ornith-1.0-9b-Q4_K_M.gguf", "served_file_sha256": "5720d1f671b4996481274fffe01868c3c36e87c135cc8538471cc7bd6087b106", "served_file_commit": "3296bc7a404871a72ac3f1903f561459c09b5c17", "license": "MIT, per the model card at the tokenizer snapshot used (83dc1f5e; the repository has since moved to ornith-ai/Ornith-1.0-9B); recorded as Apache-2.0 on 2026-09-23, both permissive; attribution shipped", "role": "turn-level policy (act / tool / keys) at decision positions; no weights carried" }, "pruning": { "sidecar": "PRUNE.json", "sidecar_sha256": "0356aae40e33a54d4560aae68d7d70d01ea61d873d53db8bc11918664bb4b7ac", "method": "Minitron-style activation importance: global hidden channels by summed mean|x|; FFN neurons per layer by mean|h| at down_proj input; heads per KV group by mean o_proj-input norm; no training", "calibration": "64 x 2048 = 131072 tokens of corpus_run9 train split; eval split never read", "capture_sha256": "6e4f0e30eebe1fd2d2d99ebc4a051c6bbc597c41f547ad4393dc44ca86be8fea", "widths": { "hidden": [ 6656, 5760 ], "ffn": [ 19968, 10240 ], "heads": [ 32, 24 ], "kv_heads": [ 2, 2 ], "layers": [ 52, 52 ] }, "parameters": { "embedding": [ 1344831488, 1163796480 ], "blocks": [ 25165117952, 12116188800 ], "lm_head": [ 1344831488, 1163796480 ], "total": [ 27854780928, 14443781760 ] }, "fraction_of_language_model_kept": 0.5185, "vision_encoder": "not carried (0 of 1921845760 parameters)" }, "general_kd": { "run_record": "evidence/kd/train-w14b-v1/run.json", "resume_chain": "muse-w14b-pruned -> train-w14b-v1/checkpoint-06000", "steps": 6000, "windows_per_step": 4, "predicted_positions_per_window": 4095, "tokens": 98280000, "wall_seconds": 583590.7, "objective": "top-64 forward KD to the frozen BF16 Muse parent, tail bucket, softcapped logits; embed/head frozen; final norm + all blocks trainable", "optimizer": "AdamW betas (0.9, 0.95), fp32 master on host, bf16 moments, per-block clip 1.0, cosine to 10%", "lr_peak": 0.0001, "warmup": 50, "seed": 20260915, "corpus": { "name": "corpus-v1", "sequences": 59946, "seq_len": 4096, "tokens": 245538816, "tokens_sha256": "0731731533ca5ab7be57d607005c8f43d5a26a8c831c551bbf72c0a0ca5f0a7a", "render": "plain text, role prefixes, EOS-joined", "fraction_used": 0.4, "tool_calling_documents": 0, "manifest": "evidence/corpus/corpus-v1.manifest.json" }, "trainer_sha256": "ac7c0352e99b8b1e6b073f6cd2acec78d936833a8cdab2a5e6b18e09e693511b", "dev_proxy": { "sequences": 16, "top64_kl_first_last": [ 1.9223216295533418, 0.5430079750088982 ], "top1_first_last": [ 0.5136446886446886, 0.7521978021978022 ] } }, "envelope": { "run_record": "evidence/envelope//run.json and train.jsonl, one pair per segment", "resume": "train-w14b-v1/checkpoint-06000 (weights and optimizer state) for the first segment; each later segment resumes a checkpoint of the one before", "steps_budgeted": 1500, "steps_run": 1440, "checkpoint_released": "checkpoint-01440", "run_released": "train-w14b-envelope-a12", "segments": [ { "run": "train-w14b-envelope-a2", "resume": "train-w14b-v1/checkpoint-06000", "steps": [ 1, 200 ], "general_steps": 100, "policy_steps": 100, "general_positions": 1638000, "policy_loss_positions": 152602, "wall_seconds": 19906.1, "general_per_policy": 1, "kind_weights": null, "terminator_weight": 1.0, "early_stop": "on", "trainer_sha256": "b0b47549dc2ecf4932698075844f19be879deef8d298e6ca6beac7b7c9761fc0", "note": "attempt 2 (attempt 1 had masked the <|message|> after each recipient; fixed in agentic_targets.py before this run). The early stop fired at step 200 (dev proxy 0.5344, 0.5371, 0.5449, 0.5534); gated there it failed E1 (162/200), E2 and E5, so attempt 3 continued it (2026-09-24 03:35)" }, { "run": "train-w14b-envelope-a3", "resume": "train-w14b-envelope-a2/checkpoint-00200", "steps": [ 201, 350 ], "general_steps": 75, "policy_steps": 75, "general_positions": 1228500, "policy_loss_positions": 110390, "wall_seconds": 14783.1, "general_per_policy": 1, "kind_weights": null, "terminator_weight": 1.0, "early_stop": "off", "trainer_sha256": "ac3df6cd53c1c40abef833c5897c404cbe37e077d14108724d6130719e1ab390", "note": "attempt 3, continued attempt 2 to step 400 with the early stop off; gated at 400 it failed E1 (164/200, every failure an answer after a tool result). Attempt 4 started from its checkpoint-00350, the latest with E4 headroom (0.841 against the 0.865 cap), so steps 351-400 of this run are not in the lineage (2026-09-24 11:45)" }, { "run": "train-w14b-envelope-a4", "resume": "train-w14b-envelope-a3/checkpoint-00350", "steps": [ 351, 450 ], "general_steps": 66, "policy_steps": 34, "general_positions": 1081080, "policy_loss_positions": 54407, "wall_seconds": 10809.5, "general_per_policy": 2, "kind_weights": { "turn:answer": 2, "flight": 2 }, "terminator_weight": 1.0, "early_stop": "off", "trainer_sha256": "4fca29a35dc14f970aef120d040e1a3439b0733d56e727b4ccc08d1db9e8f41a", "note": "attempt 4: answer turns and flight tasks sampled x2, two general steps per policy step. Stopped after checkpoint-00450 without a gate when a 20-prompt probe showed the same failure (5 of 9 well formed); attempt 5 started from checkpoint-00450, the latest with E4 at most 0.85 (0.839) (2026-09-24 14:40)" }, { "run": "train-w14b-envelope-a5", "resume": "train-w14b-envelope-a4/checkpoint-00450", "steps": [ 451, 600 ], "general_steps": 100, "policy_steps": 50, "general_positions": 1638000, "policy_loss_positions": 92803, "wall_seconds": 16987.1, "general_per_policy": 2, "kind_weights": { "turn:answer": 2, "flight": 2 }, "terminator_weight": 10.0, "early_stop": "off", "trainer_sha256": "f28323e0fbfd42ba731837c37cb85e4153c1edc775942b6c9de15698fc2c59ce", "note": "attempt 5: attempt 4's recipe plus terminator positions weighted x10; 150 preregistered steps to 600 (2026-09-24 14:40). Gated at checkpoint-00600 it failed E1 (187/200: 7 reasoning turns closed with <|eot|> on missing_tool first turns, 6 never terminated); attempt 6 continued from checkpoint-00600 (2026-09-24 21:59)" }, { "run": "train-w14b-envelope-a6", "resume": "train-w14b-envelope-a5/checkpoint-00600", "steps": [ 601, 720 ], "general_steps": 80, "policy_steps": 40, "general_positions": 1310400, "policy_loss_positions": 45027, "wall_seconds": 14021.5, "general_per_policy": 2, "kind_weights": { "turn:answer": 2, "flight": 2 }, "terminator_weight": 10.0, "early_stop": "off", "trainer_sha256": "f28323e0fbfd42ba731837c37cb85e4153c1edc775942b6c9de15698fc2c59ce", "note": "attempt 6: on-policy distillation from attempt 5's checkpoint-00600. 300 rollouts of the student's own reasoning (200 at training-set states where a tool result came back and the trajectory answers, 100 at first user turns answered with no tool call), each a loss-bearing span trained toward the frozen parent's top-64; a rollout the student stopped, on either terminator, ends in <|eom|>; mixed 1:1 by weight with the original examples; attempt 5's recipe otherwise; 120 preregistered steps to 720 (registered 2026-09-24 17:00, amended 21:59 and 23:19)" }, { "run": "train-w14b-envelope-a7", "resume": "train-w14b-envelope-a6/checkpoint-00720", "steps": [ 721, 840 ], "general_steps": 80, "policy_steps": 40, "general_positions": 1310400, "policy_loss_positions": 45362, "wall_seconds": 14298.8, "general_per_policy": 2, "kind_weights": { "turn:answer": 2, "flight": 2 }, "terminator_weight": 10.0, "early_stop": "off", "trainer_sha256": "f28323e0fbfd42ba731837c37cb85e4153c1edc775942b6c9de15698fc2c59ce", "note": "attempt 7: on-policy distillation from attempt 6's checkpoint-00720. 390 rollouts of the student's own turns (reasoning after tool results and at first user turns, plus 180 whole-generation call states including find_flight calls), each a loss-bearing span trained toward the frozen parent's top-64; mixed 1:1 by weight with the original examples after the close-mass fix (a call is never taught to close before its schema's required arguments); terminator x10; 120 preregistered steps to 840" }, { "run": "train-w14b-envelope-a8", "resume": "train-w14b-envelope-a7/checkpoint-00840", "steps": [ 841, 960 ], "general_steps": 80, "policy_steps": 40, "general_positions": 1310400, "policy_loss_positions": 52608, "wall_seconds": 14368.1, "general_per_policy": 2, "kind_weights": { "turn:answer": 2, "flight": 2 }, "terminator_weight": 10.0, "early_stop": "off", "trainer_sha256": "f28323e0fbfd42ba731837c37cb85e4153c1edc775942b6c9de15698fc2c59ce", "note": "attempt 8: on-policy distillation from attempt 7's checkpoint-00840. 343 rollouts (reasoning after tool results and at first turns, call states, post-tool calls and answers, first answers, plus answer turns on forecast and weather2 at 640 tokens); rollouts that ran to the length cap kept whole; mixed 1:1; terminator x10; 120 steps to 960" }, { "run": "train-w14b-envelope-a9", "resume": "train-w14b-envelope-a8/checkpoint-00960", "steps": [ 961, 1080 ], "general_steps": 80, "policy_steps": 40, "general_positions": 1310400, "policy_loss_positions": 52910, "wall_seconds": 14087.6, "general_per_policy": 2, "kind_weights": { "turn:answer": 2, "flight": 2 }, "terminator_weight": 10.0, "early_stop": "off", "trainer_sha256": "f28323e0fbfd42ba731837c37cb85e4153c1edc775942b6c9de15698fc2c59ce", "note": "attempt 9: on-policy distillation from attempt 8's checkpoint-00960. 370 rollouts of the same mix without the extra forecast shards; a rollout that ran to the length cap is cut at the last sentence end before its loop onset (the first repeated 5-word span) and closed with its turn's terminator (15 such); mixed 1:1; terminator x10; 120 steps to 1080" }, { "run": "train-w14b-envelope-a10", "resume": "train-w14b-envelope-a9/checkpoint-01080", "steps": [ 1081, 1200 ], "general_steps": 80, "policy_steps": 40, "general_positions": 1310400, "policy_loss_positions": 54335, "wall_seconds": 14048.6, "general_per_policy": 2, "kind_weights": { "turn:answer": 2, "flight": 2 }, "terminator_weight": 10.0, "early_stop": "off", "trainer_sha256": "f28323e0fbfd42ba731837c37cb85e4153c1edc775942b6c9de15698fc2c59ce", "note": "attempt 10: on-policy distillation from attempt 9's checkpoint-01080, attempt 9's recipe plus 10 one-position 'antirepeat' examples: at the first token of a loop's first repeat the target is the parent's top-64 without that token. Six of the ten onsets were a correctly restated calculator value, so this segment taught the student to write a different number there (measured afterwards as '840 x 32 / 6 = 8.40'); 120 steps to 1200" }, { "run": "train-w14b-envelope-a11", "resume": "train-w14b-envelope-a10/checkpoint-01200", "steps": [ 1201, 1320 ], "general_steps": 80, "policy_steps": 40, "general_positions": 1310400, "policy_loss_positions": 55986, "wall_seconds": 15908.7, "general_per_policy": 2, "kind_weights": { "turn:answer": 2, "flight": 2 }, "terminator_weight": 10.0, "early_stop": "off", "trainer_sha256": "f28323e0fbfd42ba731837c37cb85e4153c1edc775942b6c9de15698fc2c59ce", "note": "attempt 11: on-policy distillation from attempt 10's checkpoint-01200, attempt 10's recipe with the antirepeat target skipped on numeric onsets (8 examples kept, 1 skipped) and damaged calculator answers cut at the first wrong number, so their last loss position is where the parent writes the right number (6 such); 370 rollouts, 9 capped rollouts cut at the loop onset; mixed 1:1; terminator x10; 120 steps to 1320" }, { "run": "train-w14b-envelope-a12", "resume": "train-w14b-envelope-a11/checkpoint-01320", "steps": [ 1321, 1440 ], "general_steps": 80, "policy_steps": 40, "general_positions": 1310400, "policy_loss_positions": 56138, "wall_seconds": 15211.9, "general_per_policy": 2, "kind_weights": { "turn:answer": 2, "flight": 2 }, "terminator_weight": 10.0, "early_stop": "off", "trainer_sha256": "f28323e0fbfd42ba731837c37cb85e4153c1edc775942b6c9de15698fc2c59ce", "note": "attempt 12: on-policy distillation from attempt 11's checkpoint-01320, attempt 11's recipe plus 80 calc rollouts at the answer turn after the tool result (640 tokens) and two more cuts on two-decimal calc tasks: an extra example ending at the padding decision (21) and a cut at the first wrong bold number (6), with 12 cuts at a wrong number, 18 antirepeat targets (9 skipped as numeric); 450 rollouts, 27 capped rollouts cut at the loop onset; 488 examples mixed 1:1; terminator x10; 120 steps to 1440" } ], "general_steps": "4 general windows x 4096 tokens from corpus-v3-muse, parent top-64 targets", "policy_steps": "8 policy examples x up to 1536 tokens, sampled in proportion to example weight (times the segment kind weights); loss on assistant positions only; mixture targets at decision positions; terminator positions weighted where the segment says so", "general_per_policy_note": "segments without general_per_policy alternate one general step and one policy step; with it, N general steps per policy step", "lr": 3e-05, "warmup": 20, "seed": 20260922, "general_tokens": 14758380, "policy_loss_positions": 772568, "policy_examples_loaded": 4442, "mean_step_seconds": 104.2, "wall_seconds": 164431.0, "trainer_sha256": "f28323e0fbfd42ba731837c37cb85e4153c1edc775942b6c9de15698fc2c59ce (the last segment; each segment lists its own)", "early_stop_rule": "preregistered: stop if the dev proxy of E4 (8 held-out windows, every 50 steps) rises for three consecutive evaluations. Implemented 2026-09-23 before the first envelope run reached step 200. It fired at step 200 in attempt 1 and in attempt 2; from attempt 3 the preregistration turned it off and read E4 on every checkpoint instead, the gate taking the latest checkpoint inside the E4 cap", "corpus": { "name": "corpus-v3-muse", "sequences": 9325, "seq_len": 4096, "tokens": 38195200, "tokens_sha256": "34a37a0ba8086c1144e437a87a9f02becffe1a2f69ba2afae24ebafb61d52fab", "render": "agentic_render.render_muse_chat (Runner muse template at 676d3a7); wikitext plain; EOS-joined", "contamination_filter": { "window": 64, "forbidden_windows": 87790, "boilerplate_frac": 0.02, "boilerplate_sample_docs": 2000, "exempted_windows": { "tool-calling": 166, "casual-dialogue": 16, "general-knowledge": 3, "all_other_domains": 0 }, "tool_calling": { "docs_used": 5342, "docs_rejected": 425, "sequences": 732 } }, "manifest": "evidence/corpus/corpus-v3-muse.manifest.json" }, "policy_examples": { "file": "data/envelope/examples-muse-all.jsonl", "lines": 3955, "sha256": "6a3b651ccceaed113d95b44785680c321ac3ac7cb701c399ce1988a40e04332e", "bytes": 27279681, "tree_paths_with_examples": 1880, "distinct_tasks": 1715, "e3_exclusion": "19 examples from tasks 02000-02007 removed before training (E3 scores 02000-02059); prior file 3,974 lines, sha256 cb1dd6fa5e6c516d1d1cc70f4e8022fefc1c61826cd5eb84fc100b4d6f154cf4", "tree_task_range": [ 0, 2007 ], "weight_sum": 3507.527, "dedup_note": "data/envelope/DEDUP-README-20260923.txt", "held_out_task_ids": { "file": "data/envelope/holdout-tasks.json", "n": 200, "paths": 223, "seed": 20260922, "present_in_examples": 0 } } }, "measured": { "k_gates": { "instrument": "score_student.py: KLD student||parent over the full vocabulary (mean_kld_house), top-1, margin-qualified top-1 at tie band 0.5 nats, top-8 overlap; frozen BF16 parent (safetensors); 11 held-out sequences, 45056 positions", "note": "the verdict field in each score file is the house quant bar (KLD <= 0.05, margin-q >= 0.97) and is not the K gate", "K1_pruned": { "file": "evidence/kd/K1-w14b-pruned-score.json", "mean_kld": 3.238747178153558, "top1": 0.5704234730113636, "margin_q": 0.6015181107954546, "top8": 0.446685791015625, "rule": "top1 >= 0.20", "verdict": "PASS" }, "K2_step2500": { "file": "evidence/kd/K2-w14b-step2500-score.json", "mean_kld": 1.4405753988433967, "top1": 0.7257856889204546, "margin_q": 0.7700417258522727, "top8": 0.5530811656605114, "rule": "mean_kld <= 1.62 (half of K1)", "verdict": "PASS" }, "read_step3500": { "file": "evidence/kd/mid-w14b-step3500-score.json", "mean_kld": 1.1553246412764897, "top1": 0.7530850497159091, "margin_q": 0.7991166548295454, "top8": 0.5782859108664773 }, "read_step4000": { "file": "evidence/kd/mid-w14b-step4000-score.json", "mean_kld": 1.0353666365315968, "top1": 0.7653364701704546, "margin_q": 0.8109685724431818, "top8": 0.5914362127130682 }, "K3_step6000": { "file": "evidence/kd/K3-w14b-step6000-score.json", "mean_kld": 0.7863921417685394, "top1": 0.7922807173295454, "margin_q": 0.8390447443181818, "top8": 0.6186911843039773, "rule": "margin_q >= 0.80 for a model release", "verdict": "PASS" } }, "envelope_gate": { "preregistration": "evidence/preregistration/ENVELOPE-GATE-PREREG-2026-09-22.md (approved 2026-09-22; owner decisions 2026-09-23)", "headline_order": [ "E4", "E3", "E1", "E2", "E5" ], "E4_retention": { "file": "evidence/gate/treated/E4-train-w14b-envelope-a12-checkpoint-01440.json", "mean_kld": 0.7617267849855125, "top1": 0.7918812144886364, "margin_q": 0.8396883877840909, "top8": 0.6227694424715909, "rule": "mean_kld <= 0.865 and margin_q >= 0.80", "verdict": "pass" }, "E3_closed_loop": { "instrument": "agentic_closedloop.py --raw-muse, greedy, max 6 calls, seed 20260915, offset 2000; runner main 0bfa2ad on the 24 GB GPU slice (parent: automatic fit; both 14B BF16 files: --gpu-layers 38)", "task_set": "offset 2000, 60 tasks", "n_tasks": 60, "reference": { "file": "evidence/gate/reference/reference-e3-gpu-closedloop.json", "success_n": 60 }, "control": { "file": "evidence/gate/control/control-e3-gpu-closedloop.json", "success_n": 0, "timing": "on the GPU slice after the envelope training (owner decision 2026-09-23 13:45); it measures the frozen checkpoint-06000" }, "treated": { "file": "evidence/gate/treated/treated-train-w14b-envelope-a12-e3-gpu-closedloop.json", "success_n": 57 }, "rule": "treated >= control + 15 points and treated >= 0.9 x reference", "min_required_n": 54, "verdict": "pass" }, "E1_E2_E5": { "instrument": "envelope_preflight.py, greedy, 512-token cap, repeat_penalty 1.0, E2 on the first addressed turn after up to 3 reasoning turns; runner main 0bfa2ad", "reference": { "file": "evidence/gate/reference/reference-e1e2e5.json", "scored": 100, "E1": [ 100, 100 ], "E2": [ 53, 53 ], "E5": [ 0, 100 ], "terminator_not_reported": 0 }, "control": { "file": "evidence/gate/control/control-e1e2e5.json", "scored": 200, "E1": [ 0, 200 ], "E2": [ 0, 11 ], "E5": [ 192, 200 ], "terminator_not_reported": 0 }, "treated": { "file": "evidence/gate/treated/treated-train-w14b-envelope-a12-e1e2e5.json", "scored": 200, "E1": [ 199, 200 ], "E2": [ 99, 99 ], "E5": [ 3, 200 ], "terminator_not_reported": 0 }, "rules": { "E1": "well formed >= 0.98 of scored and strictly above the control", "E2": "valid >= 0.95 of calls; argument names match on >= 0.95 of validated calls (holds by construction in this instrument)", "E5": "loop rate <= control's and <= reference's + 5 points" }, "verdicts": { "E1": "pass", "E2": "pass", "E5": "pass" }, "reasoning_length_report": { "what": "a report, not a rule (prereg 2026-09-24 18:36): reasoning-turn tokens on non-call prompts where treated and reference both closed, parent tokenizer", "file": "evidence/gate/treated/reasoning-length-train-w14b-envelope-a12.json", "paired_n": 43, "treated_median": "48", "treated_p90": 86, "reference_median": "52", "reference_p90": 87, "treated_longer_n": 13 } }, "outcome": "all five arms pass" }, "gguf_fidelity_vs_own_bf16": { "instrument": "xyntetik-runner scripts/kld-compare-raw.py, both sides in Runner, tie band 0.5 nats, bar KLD <= 0.05 and margin-q >= 97%", "positions": "500", "corpus": "evidence/muse/harness-eval-split.txt", "Q8_0": { "mean_kld": "0.0001", "top1_pct": "100.00%", "margin_q_pct": "100.00%", "top8": "0.998", "verdict": "PASS" }, "Q5_0-mix": { "mean_kld": "0.0012", "top1_pct": "99.60%", "margin_q_pct": "100.00%", "top8": "0.980", "verdict": "PASS" } }, "benchmark_annex": "pending" }, "runner": { "serves_from": "v0.5.7 (runner main bd20e1f72571510b26fcdc8ad4f7b213c10abeab); v0.5.6 (e3781b9) also loads the files", "E1_E2_E5_measured_on": "runner main 0bfa2ad595b82ddb22247d7d6ca16ee89c215db8", "E3_measured_on": "runner main 0bfa2ad595b82ddb22247d7d6ca16ee89c215db8 (GPU slice)", "E4_measured_on": "no runner (study PyTorch forward on the safetensors checkpoint)", "partial_offload_fix": "per-tensor upload on runner main 8e1db32 (PR #153); per-tensor path chosen up front for this layout on 9b825fc (PR #154); both ship in v0.5.7", "stop_token_report": "runner main 0bfa2ad, after the v0.5.6 tag; ships in v0.5.7", "serving_measurements_on": "runner main 53b4deb (attempt 8 loop guard and reasoning temperature) and ac5418e (released-checkpoint loop guard), both contained in v0.5.7" }, "scope_tags": { "K1_K2_K3": "supported-preregistered", "envelope_gate": "supported-preregistered", "tool_calling_capability": "supported-preregistered, scoped to the E3 environment and the raw completions path", "parent_equivalence": "not-claimed", "no_quality_loss": "not-claimed", "capability_beyond_annex": "not-claimed", "tool_calling_from_general_distillation": "not-claimed (corpus v1 and v2 held no tool-calling documents)" }, "preregistration": [ "evidence/preregistration/PREREG_STUDENTS.md", "evidence/preregistration/ENVELOPE-GATE-PREREG-2026-09-22.md" ] }