Prove trained native 32k branch and 64k request behavior

Gap and evidence

Limits declare 32,768 state-plus-branch and 65,536 aggregate tokens, but acceptance status and validation mark trained, calibrated native exposure and inference unverified. A config maximum or byte-tokenizer fixture is insufficient.

Blocked by #3; related tokenizer defect: #2.

Acceptance

Record a pretrained native artifact, parent/training-exposure chain and hashes proving actual optimization at both 32,768 branch and 65,536 aggregate limits, followed by calibration on separate data.
Run exact-length requests and over-limit rejection with the pinned tokenizer and trained artifact; no silent truncation, dense-kernel fallback or state double-counting.
Cover evidence at beginning/middle/end, unequal branches, mixed Choice/Score/Noul, distractors, 255 choices and 10 score levels.
Report tokenizer length, serialized length, training exposure, inference result, peak GPU memory and latency separately, including any hardware feasibility blocker. Keep parity unknown until all gates pass.

Workstation harness delivery - 2026-09-27

PR #15 delivered the native/context harness guards, with verified SSH-signed source commit 150a10bcdf9cca16cfd18de1e60abd84b539b02c and verified merge commit f9304ec25f8e9722654f4cb4d3a2e1d4051f6b2d. The clean primary checkout is synchronized with main. The package's earlier uncommitted status and denied issue update were historical; this is the subsequent workstation delivery.

The report is versioned native-context-v3. Exact branch and aggregate ContextOverflow details are required; unrelated exceptions, omitted branch outputs, broadcastable shapes, nonfinite tensors and invalid temperatures cannot supply a passing result. The oracle checks installed/imported reference-version consistency and complete hidden tensors. CUDA measurement APIs use the selected device. Integration review also corrected BF16 autocast construction/execution to enter that selected device context, with three regressions for success, constructor failure and model failure that verify restoration of the caller's device. Tolerances and the 0.80 diagnostic threshold are unchanged; the separate 0.70 fixture threshold and known case=99999 failure are unchanged.

The final complete Windows suite passed 621 tests with five explicit skips: official SDK, two CUDA tests in the CPU environment, and two Windows symlink-privilege cases. The real pinned tokenizer snapshot was supplied. The 76 focused harness tests are included in that total. Source integrity, compilation and pinned critical Ruff checks passed. PR CI passed all five jobs, with 622 passed / four skipped in each Ubuntu/Windows Python 3.12/3.13 suite, followed by the numerical, resume/calibration/monitoring, wheel, HTTP and schema gates. Post-merge CI also passed on the merge commit.

The model environment initially resolved a historical installed Layev wheel; it was refreshed to an editable install of the current checkout without dependency changes and the full suite was rerun. The direct current-source native command exits 1 at its CUDA prerequisite, executes zero native cases and creates no output directory. This environment still has CPU-only PyTorch; that does not imply the workstation lacks CUDA-capable hardware.

These are acceptance-harness software results and explicit model/oracle/device doubles, not actual pretrained execution or measured trained context. All four original acceptance boxes were unchecked at this 2026-09-27 readback, and the issue remained open, retaining its dependency on #3 and relationship to #2. Real pinned weights, trained/calibrated exposure, exact native requests, measured GPU memory/latency and any hardware limits still need evidence. No Overwatch source, live service, remote model, private index or WSL workload was modified.

Current partial readback - 2026-09-29

The first acceptance item is supported by the v6 training checkpoint SHA-256 8472f22b8f02b14f27ccded353c2f378d71579a927324a25d9e5c818d4aaad90 and separately calibrated artifact SHA-256 a731903c8e91adc6206ecf60325b6ebbf4cdedd1349f0ab831c85069cf37df3d. Their parent manifests and native exact-limit exposure verified. PR #49 records the bounded public evidence; all five required and all five postmerge checks passed on verified signed revisions.

The unchanged BF16 six-stratum validator passed 6/6 numerical and 6/6 overflow checks but failed overall at 6/18 diagnostic marker decisions against the frozen 80% threshold. The fixed cases were observed during development, and #3's broader native parity gate remains open. The other original acceptance boxes remain unchecked pending full integrated acceptance. Overwatch remains archived by user direction.

Bounded native context feasibility readback, 2026-09-28. The pinned real Qwen tokenizer encoded the native-context-v3 beginning/en diagnostic at exactly 32,768 maximum branch tokens and 65,536 logical aggregate tokens across 19 branches. The reviewed two-step BF16 trained checkpoint is dd996659ff2453bd2995258b71a40c5ab8d4fbc42e0d73a32d901bbca1559ec5.

Its frozen batch policy (max_padded_tokens=2048, max_cache_bytes=67,108,864) rejects the request before model work: the first failing singleton has 27,306 padded tokens. Holding the frozen cache cap but allowing 65,536 padded tokens then rejects a 6,038,495,232-byte estimated cache requirement for that singleton. The largest singleton in this case estimates 7,112,368,128 bytes of parent KV, row KV and materialized GQA at FP32 parameter-size accounting. Content-free private receipt SHA-256: 824d1e13bfb146a81b1d65a7882ce3878682dcf74b4ca87d2b08112a976c00e5.

This is encoder/planner-only evidence for the existing short-run checkpoint. No long-context model forward, optimization exposure, calibration, GPU peak measurement, or native acceptance is claimed. Issue #4 remains open pending a deliberately frozen feasible policy and the full exposure/validation chain.

CPU-only pinned-tokenizer feasibility probe (no model invocation): the existing native context case builder produced a 19-branch diagnostic with exact longest state-plus-branch length 32,768 and logical aggregate 65,536. The pinned tokenizer SHA-256 is c0382117ea329cdf097041132f6d735924b697924d6f6fc3945713e96ce87539; Qwen config SHA-256 is 479dcf0c5286339e41ad3992cd08ae88a467c4187587936248e2b7c96283484b.

At the actual native model's FP32 parameter-byte planning rate, the unchanged default max_cache_bytes=536870912 rejected the request before a forward pass: first singleton estimate 6,038,495,232 bytes; largest singleton estimate 7,112,368,128 bytes. The GPU showed 8,570 MiB free before this CPU probe, but these estimates omit weights, activations and other work; this does not prove a larger frozen policy will fit. No policy, training run or artifact was changed.

Private reproducible script SHA-256 0c3010704843488fbd898b395b045ef70f8214db30956451c8bde1b585c07662; JSON receipt SHA-256 e6a48cf53373a8e9cded8fbf6b924ee93a5b8ea7dd1d71d0fbcdffa285279c65. This is a policy preflight, not trained exposure, calibration or inference. All #4 acceptance boxes remain open.

Pinned-tokenizer CPU-only negative extension: the exact 32,768 branch / 65,536 aggregate diagnostic now rejects a 32,769 state-plus-branch request and a 65,584 aggregate request with ContextOverflow. Updated private script SHA-256 8dd857307d152ab57936c9f32b19537cdb7a5a47994748d859d56cb52b42921d; JSON receipt SHA-256 31358e89560a939104f227db456ac867d695a455283c795981e7c81a921b7076. No trained checkpoint inference, calibration or GPU work occurred; #4 remains open.

Bounded #4 feasibility update on the local RTX 3060 / torch 2.10.0+cu128: this Windows build reports Flash Attention unavailable, but an explicit Efficient Attention-only BF16 one-layer probe passed forward/backward at 8,192 causal tokens (q14/kv2, head64, peak allocated 210,174,976 bytes). A second Efficient-only one-layer probe using the long suffix shape with 32,768 query tokens, 60,027 key tokens and causal_lower_right also passed finite forward/backward; 2.059 seconds and 952,350,720 peak allocated bytes. Raw private receipts SHA-256: 4259a23116779b0de1761f178061e174fd3f2f16795b082b8b2cb51e9a00adb9 and 1cb5f48ff274ea249a04cd22a0892f3f30ccb078d1b4edd7fe5c54c2529d7947.

This removes the unverified fused-kernel availability question for these single-layer shapes. The current checkpoint policy still rejects the actual 32k/64k case before model work, and these probes do not establish whole-model memory, optimization, calibrated exposure, exact/overflow inference, or issue completion. Keep #4 open.

Fresh bounded whole-model #4 inference diagnostic on clean signed PR #36 head c240776b66766bc6d2a8519ca1983de840dfa987: loaded the actual pinned-weight two-step full-weight checkpoint SHA-256 57428ad38b1e502cf23b6fbe8f019b2842d859dd4792d2fbd72c60b993c8b088, constructed the unchanged native-context-v3/beginning/en synthetic request, and ran BF16-autocast inference under a 70% RTX 3060 process-memory cap. The request had 19 questions, an exact 32,768-token longest branch and 65,536 logical tokens. All logits were finite. Forward time was 14.997 s, peak allocated 4,173,367,808 bytes. Private script/JSON SHA-256: 87d52d5b9e49daca87cd44668d93771d8827d4c36870816ebad78f2f6a03034c / 38e0cd3fedec0adb442ee0a28b6ff782a1bc4bdfca8395c4fd26e96a33e990d1.

This used an explicit diagnostic policy (max_padded_tokens=65,536, max_cache_bytes=8 GiB, max_branches=3) instead of the frozen trained checkpoint's 2,048-token/64 MiB policy. The conservative planner chose 19 singleton branch passes despite max_branches=3; it reused the prefix 18 times. Therefore this is feasible exact-length inference under an override, not trained 32k/64k exposure, multi-question batching, calibration, strict numerics, or full native validation. #4 remains open.

Follow-up bounded #4 inference diagnostic on the same exact 32,768/65,536-token constructed request and trained full-weight checkpoint: an explicit planner estimate cap of 16 GiB (max_branches=2, max_padded_tokens=65,536) allowed actual paired execution. Nine batches contained two questions, one contained one; one shared-prefix pass and ten branch passes returned finite logits for all 19 questions. BF16-autocast forward time was 14.537 s and peak GPU allocation remained 4,173,367,808 bytes under the unchanged 70% physical process cap. Private script/JSON SHA-256: 4d8c71f25341740fb07f5eb2226f441f94eb3de46ead6ff76a6008403ca7fec3 / 937f1dc664a6803422864df7422e12f583db8d1509f34a770ecb29730a661006.

The 16 GiB value is an explicit conservative planner estimate allowance, not observed physical usage or a production budget approval. This supports exact-length, paired inference feasibility only. No actual 32k/64k optimizer exposure, separate calibration, overflow response, strict reference parity, or whole validation pass exists. #4 remains open.

Bounded actual optimization feasibility on the same exact synthetic 32k/64k training case now has a measured local failure. A frozen four-partition generated suite (one train marker case, no held-out evaluation) has manifest SHA-256 d9562c3b059c4d54b1e31cc44a37f575902707454f5c8aed071ccf50e2f0e292 and verifies 19 questions, longest branch 32,768, aggregate 65,536. On clean signed source c240776, the pinned-weight LoRA-rank-8 activation-checkpointed BF16 trainer under a 70% RTX 3060 process cap reached the model forward but OOMed in the MLP before any optimizer step/checkpoint. With explicit paired policy the error requested another 104 MiB when 7.26 GiB was PyTorch-allocated; with explicit singleton policy it requested 52 MiB when 7.49 GiB was allocated. Both had 8.40 GiB allowed. A separate expandable-segments attempt reported that allocator mode is unsupported on this Windows build and reproduced the 104 MiB OOM. The three failed attempt identities and telemetry files remain preserved. No optimizer exposure was recorded.

This is a measured limit of these three configurations under the chosen shared-GPU cap, not proof that all hardware, offload strategies, or source architectures are infeasible. The earlier 32k/64k BF16 inference passes remain valid. #4 still requires actual trained exposure, calibration, exact/overflow requests and the full native gate; it stays open.

The separate private 32k/64k synthetic training diagnostic with 1,024-row MLP slicing still did not complete an optimizer step. Under the same 70% RTX 3060 process cap, pinned weights, BF16 LoRA and activation checkpointing, it OOMed later in Attention.forward while materializing repeated grouped-query values: 112 MiB requested; 7.04 GiB allocated and 1.27 GiB reserved but unallocated against 8.40 GiB allowed. The wrapper SHA-256 is cea1ede635887e03b21447c5247e9044a580f50a2712c69e7622b3dbe9ee082f; attempt lineage, provenance, config and telemetry remain in the private run directory. This was an in-memory monkeypatch, not merged Layev source or evidence of trained long-context exposure.

The installed Windows torch 2.10.0+cu128 SDPA documentation limits its enable_gqa implementation to Flash or math attention. Flash was unavailable in the measured long-request setup, while Layev's long path excludes dense math fallback. No grouped-query source change is established by this diagnostic. #4 remains open for actual trained exposure, calibration, exact and overflow requests, and the full artifact chain.

Partial-calibration status repair was delivered through PR #38 after the repository transfer. Exact head 5610ce2fce02f08ed9f6337cc2f73ccc059bac1e passed all five required jobs in workflow attempt 2 (36427957616); attempt 1 was cancelled before runner assignment with GitHub annotation repository transferred. Both source commits and merge commit a3be05fcd6c59c65b25662d9ddb5e67f00e93576 are verified signed. Main is protected and the PR merged normally without bypass. Local main and origin/main are synchronized at the merge SHA with divergence 0/0; the active native diagnostic remains on the unchanged PR #39 source checkout.

The repair reports a generated Noul-only calibration artifact as partial-held-out and refuses it in native acceptance. This is a status-integrity fix. Representative three-type calibration, 32,768/65,536 trained exposure, native exact-boundary parity and the other original #4 criteria remain open. The post-merge workflow on the merge SHA has started and will be read back separately.

BF16 projection repair is merged via PR #43 at GitHub-verified main 39758e61241923800b34b4d54398b38d787042db; all five PR checks and all five post-merge checks passed. The source uses fixed 1,024-row CUDA BF16 backbone projection tiles for ordinary and LoRA paths while retaining fused attention and frozen tolerances.

A fresh complete native BF16 validator on the pinned trained/calibrated actual-weight checkpoint produced six exact 65,536-logical-token English/Spanish beginning/middle/end cases. All 6/6 passed the unchanged 1e-5 parallel/full-row numerical comparison with zero maximum logit and probability error; all 6/6 passed structured overflow checks. The independent pinned Transformers short hidden-state oracle passed (max absolute error 0.000152587890625), and the 32k/64k training exposure chain verified. Private full report SHA-256 0cba4a5181ecd55176a73fe32166952a7411913b12adbfecee4fb8a6c756df. Measured optimized times ranged 16.50–19.00 s, with PyTorch peak allocations about 3.47 GB under a 70% allocator cap on the local RTX 3060.

Overall validator failed: diagnostic marker accuracy was 0/18, below the unchanged 80% criterion. FP32 parallel parity and the rest of the native precision/shape/gradient matrix also remain open. This is a real BF16 numerical repair, not #3/#4 completion or representative quality. The failed result and its private artifacts are retained.

Signed PR #44 merged FP32 fixed-64-row backbone projections at GitHub-verified 7e207d0c488466615a295385efea439a9c228d60. Five PR and five post-merge checks passed. On the pinned actual-weight checkpoint, the original 65,536-logical-token/19-question FP32 batched/full-row comparison passed unchanged 1e-5 combined tolerance (maximum logit error 1.1444091796875e-05, probability error 1.6229137018985984e-07); short HF oracle passed at 0.000152587890625. BF16 numerical parity stayed 0/0. FP32 fixed-row timing was 87.52 s batched versus 525.22 s full-row in one local measurement.

The full-weight FP32 fused-AdamW two-step stop/resume and reload were finite under the 70% process GPU cap, but trained long-context backward is not established. A fresh full six-case FP32 native validator is running. The complete BF16 native validator passed 6/6 numerical parity and overflow, yet its overall report failed marker accuracy at 0/18 versus the required 80%. #4 therefore remains open pending complete applicable native, gradient, and quality evidence.

The FP32 native 32k-branch/64k-aggregate matrix on merged source 7e207d0c488466615a295385efea439a9c228d60 now has six complete exact-65,536-logical-token English/Spanish beginning/middle/end cases. All six passed unchanged numerical parity and structured overflow checks. Short pinned Transformers oracle and the existing trained/calibrated exposure chain also passed. Maximum logit/probability errors were 1.9073486328125e-05/9.822982885293596e-07; optimized/reference latency ranges were 77.84–87.37 s / 482.64–522.24 s under the 70% RTX 3060 PyTorch cap.

Overall report passed=false because only 1/18 diagnostic marker decisions were correct (5.6% versus 80% required). The prior BF16 complete report likewise failed the overall marker gate at 0/18 despite 6/6 numerical and overflow passes. Long-context backward and representative quality remain unresolved, so #4 stays open.

The FP32 result is a recovery-labeled composite, SHA-256 50cbd14f93f7e8e92343d7cac6fce079d89f9ef783c89db5814c182fc263b919: the original process stopped after five case receipts, one continuation produced only es/end, and a finalizer hash-verified all six request/result pairs before assembling the report. A temporally adjacent NVIDIA driver event does not establish cause. No acceptance tolerance was changed.

Current merged-source long-backward update (2026-09-28): on clean f2589b9978e24cdcd4170a152d713370e8b08870, a generated two-example BF16 LoRA/activation-checkpointed run completed two parent-linked optimizer steps at exact 32,768 maximum branch / 65,536 aggregate tokens. Verified exposure: 2 examples and 2 steps, 131,072 useful forward tokens; final checkpoint SHA-256 99f797ffd0521adbee964d130d73eb381c9b6bd3b7896ce61b5007886d19ccf9, telemetry SHA-256 98f84b03b0a9d53711d09d5039d49f80e921b8db04cb759e1c2d36fbcbc87f9e. Finite loss/gradient and 8,647,482,368-byte peak PyTorch CUDA allocation under the 70% RTX 3060 cap. A separate CUDA reload produced finite short Choice/Score/Noul logits (private receipt SHA-256 429fc084b9546190b660278662b68c899a88a9610110b956db002daf9d0bb0f9). It remains uncalibrated and synthetic.

This fills the current-source long-backward mechanics gap only. The prior six-case FP32/BF16 validators still fail the frozen diagnostic marker criterion (1/18 and 0/18), and there is no reviewed representative calibration or quality result. #4 stays open with all acceptance boxes unchecked.

Current-main generated marker-curriculum diagnostic (2026-09-29). Source remained clean at dfaed82c69c4848f579297e020c25e9eab5fc39b; no acceptance threshold or model source changed.

A frozen short generated curriculum (160 train / 24 development / 24 calibration / 24 test groups; manifest SHA-256 40e9d6b5bd43c962c26f3ef1544ea332f6048f6a7a84d7f2c2a0143d44b45963) ran 120 resumable BF16 LoRA optimizer steps, 166,610 useful forward tokens. Its terminal checkpoint SHA-256 was 8c44d74dc0b3dd05ea2c004033999b23965996798df5123e605511f2b6619890. The exact-limit independent long curriculum (manifest SHA-256 984ee9c10ecd39df94bb0b8e261ec8ed311b6ecc117148fea15f10d6989a184b) then completed two parent-linked BF16 LoRA/activation-checkpointed optimizer steps on two distinct generated examples, each at 32,768 maximum branch / 65,536 aggregate tokens; 131,072 useful forward tokens, finite loss/gradient, monitoring export failures zero. Final resumable checkpoint SHA-256 f939a66c1f722f256665441d79c016c3acd1b7d35ce968443ef039c6b45f8035. A separate 24-group generated calibration produced checkpoint SHA-256 7a53dffc893bb12aa4db4ffbabd82e01f5cc15aa70bbafad40da58aabfe80dd0; verify-exposure passed the parent chain and exact-limit claims. The earlier 20-question long design OOMed before step one under the frozen 70% RTX 3060 PyTorch cap; the 19-question resource revision was frozen separately before its successful two-step run.

On that calibrated terminal checkpoint, the unchanged BF16 native validator completed all six English/Spanish beginning/middle/end cases at exact 32,768 maximum branch / 65,536 aggregate tokens. All six passed unchanged numerical parity (zero maximum logit/probability difference) and structured overflow checks; the short pinned Transformers oracle and exposure check passed. Overall passed=false: only 2/18 marker decisions correct (11.1% versus frozen 80%). Report SHA-256 0fba22628bd1c114d5ea39b19fbd7f50f273782a9d9ad4e0e7306aa044a9f73b. A separate generated development split scored 50/72 (69.4%; NLL 0.573, ECE 0.105); receipt SHA-256 a9c6e0953e5bf633cce9e24ffeee9c68678c8d3bd514934126109dd8749d29be.

These are generated mechanics/development diagnostics, not reviewed representative quality or Jev comparison. The fixed validator has been observed during development, so it is not an untouched quality test. The prior short real-weight BF16 gradient parity failure also remains. No production repair or #4 acceptance is claimed; keep #4 open.

Additional current-main native long-training evidence on clean Layev c7c6d26d80534035c67618e1f245af5f2fdbab9a (model source SHA-256 fb9e21d067f199a20d06b1a0a0a6e2222b1a577c0f6858e0c33b04942d39e024):

A separately frozen generated 24-case curriculum used exact 32,768 state-plus-branch and 65,536 aggregate tokens per example, with en/es, beginning/middle/end and group-distinct labels. Suite manifest SHA-256 7c515c4678e8b372e7c499a0a8249ebcc66b24066101ce1682c52eefd90370db; the validator's (billing,7,true) label combination was excluded from training. The v4 first attempt OOMed during step two before a checkpoint under the fixed 70% RTX 3060 PyTorch cap; failed telemetry SHA dc4ff6555012d731bb91e266239650088c94f4427812038a7ad616990b262f86. It supplies no durable two-step claim.
A separately frozen v5 same-cap attempt kept the same suite/config and initial parent checkpoint. A source-free allocator cache cleanup after gradient zeroing allowed step two in one process; the same run then resumed from its durable step-two checkpoint to predeclared step 24. Protocol SHA 28173b40e250618520de57771efd49f5aa96cbfb940ff665657b368d0443b6d3; final checkpoint SHA 799959df7fe48a54b82c5e7fd87fe706760645517327001a8baeb7050a2eaa36; telemetry SHA d006d0d7faf25853441ed938d648c6824a21dac51c13a86eccce709a7bce5e70. The trainer observed 24 optimization steps/24 exact-limit examples/1,572,864 useful forward tokens, finite terminal loss 3.06518 and gradient norm 366.715, peak allocation 8,601,058,304 bytes under the unchanged cap, and zero monitoring export failures. Every saved step 2/4/8/12/16/20/24 checkpoint matched its manifest and immediate parent hash. verify-exposure returned verified native 32k/64k on the final training artifact. The comparison does not prove the allocator cleanup caused the v4 OOM difference.
Separate generated-split calibration produced artifact SHA c1b415beb5c5d498103fe1d7595fd4eb2b9ed60768235c5c665f91a9003ea60d; exposure verification passed the calibration-to-training chain. The unchanged BF16 six-stratum native validator report SHA 83e53328d1d123b12f8a7fbae1013fff3a304a15a661bf7113ee56f9118649ca passed 6/6 exact-limit numerical comparisons (maximum logit and probability error zero), 6/6 structured overflow checks, and the short pinned Transformers oracle. Optimized case latency was 14.459-15.284 s versus independent full-row reference 84.482-88.059 s; peak allocation 3,470,024,192 bytes. Overall passed=false: 2/18 marker decisions (11.11%) versus the unchanged 80% diagnostic threshold. Generated development-only accuracy was 48/72 (66.67%), NLL 0.66158, ECE 0.14728; the earlier v3 artifact had 50/72 on that split.

These are generated mechanics and a fixed validator already observed during development. They do not establish representative quality, untouched validation, BF16 backward parity or Jev parity. No source, tolerance or issue acceptance box is changed. Keep #4 open.

Follow-up generated diagnostic on unchanged clean Layev main c7c6d26d80534035c67618e1f245af5f2fdbab9a:

The previous v5 fixed validator made all six route predictions option-099, all six escalation predictions false, and read urgency correctly only when the fact was at the end. Its long training option descriptions differed from the fixed validator's distractor descriptions. To test that coverage gap, I froze a separate 160-case short curriculum with 255 route options per case, varied option order/descriptions, en/es, beginning/middle/end and distinct route/urgency/escalation labels. The exact fixed-validator (billing,7,true) combination was excluded from training. This design followed the observed v5 failure, so it is diagnostic and cannot make the fixed validator untouched.

Short suite/config/protocol SHAs: 2a10bfdd8368e61da0cdeadf180e0ec6b3e127a1afc1ade9acccb0daf4825c2f / 15a83f3d3a2b00f7aa66009c11e40c93d115364aceb17d8839b3c6df2b51397a / f339790136e91f98436c75653f13e21ffb9f2870457088bf67da5d088ef86480. From the verified v5 step-24 parent, one run completed 160 planned short optimization steps across a durable step-40 resume, with 961,078 useful forward tokens, finite terminal loss/gradient and zero monitoring export failures. Final short checkpoint SHA 12b172f41124a71438a574e3bd529b8e41bbad77d1b464682182267a45901413. Its own observed lengths were short, so this intermediate checkpoint is not the exact-limit exposure receipt.

A separately frozen v6 continuation then trained two distinct exact 32,768-branch/65,536-aggregate generated examples from that short checkpoint under the unchanged 70% RTX 3060 PyTorch cap. Long suite/config/protocol SHAs: 0c6e05e7d4485dbcbdcb5a9b9d467564ede76f973b3cab4e58ee1d25ad8a0042 / d91870a62c43a0bd8f7077cf69ec1acd63de21a46cd89f2887c3ee2e0b8949da / 3695e2017ad0608cfdecb82cf6fa97e49f02f2f12be288f22996cee74c74159c. Final training checkpoint SHA 8472f22b8f02b14f27ccded353c2f378d71579a927324a25d9e5c818d4aaad90, telemetry SHA 6f7b561b711dfc526ef6e916f5c3e040f7703fc300d783d50e86fc4c86746a43; 2 optimizer steps, 131,072 useful forward tokens, finite terminal loss 2.80442/gradient norm 531.565, peak allocation 8,596,272,128 bytes, zero export failures. Both checkpoint hashes matched their parent manifests, and verify-exposure returned verified native 32k/64k.

Separate generated-split calibration produced artifact SHA a731903c8e91adc6206ecf60325b6ebbf4cdedd1349f0ab831c85069cf37df3d; calibrated exposure chain verified. The unchanged BF16 six-stratum validator report SHA 9b3f05edb105c2a09f80d0acf6cb28178ae64dd3329ba93415b505fcc240bc60 passed 6/6 numerical comparisons (maximum logit/probability error zero) and 6/6 structured overflow checks. Overall passed=false: 6/18 marker decisions (33.33%) versus unchanged 80%. This rose from v5's 2/18 on the same fixed cases, but remains a failure. Route is now correct in both beginning and middle language strata, while both end routes fail; urgency is correct only at the end; escalation predicts false in all six. A separate v2 generated development split scored 52/72 (72.22%), with Choice 11/24, Score 24/24, Noul 17/24, NLL 0.97448 and ECE 0.13591 (receipt SHA fb69e4f47a7207505aa25994af62930279e417b56898ceddcfb7e64bdb204a19). That split differs from v5's development data, so the two development percentages are not a paired comparison.

These results do not prove representative quality, untouched validation, BF16 backward parity, or Jev parity. No source or threshold changed. Keep #4 open.

Post-merge #4 readback after the focused short BF16 attention forward repair in PR #50: the unchanged six-stratum native validator was rerun on the same calibrated v6 actual-weight artifact, exact 65,536-logical-token cases, unchanged 1e-5 numeric/overflow gates and 70% RTX 3060 PyTorch cap. All six cached/full numeric comparisons and all six structured over-limit checks passed with zero maximum logit/probability error; the pinned FP32 Transformers hidden-state oracle passed. Overall still failed: diagnostic markers 6/18 = 33.3% versus the frozen 80% minimum, with the same six marker decisions as before. Private report SHA-256 3bba0d6d9f75c2c42ec03976438d8bd7e88e912faaa6e93ecb580ccc9e55c62e. The validator's exit 1 reflects that overall marker failure. All five PR and all five exact-merge post-merge jobs passed on GitHub-verified merge 5fedf05. #4 remains open; no marker or representative-quality acceptance is claimed.

A real optimization step at the exact context limits is now verified on pinned actual weights: 32,768 maximum branch tokens and 65,536 logical aggregate tokens on a frozen generated 19-question row. The run used the unchanged 70% RTX 3060 process cap, completed one optimizer step, had zero recomputed-logit delta, and wrote checkpoint SHA-256 ad921524f8aadd67ba1c117caada476bd3b3f5364d55ebd8c398f0bf5af7f27e. CUDA reload passed, and verify-exposure returned verified with native_32k_64k=true. Peak PyTorch allocation was 7,254,551,040 bytes. PR #53 merged with all required checks and exact-merge post-merge CI passing; the portable receipt records hashes and limitations.

This is generated-data training mechanics evidence. Separate calibration, the complete exact/overflow inference matrix, and remaining native/quality evidence have not been established. The issue stays open; acceptance boxes are unchanged.

Fresh merged-main calibrated native service edge-shape probe on source 9504b6e and pinned calibrated checkpoint SHA-256 a731903c8e91adc6206ecf60325b6ebbf4cdedd1349f0ab831c85069cf37df3d passed under the unchanged 70% RTX 3060 process cap. A generated mixed Choice/Score/Noul request exercised 255 Choice options, 10 Score levels, and 2 Noul options. The local authenticated service returned all three typed answers, 255 Choice probabilities, 10 Score probabilities and legend entries, fitted-held-out calibration status, and exact 5,492 input/forward token accounting. Branch totals were 5,274/181/65 tokens; serialized request 11,219 bytes; elapsed 2.263 seconds; peak GPU allocated 2,403,747,328 bytes. Private reproducible script SHA-256 c703e9e26ffd3863e62f491b8c6ec26d068eea3f3e273954e604ae56a186c04d, receipt SHA-256 080f23617c529f49bce6b87ee7cf94fc9c60c0713bd922546c3580b19c4cda1f. This covers edge response shapes on generated data, not exact-limit requests, representative quality, or the failed marker-decision threshold. The original remaining acceptance boxes stay open. Overwatch remains archived and untouched.

Fixed-limit native service readback on clean merged main 9504b6e, calibrated checkpoint SHA-256 a731903c8e91adc6206ecf60325b6ebbf4cdedd1349f0ab831c85069cf37df3d, pinned tokenizer and unchanged 70% RTX 3060 cap: exact generated request had 27,259 state tokens, maximum branch 32,768, aggregate 65,536, and 19 questions. The local authenticated service returned HTTP 200, all 19 answers, and input usage 65,536; 68.385 seconds, peak allocated 4,666,471,936 bytes. A measured 65,537-aggregate request returned HTTP 422 aggregate_input (actual 65,537, maximum 65,536). A measured 32,769-branch request returned HTTP 422 state_plus_branch (actual 32,769, maximum 32,768). The current long CUDA inference path requires fused SDPA, and usage counts shared state once. Private reproducible script SHA-256 f40ab75cf7cdfbfa66db4dc188bdcde79bc8d9f8af6687290c52113b0888e45f; receipt SHA-256 36119e126f461154651a29728c0b8e92e555ee5d555af1b0010343c85f0ee1ba. I checked only the original exact/overflow acceptance item. The six-stratum diagnostic marker threshold still fails, and the remaining boxes and issue stay open. Overwatch remains archived.

Portable fixed-limit service evidence is merged in PR #55 at verified merge e4822909680080bc300371782f6b439ff26fc8a8; signed source 636ab1203d78e001203b51f7a5e2134323a7f1a4. The public receipt and portable scripts/probe_native_context_service.py bind the calibrated checkpoint, trained exposure chain, exact HTTP 200 request, measured one-token-over HTTP 422 cases, 255-choice and ten-level shape evidence, GPU memory and latency. The actual portable run passed 200/422/422; local suite 858 passed/31 skipped; all five PR checks and all five exact-merge post-merge checks passed (push workflow 36591448821). The six-stratum numerical/overflow checks passed, but marker decisions remain 6/18 against the frozen 80% threshold. #4 stays open; the remaining boxes are unchanged. Overwatch remains archived.

Numerical/source repair delivered; context acceptance remains incomplete

PR #56 merged at 4f31cc76e38ef9848950e1667875571f8e190903. All five PR jobs and all five post-merge jobs passed on verified signed revisions. Source repairs and generated numerical evidence are distinct from trained context/quality acceptance.

Fresh FP32 actual-weight 8191/8192 prefix boundary and 10520-token state comparisons passed all 100 trainable LoRA/head gradients across batched/serial/full modes at unchanged tolerances, with exact logits/probabilities. These generated forward/backward probes do not establish exact-limit optimization, representative marker accuracy or full-weight parity.

The separate frozen maximum-branch 32768 / aggregate 65536 FP32 training resource probe failed during forward at the unchanged 70% allocator cap, before objective/backward. Peak allocated/reserved memory was 8,382,291,968 / 9,017,753,600 bytes; worker elapsed time was 148.098410 seconds. Portable receipt and scope preserve the measured failure.

The separately demonstrated BF16 graph lifetime repair enables a short full-weight first step and reload but does not fix resumed backward feasibility. The existing diagnostic marker threshold remains unmet; no frozen tolerance or quality threshold was relaxed. Keep #4 open and preserve its existing exposure/calibration history. No representative quality, Jev comparison or new remote-resource authority follows. Overwatch remains archived by user direction.

Development evidence delivery — September 30, 2026

PR #58 merged normally at verified merge e4dde43. Signed source99c585cb494f6397f68e4b849f0e4b66eddd57de and all five PR jobs passed; all five post-merge jobs passed in run36668523167.

Portable observed synthetic96-case/288-task diagnostic:
https://github.com/Jev-Engineering/layev/blob/e4dde4391e436ef00ad85febc804a7b2a365bf7b/docs/evidence/context/development-factorial96-20260930.md

Choice79/96, Score90/96, Noul48/96; all96 Noul predictions are true. Complete slices/factorial cells and source/checkpoint/data/report hashes are retained. This is observed short/medium development evidence. The original v8 native marker gate remains failed3/18 against unchanged80%; v6 historical6/18 remains separate.

The separately frozen paired144 development attempt timed out at its5400-second cutoff. Latest durable checkpoint96/144 has SHA256 f8e479797f721f51628d82c72af66fbed27e2cae4b6a47696400ab4a8d0b814d and a verified48-to-96 parent chain. Unsaved updates are unknown. No completed worker terminal, calibration or repair readout was observed:
https://github.com/Jev-Engineering/layev/blob/e4dde4391e436ef00ad85febc804a7b2a365bf7b/docs/evidence/context/paired144-timeout-20260930.md

Acceptance flags and original protocol thresholds remain unchanged. This supplies no representative quality, Jev benefit or completed repair/full-context acceptance. The expired window has not been restarted or extended. Overwatch remains archived per user direction.

Navigation Menu

Repository navigation

Prove trained native 32k branch and 64k request behavior #4

Gap and evidence

Acceptance

Workstation harness delivery - 2026-09-27

Current partial readback - 2026-09-29

8 remaining items

Numerical/source repair delivered; context acceptance remains incomplete

Development evidence delivery — September 30, 2026

Metadata

Assignees

Labels

Type

Projects

Milestone

Relationships

Development

Prove trained native 32k branch and 64k request behavior #4

Description

Gap and evidence

Acceptance

Workstation harness delivery - 2026-09-27

Current partial readback - 2026-09-29

Activity

CompleteDotTech commented on Sep 28, 2026

CompleteDotTech commented on Sep 28, 2026

CompleteDotTech commented on Sep 28, 2026

CompleteDotTech commented on Sep 28, 2026

CompleteDotTech commented on Sep 28, 2026

CompleteDotTech commented on Sep 28, 2026

CompleteDotTech commented on Sep 28, 2026

CompleteDotTech commented on Sep 28, 2026

8 remaining items

CompleteDotTech commented on Sep 28, 2026

CompleteDotTech commented on Sep 28, 2026

CompleteDotTech commented on Sep 28, 2026

CompleteDotTech commented on Sep 28, 2026

CompleteDotTech commented on Sep 29, 2026

CompleteDotTech commented on Sep 29, 2026

CompleteDotTech commented on Sep 29, 2026

CompleteDotTech commented on Sep 29, 2026

CompleteDotTech commented on Sep 29, 2026

CompleteDotTech commented on Sep 29, 2026

CompleteDotTech commented on Sep 29, 2026

CompleteDotTech commented on Sep 29, 2026

CompleteDotTech commented on Sep 29, 2026

CompleteDotTech commented on Sep 30, 2026

Numerical/source repair delivered; context acceptance remains incomplete

CompleteDotTech commented on Sep 30, 2026

Development evidence delivery — September 30, 2026

Metadata

Metadata

Assignees

Labels

Type

Projects

Milestone

Relationships

Development

Issue actions