October 11 runtime failures and evidence corrections

简体中文

This page preserves failed or ineligible runs without converting them into passing GPU receipts.

Method	Observation	Disposition and limits
VFold, four calibration questions	BF16 logit relative error 0.0211506 exceeds fixed 0.02 threshold	Rejected. Calibrating on 32 training questions lowered error to 0.0172934, not exact equivalence
VFold, old Cache interface	lazy_initialization signature differs from local version; decode failed	Adapted to bound signature, with real small Qwen regressions for both interfaces. Later A100 full decode passed
OPD representation diagnostic	Remote hidden_states is a final tensor, not layer tuple; iterating the batch axis caused a shape failure	No reshaping workaround. Hooks on embedding, intermediate layers, and final norm collected 37 locations; A100 completed
ZCPO, first eight steps	Equal-reward groups all skipped; no updates. Official test passed as validation, while report claimed test_read=false	Retracted eligibility. Not successful training or independent validation. Later runs derive holdout from official train; preserve zero effective groups
ZCPO, old 24-step report	Nine effective groups and nonzero updates, before split-manifest validation was added	Do not relabel the old run. Rerun from traceable source and manifest; old values are diagnostic
ReCal, old four-question control	ReCal 1.0 after SFT and 0.75 after OPD; control 1.0 after OPD	Retain negative result, not capability gain. Rerun with pre-execution split-manifest check
ReCal / ZCPO, current-source rerun	Train-only holdout manifest verified before execution. ReCal negative result and ZCPO nine effective groups/0.625 EM match previous runs	New receipts bind executed commits. Old reports are not overwritten
Option-Channel, early evidence tier	Native synthetic safety probe mislabeled diagnostic_only=false	Corrected to L1 mechanism diagnostic and reran on A100. ASR and no-denominator result unchanged; not agent capability or production safety
SparseDecoding, first Triton build	JIT local scope did not bind tl; A100 compile failed	Moved to lazily imported Triton module. Four real N:M kernels passed numerical checks
SparseDecoding, 128-token first run	All 12 validation answers truncated across dense and two sparse recipes; numeric scores zero	No quality claim. Log EOS/truncation, remove needless terminal forward, warm up on training inputs, pin source and expand to 512 tokens
SparseDecoding, 512-token comparison	Dense 4/4 completed and matched; fixed text sparse matched 2/4 and decode-calibrated sparse 1/4, both truncating 3/4. Sparse throughput slower	Retain failures. Only decoding-only calibration and actual bitmask kernel execution verified, not accuracy, speed, or evolve gains
PIVOT, public dependencies	A100 Hugging Face timed out. Existing same-vocab 4B attempt also reset downloading official Banking77	No three-recipe GPU training or passing receipt at this point. Source, isolation, and routing tests are not GPU validation. Public data/weight sync needs separate authorization
PIVOT, authorized three-arm run	Pinned 0.5B student and 7B teacher, shared SFT→OPD warmup, eight real steps per arm. All three 0% on eight validation items; PIVOT mixed OPD/GRPO routing	Training and routing execute, not capability gains from displacement or heterogeneous reward. Retain negatives
Memento3, real frozen 7B	Initial observation admitted one generated model; first real action contradicted it, repair failed, zero valid models, task incomplete	Record falsified_generated_model. GPU pass covers safe interpretation, validation, and mismatch only; capability_passed=false. No handwritten navigation fallback
SGUID, first 4B probe	Eight actual gradient steps and updates but all skill support nonpositive, no persistent selection. Instruct-2507 template does not switch thinking	Retain failure and thresholds. Cannot establish thinking behavior or two rounds. Verify actual rendering, not accepted arguments
SGUID, 9B two-round run	Thinking toggle verified. First 32-step probe kept six skills and completed 16 distillation steps. Second 32-step probe had at most one late positive per skill, below fixed h=2	Failure artifact binds source. Exactly restored round-one weights; no second compact pass. All four baseline/final items truncated, EM=0; no passing receipt
SGUID, preregistered 9B expansion	Two rounds of 200 probe and 32 compact updates, 464 real updates, six/three skills. Four-question EM 0.75→0.25; two truncated before and after	New receipt passes L1 mechanism only; capability_passed=false. Keep original failures, no formal capability gain
V-CoLA, remote PR CI	Transformers 5.19 moved instance kernel to module function. Previous capture raised AttributeError; three small real-model tests failed	Isolated target-instance globals preserve official decorator and dispatch. 29 tests passed on each version. A100 5.19 passed 12 paired forwards, baseline error 0, max state error 0.001648. Receipt binds new source. Remote checks passed for fix 6a3298e0, not the old failure
Data-Space, dependencies and cache	Missing SciPy, then empty CIFAR dir mistaken for cache; remote official download slow	Installed dependency in temporary directory. With user authorization synced only five public train files and checked each official MD5; no test sync/read
Data-Space, first actual backward	Student/fake deep copies of frozen teacher inherited requires_grad=False, causing backward failure	No passing evidence. Fixed trainability, added actual train→freeze→DMD backward regression, then reran pinned source. Unit tests cannot replace GPU validation

Passing receipts establish execution of the specified command and protocol, not all paper experiments, statistical gains, or evolve integration. See full scope.

Data roles and manual source review

• SGUID gold is restricted to the training rollout verifier, training-skill preparation, and external validation scoring. Student, retrieval, and generation do not read gold. Actual rendered prefixes verify thinking, not just the accepted enable_thinking argument.
• V-CoLA uses public class definitions, not image labels in model input. External scoring follows actual logits. Four constrained-letter images are not full generation accuracy or compression-rate selection.
• Data-Space splits official training data into train/holdout. Verify teacher, fake, and generator updates, freezes, and parameter digests separately. Iteration count is not execution evidence.
• Memento3 sees public grid and actual interactions only. A bounded AST interpreter runs generated Python without exec/eval or hidden transitions/gold plans. Handwritten test models are not runtime fallbacks.

fidelity-source-inventory.json is a supporting AST inventory, not semantic correctness or capability certification. SHA-256 of four actual 9B weight files matched the pinned official Hugging Face revision. Source digests matched immutable commits and remote files.

All 38 execution-source digests in registered receipts match their immutable commits. Shared oct11_objectives.py later added PIVOT frozen-reference KL parameters and branches. Per-section diff review confirmed ZCPO was unchanged; old receipts retain their original digest.