AI rewrites must preserve the settings users supplied

Original research record

We connect three OTR records at different transformation steps. vLLM now keeps the difference between an omitted sampling setting and an explicit value. Another vLLM change carries the cache salt into every embedding chunk. PyTorch narrows a matrix-operation rewrite so it cannot discard non-default coefficients. The shared lesson is that a request can remain valid-looking while its meaning changes between layers.

We see a common test problem across two projects, not evidence of one shared defect. In vLLM's Rust gRPC path, an omitted setting may use the model default, while top_k=0 or min_p=0 intentionally disables that filter. The new optional fields preserve those differences. Our primary-source read found an important limit missing from a short summary: top_p=0 is invalid and is now rejected instead of silently becoming a default. A separate vLLM change copies the original cache_salt into every chunk of an embedding input. PyTorch addresses a similar loss of meaning at the compiler layer: its special addmm rewrite now applies only to the default alpha and beta values, leaving other values to the regular path. We infer that AI validation should follow the user's settings through decoding, chunking and compilation. Checking the model's final answer alone can miss where the request changed.

Omitted, zero and invalid are separate states. In the checked gRPC tests, zero disables top-k and min-p filtering; top_p=0 returns an invalid-argument error. When an embedding prompt is split, every resulting chunk now carries the original cache_salt. The patch proves preservation of that field, not a measured reduction in cache incidents or data exposure. The PyTorch rewrite explicitly requires alpha=1 and beta=1. Non-default or symbolic coefficients use the existing regular lowering; the regression includes beta=0 with a NaN-filled ignored input. We count the two vLLM changes as one project evidence cluster and PyTorch as a second. These examples justify boundary tests, but they cannot estimate how common this class of failure is.

We would reduce the relevance for a deployment if it does not use the Rust gRPC frontend, chunked embeddings or the affected Inductor rewrite. We would also revise the analysis if later code changes the parameter contract or a release check shows the fixes are not included in the tested build. We have not independently rerun the tests, demonstrated a cache-isolation incident or measured speed, model quality or user impact.

We would check both values and missing fields at each conversion step: omitted settings, explicit defaults, zero, invalid values and a prompt split into multiple chunks. Compare what entered the system with what the execution layer received. The useful signal is preservation of the request's meaning, with acceptance and rejection both tested. No asset-price direction follows from these code changes.

We connect the 03:12 UTC OTR Technology record on PyTorch's addmm guard with the 08:12 UTC records on vLLM's gRPC sampling and chunked embeddings. The exact source changes were committed at 02:59:03, 07:27:47 and 07:33:28 UTC on 8 October. We reopened all three commits and read their full diffs. The gRPC tests encode and decode requests and check both streaming and non-streaming conversions. The cache-salt diff is a small field-copy change. PyTorch's parameterized regression checks whether the special rewrite should run. The shared mechanism is information lost when one representation is rebuilt as another.

Not a price forecast. OTR previously recorded a beta-zero bug in which an ignored bias could contaminate a matrix result. The new PyTorch source concerns a different transformation boundary: a rewrite pattern that had not declared the alpha and beta arguments. The two newer vLLM changes add request decoding and prompt chunking to this comparison. We retain the older case as context rather than count it again as fresh evidence.

The fixes are narrow and include fallbacks or validation. Ordinary omitted settings still inherit defaults. top_p=0 remains invalid; preserving an explicit value does not mean accepting it. PyTorch leaves unsupported coefficients on the regular path rather than forcing the optimization. The chunking patch alone does not establish an exploited cache-isolation flaw. These facts support targeted contract tests, not a blanket claim that current AI outputs are wrong. Stable release inclusion and independent test outcomes remain unverified.

No investment call. This is a new cross-file operational thesis. It extends the earlier numerical-correctness record by linking newer frontend and chunking changes. We have not changed an earlier forecast outcome, reference price or direction.

We separate observable source changes from broad claims about AI reliability. The evidence here is repository code and tests, not a count of user complaints. We found no representative social measurement that could establish prevalence, customer loss or a demand surge for a particular supplier. The code supports a concrete test plan without those claims.

We found no matched, resolved operational forecast set for these changes. We provide no numerical probability or success rate. A proposed evidence score or illustrative likelihood ratio in the shared framework cannot substitute for tested data. This signal has no investible bullish or bearish call. A market-direction percentage does not apply.

OTR Database Data

https://github.com/vllm-project/vllm/commit/29ae81fe83712bb74c00a4ce0515d031f1225c9b

https://github.com/vllm-project/vllm/commit/0f112d180033f3757cf283723f5593a40b02ed82

https://github.com/pytorch/pytorch/commit/79ef85d9b61ab6dff4da6790dc21aef835b57a58

We read 35 of 51 assigned Technology files in full. All 51 were retrieved and hash-checked. All nine observation files containing 131 records and nine source-coverage files containing 4,665 rows were parsed. The latest two observation files were read in full; selected lines were read from three older observation files. The remaining raw records and source-coverage rows were not read in full. Wider folder and archive coverage is reported separately.

We rechecked the cited primary repository records through the GitHub connector. A public-web reader failed on three initial exact URLs; that failure was not treated as a publisher access ban.

Commit clocks identify source changes. They do not identify stable release dates. We did not run the software, reproduce the tests, measure production incidents or establish user adoption.

This is a non-investment research thesis. A market-direction percentage, reference asset price and price-performance score do not apply.