Skip to content

docs(research): add the 1.4 to 1.5 research map - #117

Merged
tc3oliver merged 5 commits into
mainfrom
docs/research-map
Sep 26, 2026
Merged

tc3oliver merged 5 commits into
mainfrom
docs/research-map

Conversation

@tc3oliver

Copy link
Copy Markdown
Owner

Adds research/README.md, an index of the GPU + ANE research line from #16 to #105. It is the entry point for citing this work. It adds no measurement, changes no verdict, and does not edit any study's README.

What it holds:

  • a Mermaid journey from the first interference measurement to 1.5 adaptive execution. The coremltools GIL finding is marked as an upstream opportunity only; no upstream issue exists yet;
  • a 19-row map: question → primary evidence (with its kind: observation, intervention, gate, post-hoc or production) → finding → consequence;
  • key findings, the production outcome, evidence boundaries, and a track index to each study directory.

Every number was checked against its study's own README, tables or JSON. Negative results and the recorded FAILs stay as recorded (#57, #77, #88, #92, #99, #102, #103).

The boundaries it states:

  • E-core residency is a strong correlation, not a scheduler cause.
  • The breaker's recovery evidence is the controlled Phase 1 only; validation had no natural trip.
  • Recovery from a mid-episode slow state (the research(execution): record the closed H64 production evaluation #103 kind) is untested.
  • The C3 detector was chosen on the replay data, not preregistered.

Docs only: no runtime code, benchmark data or research data changes.

research/ had no index. The GPU + ANE line from #16 to #105 is spread over
twelve study directories and forty issues and pull requests, so the chain
from GIL causality to the adaptive breaker, and the failed routes between
them, had to be rebuilt from each study's README.

research/README.md maps each question to its primary evidence, verdict and
consequence. It keeps the negative results and states the evidence
boundaries: E-core residency is a correlation, recovery was measured only in
the controlled Phase 1, and mid-episode recovery is untested. No study,
runtime code or benchmark data changes.
The README pointed only at the adaptive-breaker study, so a reader had no
path to the research line that led to 1.5 or to its failed routes. The
Chinese READMEs get the same line.

The research map also stops asserting that no upstream coremltools issue or
pull request exists; it now says only that this repository records none.
The research map called the coremltools GIL release an upstream
opportunity, but apple/coremltools#2876 is now open. The journey diagram,
the GIL causality section and the link references now cite it, and say it
is open and unmerged, so released coremltools still holds the GIL.
…quest

#46 measured CompiledMLModel.predict, while apple/coremltools#2876 changes
MLModel.predict() / Model::predict and keeps predict() synchronous. The
map now says which path each one covers.
@tc3oliver
tc3oliver merged commit 7cb5c87 into main Sep 26, 2026
2 checks passed
@tc3oliver
tc3oliver deleted the docs/research-map branch September 26, 2026 16:25
Sign up for free to join this conversation on GitHub. Already have an account? Sign in to comment

Labels

None yet

Projects

None yet

Development

Successfully merging this pull request may close these issues.

1 participant