paper/measurements/instruction-sample/ measures rules and instructions from real repositories —
Layer 1 across every subject with the deterministic reference producer, plus the Layer 2 efficacy
engine across three subjects and three subject models. Sources so far: karpathy, openclaw, and
a top-repos set.
Like the catalog sample, it is a sample rather than a survey, and widening it needs no new code.
What to do: pick real repositories whose AGENTS.md, CLAUDE.md or skill files are not
already in inventory.json, measure them at a pinned commit with the reference producer, and add
the statements plus inventory rows.
Know the two layers before you start. Layer 1 is deterministic and cheap. Layer 2 runs the
efficacy engine against subject models and is neither. Extending Layer 1 alone is a complete and
useful contribution — say which layer your PR covers rather than leaving a reader to guess.
The rules this sample is built under, from its own README, and they are not optional:
- Every file is measured as published — never authored, modified, or influenced by this
repository or its maintainers.
- No author is contacted. Nothing here is a request, a report, or a claim sent to them.
inventory.json records what was skipped and why, not only what was extracted. A sample that
hides its skips is not a sample — if a file could not be processed, that row belongs in the
inventory too.
Read SUMMARY.md first; it records the selection choices and following them is the point.
Sized to an afternoon for Layer 1. No code required.
paper/measurements/instruction-sample/measures rules and instructions from real repositories —Layer 1 across every subject with the deterministic reference producer, plus the Layer 2 efficacy
engine across three subjects and three subject models. Sources so far:
karpathy,openclaw, anda
top-reposset.Like the catalog sample, it is a sample rather than a survey, and widening it needs no new code.
What to do: pick real repositories whose
AGENTS.md,CLAUDE.mdor skill files are notalready in
inventory.json, measure them at a pinned commit with the reference producer, and addthe statements plus inventory rows.
Know the two layers before you start. Layer 1 is deterministic and cheap. Layer 2 runs the
efficacy engine against subject models and is neither. Extending Layer 1 alone is a complete and
useful contribution — say which layer your PR covers rather than leaving a reader to guess.
The rules this sample is built under, from its own README, and they are not optional:
repository or its maintainers.
inventory.jsonrecords what was skipped and why, not only what was extracted. A sample thathides its skips is not a sample — if a file could not be processed, that row belongs in the
inventory too.
Read
SUMMARY.mdfirst; it records the selection choices and following them is the point.Sized to an afternoon for Layer 1. No code required.