Source notes
What exists on the cluster but is not in this repository, and why. The intent is that nobody has to wonder whether something was hidden.
The agent workspaces
7,628 intermediate kernel snapshots across 219 per-target workspaces, 5.8 GB.
Each operator's workspace holds candidate/, bench_results/, and a
snapshots/codex-NN/ directory for every revision the session produced --
gelu alone went through 14. For a dataset about kernel-generating agents this
is arguably the most interesting material in the whole run: it is the full edit
history of every kernel, successful and abandoned.
It is not shipped because 5.8 GB of mostly-near-duplicate Python would dominate the repository, and because the snapshots have not been reviewed for anything the agent may have pasted in from its context. The final winning kernels are recoverable from the bench results plus the harness; the intermediate revisions are not recoverable from anything here.
Retained at /code/users/alexsu/codex-fk-runs/ on the cluster. If the snapshot
history is wanted as a release, it should be a separate, reviewed one.
The smoke run
A 5-operator pipeline validation (rms_norm, silu, softmax, gelu,
layer_norm) run before committing to the full sweep. All five won, at geomean
1.04x - 1.61x, and the composed bench passed 25/25.
Excluded because it is a different, much smaller candidate set and mixing it
into the tables would be misleading. It appears once in REPORT.md: it is the
only configuration in this project where fastkernels eval produced end-to-end
numbers, so its Llama-3.1-8B correctness row is used as the comparison against
control/, labelled as such.
Retained at /code/users/alexsu/codex-fk-runs-smoke/.
The operational runbook
CODEX-RUNBOOK.md, kept on the cluster: how to rebuild the node, the four
pieces of codex configuration needed to reach Snowflake Cortex, token rotation,
the daily monitoring commands, and the running list of harness bugs.
Its technical content has been folded into MANIFEST.md and CANONICAL.md.
The runbook itself is not shipped because it is written for one person on one
cluster and is full of internal paths and credential-handling steps.
Credentials
The Snowflake Cortex token lives in $HOME/.snowhouse-pat on the node and is
not in this repository, any log shipped here, or any file under patches/.
patches/codex-config.toml references it only indirectly, via
env_key = "OPENAI_API_KEY". The driver logs were checked for token material
before packaging.
Evaluation logs
~/.fastkernels/evals/logs/ holds the per-scenario end-to-end logs, including
the full tracebacks for the three Dynamo failures and the two OOMs quoted in
REPORT.md. They are not shipped: the scenarios produced no results, the
tracebacks that matter are quoted in full in the report, and the logs are
dominated by engine startup noise. Retained on the cluster.
Superseded scoring executions
Executions 1 and 2 (see CANONICAL.md) wrote result files that were overwritten
in place by execution 3, so they no longer exist even on the cluster. What
survives is their driver logs, which ship under logs/ because they are the
evidence for the two harness bugs and for the gate's run-to-run instability.
The per-execution pass/fail counts and winner sets quoted in CANONICAL.md were
read back out of those logs.