Distill

Local context-control v2 operator contract

V2 is the separate prospective successor described in local-context-control-v1-status-and-v2-amendment.md. Do not run v1 again and do not execute v2 observations from an implementation or review checkout.

Portable offline package

make distill-local-context-control-v2

This prepares, validates, and summarizes the 14-base, 118-condition, 332-observation-per-model v2 package without network access or model execution.

Frozen v2 identities:

• corpus: 222d1a4f021022fdb048f705bf01fd3001833f441fc2f800ad960c41a174ffdc

• protocol: d1551d992e55c112b4e6e18fea5a018bdb360969947f2f4a45148a3ac50a1b27

• contexts JSONL: d97859a33500b10453bef8869e43327ed1e33f981865e5e35dceaaaff1e0f707

• schedule JSONL: 46ab21daf6d64eda431364cdf2c8e81cd5fce0d2dd38321a6dd5863e991edbf0

• result schema: 3f98d5c5427f46192fbf4413d90119e32fc618bd00679a19a32a0d742196ac11

• package manifest: c0ad25e2c16bb38bf80b1164696580367836248bcd4a9697d7e5e91b7a943f85

• framing conformance receipt: d864b23a23ff10402bc44a9261b31903077f6f7915a2374db7c4e122caaedae3

Outcome-free v2 preparation history

Two private v2 preparations are classified NO_RUN and are not study results:

1. The first stopped at readiness because Colima exceeded the frozen 1 GiB unrelated-process RSS bound.

2. The second passed three host-readiness samples, then the adapter check rejected an eligible Go-produced HostSnapshot: the Go producer encoded empty heavy_processes and failures slices as null, while the strict Python binding requires concrete empty arrays. It failed with local-context-control-mlx-v2: AdapterError: host preflight binding mismatch.

Both events ended before authorization, attempt-ledger creation, model load, or inference. The failed adapter-check directory did not survive, and there is no model output or valid result to interpret.

Because no authorization, attempt, model load, observation, or outcome exists, the correction is an outcome-independent implementation amendment within v2, not a v3 protocol. The frozen scientific protocol and package identities above remain unchanged. Re-execution requires a newly reviewed and merged implementation commit/tree, a rebuilt runner, the corrected adapter hash, and fresh owner-only runtime and adapter-check records. The old private binary must not be reused or represented as execution-eligible. A v3 namespace would be required only for a scientific protocol, schedule, outcome, or authorization contract change.

Clean-main execution sequence

Use a separate, explicitly authorized session from the exact reviewed and merged origin/main commit.

1. Build distill-local-context-control-v2 from clean merged main.

2. Prepare and validate the v2 offline package from that checkout and retain its path. The package is an input and may be under the repository build directory; only the private output namespaces below must be external.

3. Restore or reuse the exact pinned CPython/runtime tree and independently retained wheel-verification records. The full-tree identity is path-dependent and cannot be substituted by a new runtime at a different path; any fresh reconstruction must independently prove byte-identical custody.

4. Build a fresh runner and write owner-only v2 model and runtime manifests from the corrected merged implementation. Do not reuse the old private binary.

5. Run preflight; any frozen host eligibility failure blocks execution.

6. Choose absent, absolute, owner-private adapter-check, authorization, and run namespaces. They must be distinct and outside the repository, model, and runtime trees.

7. Run adapter-check with the reviewed commit/tree and an expiry no more than two hours after check start. It must finish with zero model loads.

8. Run authorize with the successful adapter-check directory and this exact acknowledgement:

I authorize only the pinned local 4B v2 development/calibration schedule with zero network access, zero cloud spend, and no scaling control.

9. Run the exact run command once. The runner repeats pre-load validation, consumes the ledger only after the adapter preload handshake, and then issues the explicit load command.

10. Preserve all private records under independent custody. Verify and summarize in a separate result-review session.

There is no retry flag. Any consumed failure is terminal and must remain preserved. No TypeSafe/AgentTrace held-out, provider, cloud, network, or scaling access is authorized.

Public aggregate result

The completed amendment-1 study produced a valid negative result. The public result report and canonical aggregate record publish aggregate values and public hashes only.

Validate the public package with:

go run ./cmd/distill-local-context-control-v2 validate-public \
  --input research/context-is-a-build-artifact