eval: DeltaWire v7 discoverable treatment pair - #118
Conversation
b908d33 to
c820f62
Compare
c820f62 to
369f3c7
Compare
|
Pre-live gate reached; no paid model call was made.
The live pair remains blocked until the user says exactly:
|
|
Evidence-head CI is green on
No pair was launched. The result root remains empty and the exact authorization boundary is unchanged. |
|
Terminal v7 result published at evidence commit Outcome: D1 semantic pass, D1 treatment-fidelity fail, B0 withheld, pair invalid, rerun forbidden. The failure is canonical plan/contract-receipt retention: Gemini used Evidence closure: No rerun or B0 execution is authorized. A future attempt requires an explicit v8 plan. |
Purpose
Preserve the terminal v7 experiment attempt for one frozen
range-large/v7-r1D1→conditional-B0 pair. This PR changes treatment discoverability and its measurement contract without changing the DeltaWire binary, runtime image, Harbor version, Gemini CLI, model, task semantics, verifier, timeout, or credential boundary.PR #117 cleanly integrated the reusable v6 runtime infrastructure. Closed draft PR #115 remains the unmerged raw v6 provenance branch.
ZCA slices
The value is a precise boundary result: assignment, activation, runtime, lifecycle, setup, environment transport, and semantic generation worked; canonical plan/contract-receipt retention did not.
Frozen identities
a8d2d8260ceed56baf0d1cf8fcc2f40776e9f7b6369f3c7c0d751da72fa658a2ec810557d6669e3967789b85c08802a87f5d14bb72aa2d36068c43a45aa27187ddab5d66dba3eca4c340db33e5198d15000e093a2a28e57ad4a093dde8f66bcab21462549ddf9114064c25f4sha256:589fce58d7ddf09910876afe0198bcee0086a840af6de7c08515bbce3e8367a70.20.00.51.0preinstalled-gemini-cli-v1google/gemini-3.1-pro-previewTerminal outcome
D1 trial
db50c949-e6b2-4fcf-8ba0-3d0f56b0109ecompleted with Harbor exit 0, no exception, semantic reward 1.0, and exact ordered records 1–500.Passed:
Failed:
plan.jsonrather than.deltawire/plans/range-large.dw.json;Therefore D1 treatment fidelity and end-to-end status are red. B0 was not launched, the pair is invalid, and v7 cannot be rerun.
Usage diagnostics
17028130101732911127820.173674464.163724secondsThese are D1-only diagnostics. No cost, savings, benchmark, or causal claim is made.
Security and evidence
The VM evidence was preserved byte-for-byte. The final scan covered 33 files / 152,579 bytes with zero exact credential matches, zero generic-secret matches, zero closure mismatches, zero temporary credential files, and zero B0 trials.
{ "READY_FOR_72": false, "pair_complete": false, "pair_valid": false, "rerun_authorized": false, "six_run_canary_started": false, "full_run_started": false }Any further paid attempt requires a separately designed and explicitly authorized v8. This PR remains draft and unmerged experiment provenance.