---
summary: "Ripwire-only rollout: explicit use available, automatic promotion blocked on paired agent-task evidence."
read_when:
  - "Reviewing PR09 or choosing a code-context activation policy."
system4d:
  container: "Local trusted code-discovery provider in pi-context-packer."
  compass: "Do not promote a retrieval benchmark into an agent-task success claim."
  engine: "Explicit selection -> bounded acquisition -> independently reviewable evidence."
  fog: "Automatic activation remains unapproved; reference timings are not equivalent workloads."
---

# Ripwire rollout status

**Explicit use is implemented. Automatic activation is OFF.**

The committed `RIPWIRE_AUTO_APPROVAL` record is immutable and disabled. A model
parameter or environment opt-in cannot substitute for reviewed adoption evidence.
`providers.ripwire: required` requests code; `off` prevents execution.
`PI_CONTEXT_PACKER_RIPWIRE_DISABLED=1` is an operator kill switch that also blocks
explicit requests before binary reading, source copying, cache access or subprocesses.
There is no replacement backend or automatic installation.

The policy supports relevance, available-headroom, provisioning and a trusted host
assertion that a request is already satisfied. A nonempty working set by itself does
not establish that the current objective has been answered. Automatic selection,
when approved in a later reviewed promotion, is selection within `context_pack`,
not an unconditional every-turn injection hook. Automatic-provider failures remain
omissions; explicit required failures make the packet incomplete.

## Completed retrieval-only study

The exact PR08 candidate `d5544e145aac32853ac6366fd1bf51102a60213c` was packed as
`3e4447c84dbfe0e8f9397aac792ef164678d49a452a0c7767bd7f37222f28907`.
The frozen study manifest digest was
`441074e40af6057dcd3bf10134591f122c1c0dc995bd246e4c6e2594f063ba0b`.
The complete result file digest was
`428085cf82b090c24c37a07a86c59aaea22fa50da7c6ecdfa62231ccddee0732`.

Across 30 source-inspected questions on three declared pinned public source
snapshots, the expected file appeared in the rendered packet's first ten unique
paths on 28 cases; the deterministic literal-search reference hit 24.
There was one gold file per question. This does not establish multi-file change
completeness or general semantic recall. Two failed retrievals remain recorded;
no post-result tuning changed these answers. Two interrupted acquisition attempts
were preserved separately and are not counted as complete studies.

Cold full-provider median/p95 measured about 1.73/3.19 seconds in this environment.
The literal reference used previously read source and is not an equivalent latency
workload. Neither those timings nor the selected-output size establish a warm
production SLO or total model-context savings. No model-agent trial was executed.

## Required before automatic promotion

Run the paired task pilot from `ripwire-evaluation.md` on the same host/model,
repository state, permissions and budgets, including fallbacks and repeated packets.
Retain source/target identities, actual invocations, full model-visible cost and
independently checked task outcomes. Report regressions and uncertainty. An
independent reviewer must validate the evidence and an explicit later policy change
must approve it; current tests do not invent that approval.

Review and installed-runtime verification remain separate from publication and
operator installation. No live setup has been changed by this stack.