---
summary: "Owner decision to reject the v4 positive-evidence treatment after a valid single quality ranking."
read_when:
  - "Auditing why positive-evidence filtering and automatic source-list invocation were rejected."
  - "Reviewing the v4 evidence lineage behind the durable source-evidence acquisition posture."
type: "reference"
system4d:
  container: "Owner disposition of the valid single-run v4 quality-only experiment."
  compass: "Reject a treatment that did not activate or satisfy conjunctive quality gates."
  engine: "Review fixed evidence -> apply preregistered gates -> record owner decision and non-authorizations."
  fog: "Passing precision and omission gates can be mistaken for overall treatment or production readiness."
---

# Positive-evidence source-selection decision — 2026-07-26 v4

## Decision

**REJECT the `source_list_positive` treatment tested in v4.**

Automatic `source-list` invocation, provider registration, and production wiring remain **REJECTED**. No implementation task is authorized.

V4 executed exactly once after two-stage independent pre-run acceptance. The attempt sentinel is permanent and retry is forbidden. Independent post-ranking review `dispatch-1785042159907` accepted the result bindings and arithmetic and recommended `REJECT`.

## Evidence

- Final pre-run manifest: `6e73ac83b4d1a425da390c3faed69877125c9a6f971a00b258d509d0dbf4870e`.
- Attempt sentinel: `4e4a11f9d82ba0b90261bf7dbb507abe3d6b75926bfafe73fb0aebb587c3faef`.
- Fixed result: `d1d592bc7dbb592a1b6d7b3151eefe079bdb41408cc3b02fda4c0f21d70061ca`.
- Independent audit: 250 arm-case records and 1,140 metric fields recomputed with zero discrepancies.
- Population: five declared repositories, three metadata-eligible repositories, 50 cases total, 30 eligible treatment cases.

Primary equal-repository result:

| Gate metric | Result | Threshold | Disposition |
|---|---:|---:|---|
| precision delta vs paths | +0.108333 | >= +0.10 | PASS |
| unnecessary-selection reduction | 13.829787% | >= 20% | **FAIL** |
| omission delta vs paths | -0.433333/case | <= 0 | PASS |
| zero-evidence selections | 0 | 0 | PASS |
| strict improvement over `source_list_full` | none | required | **FAIL** |

Overall: 8/10 conjunctive gates passed; the quality gate failed.

## Interpretation

The treatment was designed to remove zero-evidence backfill. It did not activate on the fresh eligible population: every eligible case had at least five positive-evidence candidates for a four-item budget. Consequently, `source_list_positive` selected exactly the same paths as `source_list_full` in all 30 eligible cases, with identical precision, recall, unnecessary selections, and omissions.

This is not evidence that arbitrary threshold tuning should follow. V4 outcomes are now visible. Threshold 1 is retired as a pi-context-packer policy candidate because it supplied no incremental behavior in the valid test; that does not prove it is universally inert outside the tested population.

## Durable disposition

There is no standing successor experiment. [Source-evidence acquisition posture — 2026-07-26](./2026-07-26-source-evidence-acquisition-posture.md) supersedes the provider-adoption/refinement frame: begin with a concrete unmet evidence need, then evaluate acquisition economics and selection-policy value separately. Any new risky experiment requires independent justification and authority; this decision does not create it.

## Non-authorizations

This decision does not authorize:

- automatic or implicit `source-list` invocation;
- production provider wiring or registration;
- implementation of `source_list_positive`;
- metadata tuning or campaign work;
- rerunning v4;
- using fusion/structural diagnostics to rescue the treatment;
- selecting a new threshold from v4 per-case score or ranking outcomes; or
- creating a successor task without a separately justified, preregistered hypothesis.

Manual or caller-requested use of the deterministic `source-list.v1` inventory remains available directly through Agent Scripts, outside pi-context-packer; it creates no pi-context-packer invocation exception.