---
summary: "Durable policy: acquire external source evidence only to resolve a named context decision; source-list is not integrated."
read_when:
  - "A context decision appears to require evidence that pi-context-packer does not already have."
  - "Considering source-list, a new selection rule, or provider invocation."
type: "reference"
system4d:
  container: "Owner policy after the completed v2-v4 source-selection evidence program."
  compass: "Start with the missing fact, not an available provider."
  engine: "Name blocked decision -> identify missing fact -> choose least-cost owner source -> prove selection value -> seek operational authority."
  fog: "A valid provider can be mistaken for useful ranking evidence or permission to integrate it."
---

# Source-evidence acquisition posture — 2026-07-26

## Policy

**Use existing evidence unless a named context decision is blocked by a missing fact. Acquire only that fact from its owner. Provider availability is never a reason to acquire, rank, register, or wire it.**

Apply four separate gates:

1. **Need** — name the blocked user decision and the missing fact. If neither is concrete, stop.
2. **Acquisition** — choose the least-cost source-owned way to obtain that fact; price latency, bytes/tokens, eligibility discovery, staleness, trust, and maintenance.
3. **Selection value** — prove that a specified ranking, budgeting, or abstention rule improves the context decision. Valid input is not proof of useful selection.
4. **Operational authority** — require quality, cost, trust, and runtime approval together. Passing one gate never passes the next.

This replaces provider adoption as the decision frame.

## `source-list.v1` disposition

`source-list.v1` is a valid deterministic factual-inventory capability owned by Agent Scripts. The completed program does **not** establish that it is useless, or that it has net value for context packing.

- A caller needing tracked-source inventory or authored factual fields may invoke Agent Scripts directly, outside pi-context-packer.
- No pi-context-packer product/runtime provider invokes, acquires, ingests, ranks, registers, or wires `source-list.v1`, including in response to a manual or explicit caller request. Retained source-selection experiment code and artifacts are audit/reproduction surfaces, not an integration or standing experiment authorization.
- The threshold-1 `source_list_positive` rule is retired: it changed no selection in the valid v4 population.
- There is no metadata campaign, threshold-tuning path, automatic-invocation path, or standing successor experiment.

An explicit request changes neither ownership nor authorization.

## Evidence limit

| Record | Established | Not established |
|---|---|---|
| v2 | Metadata-assisted ranking improved precision and omissions; unnecessary-selection reduction was `16.0494%`, below the `20%` gate and uneven by repository. | Automatic net value or a duty to keep refining. |
| unranked v3 | Integrity checks rejected the prepared run before ranking. | Cost or treatment results. |
| valid one-shot v4 | Eight of ten gates passed, but unnecessary-selection reduction was `13.829787%`; threshold 1 and the full treatment selected identical paths in all 30 eligible cases. | A better threshold, universal threshold-1 inertness, or production value. |
| AK-4207 | `REJECT`; one attempt, no retry, no implementation, no automatic invocation, no production wiring. | Authority for more work. |

Thus the only durable classification is: **valid inventory capability; unproven context-packing value; rejected integration proposition.** V2-v4 and their one-shot evidence remain immutable audit records, not a tuning set.

## Reopening boundary

Do not reopen work for `source-list` itself. A new case begins only with an observed context failure whose missing fact cannot be supplied by the objective, safe seeds, docs-list, SCI, current-session evidence, or a cheaper read-only analysis.

Before experiment design, require an independently derived policy hypothesis, the decision the result would change, and separate authority. Multi-repository or one-shot evidence work is risky/expensive: route its need, integrity plan, decision impact, and owner boundaries to AK/decision authority. This document creates no task and authorizes no experiment or production change.

## Owners

- Agent Scripts owns `source-list.v1` behavior and contract.
- Pi-context-packer owns acquisition, ranking, budgeting, and context-packing value claims.
- SCI owns its separate semantic evidence; its results cannot rescue a failed source-list policy.
- AK owns task, decision, and evidence lineage.