Resources / Retrieval Layers

Retrieval Layers

Retrieval layers help readers find source candidates. They do not decide claim status, prove physics, or replace source inspection.

Retrieval layers can surface candidate material, but source inspection decides what the website may claim.

Retrieval is useful when it shortens the path to the owner. It becomes unsafe when the lookup result is treated as the owner.

Static diagram

Keep retrieval layers below canonical source inspection.

The source-first memory diagram shows memory, semantic extracts, mirrors, caches, registries, handoffs, and the final source-inspection boundary.

Retrieval layer classes

The route keeps local, generated, and indexed material useful without giving it source status.

Layer	Examples	Use	Limit
Memory lookup	Local memory summaries, prior packet notes, and retrieved workflow hints.	Find candidate files, commands, and prior decisions quickly.	Memory lookup is a navigation aid, not claim authority.
Semantic extracts	Vector or text extracts used to locate related source objects.	Discover relevant source files or registries for inspection.	Extracts can omit context and cannot replace source reading.
Obsidian or wiki mirrors	Generated notes, object pages, and relationship mirrors.	Browse relationships and find source candidates.	Generated notes are retrieval layers unless promoted by a tracked source.
SQLite or local indexes	Local databases, source indexes, and cache tables.	Speed search and repeatable lookup across large source sets.	Index rows are not proof, source authority, or public publication state.
.local and QA caches	Ignored previews, screenshots, temporary analysis, and local artifacts.	Support local inspection, debugging, or browser QA.	.local data is not committed evidence or source authority.

Source-first workflow

The workflow separates discovery, source ownership, inspection, and uncertainty reporting.

Step	Action	Boundary
Ask the retrieval question	Use memory, semantic search, wiki, or cache lookup to find candidate material.	The retrieval layer can suggest where to inspect; it cannot decide the claim.
Locate the owner	Map the hit to a source file, registry row, handoff, completion, PRD, or manifest.	The owner, not the retrieval hit, controls claim status.
Inspect the source	Read the current tracked source or governed record before changing public copy.	If source and retrieval disagree, source wins.
Report the uncertainty	Name stale, partial, or local-only retrieval evidence before reuse.	Uncertainty remains visible instead of being hidden behind a polished route.

Common failure modes

Retrieval layers fail differently. Naming the failure mode makes the safer next step clear.

Risk	Safe response	Blocked use
Stale memory	Refresh against tracked files or state before relying on the hit.	Do not present an older memory note as current source truth.
Partial extract	Open the full source and surrounding context.	Do not infer missing assumptions from the extract alone.
Local cache	Treat cache paths as private QA or navigation support.	Do not cite `.local` or ignored cache state as public evidence.
Generated mirror	Use the mirror to locate the canonical source or registry row.	Do not edit the mirror to change canonical claims.