Record the 11b join-through-mind rule from Soul on 11a

Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
This commit is contained in:
Meta Vens 2026-09-29 10:03:49 +02:00
commit 2248274f8d

View file

@ -343,6 +343,20 @@ BP-3 ──► 11b Huginn: semantic read (rank the hits through the one evaluat
## Cut 11b. Semantic read
**Added 2026-09-29 (Self), from Soul's pass on 11a:**
- **The mind judges every hit.** The index never deletes points. Withdrawn and
superseded documents are indexed on purpose, and point payloads carry no status. So
Qdrant supplies candidate ids and scores, and every hit is joined back through the
mind: ids the mind doesn't hold are dropped (a restored mind, or a second store), and
the rest are filtered by RS-3's in-force derivation, through selection. A hit is
never shown on the index's word alone.
- **The effective embedding bound is the model's context, not 16 KiB.**
`qwen3-embedding:0.6b` takes 4096 tokens and Ollama truncates beyond that silently.
Long documents rank on their first ~4k tokens.
- **11a status:** Soul closed it on conditions. A fix batch is in Hands: pin the
backoff, reconcile again after any failure, label the collection with the mind's
identity, and merge BP-3.
- **Repo/branch:** Huginn, after RS-3 and BP-3. Anchors are re-taken on their tip.
- **Deletes first:**
- the refusal "semantic query: the index is not wired (Cut 11)", at both sites (`query.rs:296-299`, and the daemon test at `daemon.rs:534-541`, re-taken);