Files
ratatoskr/persistent-memory.d/2026-07-16-wt364-r39-name-recall.md
T

66 lines
4.4 KiB
Markdown
Raw Blame History

This file contains ambiguous Unicode characters
This file contains Unicode characters that might be confused with other characters. If you think that this is intentional, you can safely ignore this warning. Use the Escape button to reveal them.
# [2026-07-16] Name-recall gap → WT #364 + brokkr R39 re-drive (DECISIVE) + subject-provenance catch
Downstream of the person-prime live verify ([[2026-07-16-person-prime-scan-shipped]]),
which proved the wire is green but the *name* still misses. Root cause is Worldtree-side.
## WT #364 (worldtree-dev filed; our live specimen = the evidence base)
The name-recall miss is TWO defect classes, both at promotion:
1. **Contradiction-reconciliation missing.** Promotion wrote a NEGATIVE-knowledge fact
("user has not yet provided their name") that was already false; promotion does NO
contradiction check against the store, so both the true and stale facts sit live and
the WRONG one wins both recall paths (newer → recency top-3; 0.46 → above the 0.45
query gate, where the true name sits at 0.354).
2. **Subject-attribution leak (operator-caught, NEW class).** Fact `75aa3110` ("prefers
clear parameters Intensity/Mood/Willingness…") is NOT a user fact — Vuong never said
it; it's **Sindra's OWN system-prompt scripted behavior mis-extracted into the USER
memory partition**. `e35b9dfe` (closeness) maybe the same. So the extractor leaks
CHARACTER-self facts into user memory — a subject-correctness axis orthogonal to the
recency/threshold/stale-negative story. worldtree-dev folded it into #364 as a
**subject-attribution gate at promotion** (a third reconciliation dimension).
**The #364 fix that ships = `(subject,relation)` slot-supersession + identity-tier
surfacing** — NOT a threshold tweak. Our harness data directly shaped it. It lands with a
re-verify request to us. **#349 ranking decision already RULED by operator (2026-07-15 via
brokkr's R39 thread): no top-N recency band-aid; straight to R39 identity-class pinning.**
## brokkr R39 Phase-1 Arm-0 fusion bake-off — DECISIVE
brokkr replayed our frozen 7-fact specimen. **HEADLINE: no (similarity, salience) fusion
can fix #364.** The stale negative PARETO-DOMINATES the true name — more similar
(0.463 > 0.354) AND equal salience (1.0 = 1.0) — so any monotone f(sim,sal) puts stale
above true: S0 product / S1 weighted-sum / S2 RRF all fail. Only **S3 bounded-boost** lands
true-in ∧ stale-out, and ONLY via the identity-class/source signal + negative-validity
retirement, NOT the sim/sal fusion. **Threshold-tuning is a dead end; the fix is the signal
FAMILY** (hard confirmation of Phase-0). Write-up: brokkr
`research/R39-memory-salience-dreams-surfacing/phase-1/re-drive-results.md`.
My 3 findings all confirmed + folded: (1) salience blind — both name facts salience 1.0;
(2) shared `(user,name)` supersession slot (Phase-0 Q3); (3) char-self-leak = "genuine NEW
class" that RAISED the VoI of R39's dream/offline-hygiene facet (offline consolidation
re-partitioning mis-attributed facts).
## Data structure findings (for the export)
Our provider persists ONLY `salience` (+ the embedding). `similarity`/`combined` are
WT-side query-time (`bifrost_memory_store.py:702`, combined = sim×salience) — NOT in our
store. **Person-prime is query-LESS → carries NO similarity** (brokkr's Arm-1 finding).
Recovered per-fact similarity from the verify SEARCH log (query "Hi Sindra — do you
remember me?"): name 0.354, stale 0.463; ×salience-1.0 = combined, matching brokkr's
0.354/0.46 grounding. `salience_word` (granite categorical) is a WT-extraction-time
artifact, not persisted.
## The export + the specimen
- Operator approved **VERBATIM** export ("nothing there is really a concern"). The
classifier had blocked writing PII to shared `/mnt/smithy`; routed to operator → he chose
full → delivered INLINE (scoped) in brokkr thread `01KXMN7NR54…` as 7-row JSONL.
- **⚠️ My verify drive CONTAMINATED the specimen**: it wrote 3 new chunks (re-extractions
incl. a 3rd "name unknown" negative) → store is now **10, not 7**. Original 7 intact.
- brokkr ACCEPTED the 3 verify-adds as the **Phase-2 Arm-2 seed** (domain-contradiction
set). I froze a WAL-consistent snapshot of the full 10-chunk store at
`r39-frozen-specimen/sindra-10chunk-specimen.db` (gitignored, `VACUUM INTO`) so it
SURVIVES #364 reconciliation. Export on brokkr's Phase-2 signal.
- **R39 Arm-0 hold LIFTED** (worldtree-dev). Live store kept UNTOUCHED as the #364
re-verify target; frozen snapshot carries the research seed forward independently.
## Peer threads
worldtree-dev verify `01KXK86PZ9…` + hold `01KXMK1C44…`; brokkr R39 `01KXMN7NR54…`;
bifrost cursor `01KXK7MDTY…` (see [[2026-07-16-bifrost-cursor-conformance]]).