ADR-0048: Origin-Partitioned Recall + Energy-Neutral Ranking — Reaching Autobiographical Memory

Status: Proposed — v3, revised after adversarial-design-review (GO_WITH_CHANGES, 2026-07-26). v1 (situated longer query) and the v2 research-is-the-competitor premise were both corrected by measurement. Relates to: ADR-0047 (attention re-ranks within a result — bottom-up: reach → energy-neutral → awareness → attention), the research: ingest prefix, WavefrontMeta (types.rs:517 — note the existing provenance field is the entropy stamp, a name collision this ADR must avoid), and hemisphere.rs:232-247 (the full-scan resonate that ranks by similarity × energy).

Context

The origin failure was measured, not assumed. Two live probes:

• v1 (situated longer query) made recall worse — a longer query fed the large research: corpus more to grab.
• v2 (blame the research corpus) was refuted by a third probe the reviewer ran: recall "where is Kannaka Labs" --top-k 20 returns zero research: memories — the top hits are all episodic (OBC art, recognition milestones, sessions), and the lab is still absent.

The real competitor is energy imbalance, not provenance. Ranking is similarity × energy (hemisphere.rs:238), and every recall pumps the surfaced memory's energy toward the 2.0 cap. So frequently-recalled memories rich-get- richer; the never-surfaced lab sits at baseline energy and loses even to weaker- similarity episodic memories. Two forces bury autobiographical recall: (1) the semantic corpus out-competes on conceptual queries, and (2) within episodic, recall-frequency energy out-competes a cold fact. This ADR addresses both, and is honest that surfacing the lab needs a stack, not one fix.

Decision

Two independent mechanisms, each falsifiable on its own gate.

1. Origin-partitioned recall (reachability)

• A persisted typed origin field. Add origin: OriginClass { Ingested, Lived } as the last serialized field on WavefrontMeta (bincode default-fallback so old .hrm still load, mirroring types.rs:509-517). Stamped at the call site at remember time (research/OpenAlex handler → Ingested; remember/dream/ perception → Lived), backfilled once from the canonical research: prefix, and always read from metadata — never recomputed from the content string (recomputing would make it content-derived like the refuted Fano class). Do not overload the existing entropy provenance field or auto-detected Modality (which stamps autobiographical text Semantic). Fix the writer/reader prefix skew ("research: " vs "research:") and assert byte-identical predicates.
• Single-pass two-heap partition. Inside the existing full-scan resonate loop (hemisphere.rs:232), bin each scored candidate into an Ingested or Lived top-K heap by origin. One scan, two ranked outputs — no separate index, no second resonate() call (both refuted as unnecessary/costly).
• Quota merge, not weight-then-truncate. Fetch top-K from each partition (never top_k/2 — that returns 0 for top_k=1 liveness callers), union, and reserve ≥1 configurable slot per partition (round-robin by rank). Intent sets ordering within the quota; it can never exclude a partition. This preserves the ADR-0047 invariant that the fix is structural (candidate membership), not a post-truncate multiply.
• Intent is an explicit caller parameter, intent: Option<Intent>, threaded through the recall path; default Unknown → balanced. The benchmark drives the same path production uses (no test-only injection). A token heuristic is a separate, later increment that must pass negative controls before it may gate ("where do grid cells fire" MUST classify knowledge; a bare where/place token must not imply episodic) — it is not the floor, killing the circular dependency on the unbuilt awareness layer.
• Quarantine hallucinated and non-text-modality memories out of the Lived partition (reuse the re_encode_all predicate, chiral.rs:642) so dreams and raw perception don't compete as lived experience.

2. Energy-neutral ranking (within-episodic reach)

Within the Lived partition, rank by similarity (or similarity × √energy) instead of similarity × energy, so a never-recalled memory is not buried by the rich-get-richer energy of frequently-surfaced ones. This is the measured within-episodic surfacer; partitioning alone cannot do it. Flag-gated, energy array left byte-identical (ranking-only change; no write-back).

Two falsifiable gates (honest about composition)

• G1 — ADR-0048-owned, falsifiable in isolation (gates the partition flag): for an autobiographical-intent query the Lived partition contributes ≥N candidates to the merged pool; knowledge-query research recall is not regressed (Δrank ≤ 0); recall(q, 1) is non-empty for every caller. Measured on the chiral path (not the flat readonly mirror), energy byte-identical, on a corpus with zero research: memories partitioned recall == today rank-for-rank.
• G2 — composed, owned by the rollup: the cold "where is Kannaka Labs" query rank-wins the lab. Explicitly requires energy-neutral ranking + awareness (situated intent) + ADR-0047 attention — the ADR records up front that partition alone cannot pass this, because the competitor is intra-episodic energy, not research.

Confirmed (do not regress)

• Pre-fetch partition genuinely escapes ADR-0047 post-fetch inertness by changing candidate-set membership — preserve; a later refactor collapsing it to a post-truncate weight silently regresses to the Fano failure.
• The 2-class origin axis differs from Fano's 7 content-classes only because it is a persisted origin fact read from metadata — ship the field before claiming it.
• Partitioning is affordable on the existing brute-force scan (no ANN/HNSW needed).

Build order (flags default OFF; G1 gates before G2)

1. Typed origin field — struct + serialize (bincode-fallback), call-site stamps, one-time prefix backfill + an audit (how many research: matches vs known-research that don't; enumerate every ingest path).
2. Single-pass two-heap partition + quota merge + explicit intent param — with the empty-semantic == today guard.
3. G1 benchmark — chiral path, Δrank, no knowledge regression, top_k=1 non-empty, energy byte-identical.
4. Energy-neutral ranking (ranking-only, flagged).
5. Compose upward toward G2 — awareness (intent + situate), then ADR-0047.

Decision record — 2026-09-16: energy-neutral ranking is now the default

KANNAKA_RECALL_ENERGY_EXP shipped off-by-default (1.0 = the historical similarity * energy). On 2026-09-16 the default flipped to 0.0, and the amplitude → energy copy on insert and on load gained the same ENERGY_CAP every boost path already had. Evidence (kannaka-memory#965), measured on the live O1 store against its own vectors and probes:

In 80% of the misses the correct memory had the higher cosine and lost on energy; the winner carried 3.7× the target's. The cap was real but applied only to boosts — the write path copied caller-supplied amplitude in unbounded, so records entered at 7.7 and 8.5 and sat in a third of all top-10 lists whatever was asked. The rest of the medium was measured clean along the way: the codebook projection is lossless for d_eff (×0.99), the stored rows match their encodings at corr 0.994, and facet resolution returns children under their parent by design. 1.0 remains available to reproduce the old ranking.

Alternatives considered (and discarded by the review)

• Post-fetch provenance weight / situated longer query — refuted (inert / worse).
• Separate per-provenance index / two resonate() calls — unnecessary; the single existing scan bins at identical cost.
• Heuristic intent as the floor gate — deferred; self-defeats on the paired fixtures.
• Blaming the research corpus — measured false; the within-episodic energy imbalance is co-primary.