ADR-0053 — Phase-Preserving Multi-Path Recall

Status: Experimental / Proposed — v2. (v1 proposed encoder replacement as the headline; it was built on a misdiagnosis and is superseded by this document. The v1 errors are recorded in Rejected diagnoses below because they are instructive.) Date: 2026-08-01 Relates to: ADR-0046/0048 (energy-neutral ranking), ADR-0049 (facet encoding — proposed, not shipped), ADR-0050 (temporal-confirmation weighting), ADR-0036 (dream consolidation), ADR-0024 (consciousness metrics), ADR-0029 (recall envelope)

Context

The substrate is wave-mechanical; the recall interface is classical

Medium::recall_against (src/medium/core.rs) already computes interference:

// Phase DIFFERENCE (constructive interference), not absolute phase.
let phase_modulation = (self.store.phase[i] - query_phase).cos();
let mut resonance_strength = similarity * effective_strength * phase_modulation;

Per-wavefront phase exists, the query carries a phase, and cos(Δφ) can go negative — memories are already suppressed by phase opposition inside a single path. But the value that leaves the path is a real scalar:

// src/medium/types.rs:564
pub struct Resonance {
    pub id: Uuid,
    pub content: String,
    pub similarity: f32,
    pub resonance_strength: f32,
    pub effective_strength: f32,
}

Phase dies at the API boundary. Every recall path computes wave mechanics internally and then collapses to a magnitude before returning. The consequence is structural: multiple paths can only ever be combined by *rank fusion over real numbers* — classical ensembling. They cannot superpose. The physics cannot determine the output, because the output no longer carries what physics needs.

There are already many paths, and they are mutually exclusive switches

HrmStore.medium is constructed Medium::new() / Medium::load(...) — the flat medium. Hemisphere::resonate (hemisphere.rs:333) does delegate to resonate_with_weights reading KANNAKA_RECALL_ENERGY_EXP, KANNAKA_RECALL_TEMPORAL_EXP, and the half-life — so those knobs are correctly wired on the chiral path, which CLI recall does not enter.

Measured 2026-08-01: setting KANNAKA_RECALL_ENERGY_EXP=0.0 produced byte-identical CLI recall output (same ids, same scores to 3dp), consistent with the flat path having independent scoring. The mechanisms ADR-0048 and ADR-0050 measured and accepted are live code on a path not in daily use.

Rejected diagnoses (recorded so they are not re-derived)

1. "The medium is saturated; d_eff collapsed to 6%." effective_dimensionality() is a participation ratio over memories, bounded by n. d_eff = 597.97 at n = 609 is 98.2% of maximum — healthy uniformity. The ratio field divides by nominal 10,000 and is low only because 609 ≪ 10,000.
2. "The corpus is polluted with low-value audio memories." They crowd recall because they are lexically repetitive under a lexical encoder. That is an encoder property, not a statement about their worth; they are also feedback signal and plausible training supervision.
3. "The encoder is the root cause." SimpleHashEncoder::new(384, 42) is indeed what init_with_hrm constructs, and it is random indexing (lexical, not semantic). But ADR-0049 already chased this to the encoder by measurement and found the operative variable to be compound vs atomic encoding — same fact atomically stored reaches rank 1, sim 0.763. Encoder backend is a supporting change, not the headline.
4. "Temporal confirmation weighting explains new memories ranking low." It is 0.0 by default and, per above, on a path CLI recall does not take.
5. "The knobs were built but never wired." They are wired — to the chiral path.

Decision

Phase 0 — Honest measurement (blocking)

All probes must use the non-observing path. recall_resonance calls apply_observation, which writes energy back: *"If you query the medium, you change it."* Diagnostic probing during this investigation repeatedly recalled the same memories and pumped their energy — amplifying the very bias under study. recall_resonance_readonly exists precisely for this and must be the harness's only entry point.

1. Labeled probe set: (query, expected_id, relevance), including paraphrase probes and near-miss distractors. Reuse recall_bench.rs::hash_label() so it is publishable without leaking content.
2. Metrics: Recall@k, MRR, nDCG. Baseline every path in the table above separately before combining anything.
3. Record time_since_last_dream with every measurement (see predictions).

Phase 1 — Carry phase out of the paths

Extend the recall result to carry amplitude and phase (or a complex amplitude) rather than a collapsed magnitude. Resonance is a public type consumed by the ADR-0029 envelope, the radio hub, and the observatory tangle fallback — so the field is additive with a default, and the existing scalar stays byte-identical for current consumers.

Phase 2 — Superposition combinator

Combine N paths by summing complex amplitudes and taking magnitude last, so agreeing paths reinforce and disagreeing paths cancel. Off by default (KANNAKA_RECALL_SUPERPOSE, default off, byte-identical), per the ADR-0048/0050 discipline. Compare three arms on the Phase 0 harness: single-path, classical rank fusion, phase superposition.

Phase 3 — Phase commensurability

Superposition is only meaningful if paths share a phase reference. Hemispheres carry independent phase state; callosal Kuramoto coupling during dream is the existing machinery that synchronizes them. This phase establishes whether a shared frame exists, and schedules consolidation (nothing currently does — local last_dream was 13 days stale).

Testable predictions

1. Anomaly decay. If phase alignment drives the "short-lived glimpses" of unusually good recall, quality should peak immediately after a dream and decay with time since. Falsifiable with the Phase 0 harness plus a dream schedule.
2. Superposition beats fusion, or it doesn't. If phase carries real information, complex summation should beat classical rank fusion on the same probe set. If it merely ties, phase is decorative and Phase 2 should be dropped.

Consequences and risks

Destructive cancellation is the headline risk. Classical fusion can only dilute a correct result; phase superposition can null it out. A wrong path in antiphase with a right one removes the right answer entirely. This is strictly more dangerous than the status quo and is the reason Phase 2 ships off by default behind a measured comparison.

Uncalibrated phase is noise. If cross-path phases are not commensurable, summation injects noise with the confidence of physics. Phase 3 is a precondition for trusting Phase 2, not an enhancement of it.

Public type change. Resonance crosses process boundaries (ADR-0029 envelope, radio hub, observatory). Additive-with-default only.

Not addressed here: encoder backend (supporting change), ADR-0049 facet decomposition (independent, proven, unshipped — likely the single highest-value separate work), codebook entropy quality (Codebook::new(384, 10_000, 42) and SimpleHashEncoder::new(384, 42) share the constant seed 42; whether basis near-orthogonality is materially affected, and whether QuantumOS's quantum_set_boot_entropy QPU path is worth wiring in, is its own experiment).

Open questions

1. Is per-wavefront phase semantically meaningful, or incidental? If phase is set arbitrarily at insert, interference is noise dressed as physics. This gates everything.
2. Does query_phase derive deterministically from query content? Two paraphrases must land near the same phase or superposition is unstable across wordings.
3. Should the flat path remain CLI default, or should the CLI itself superpose? A change of default recall behaviour affects every downstream consumer.
4. Can flat and chiral phases be compared at all without a callosal coupling pass having run recently?