◈ Kannaka Library
Kannaka Library / kannaka-quantum / ADR-0002 — Controlled-delay experiment: physical decoherence as a model of forgetting
kannaka-labs/kannaka-quantum docs/adr-0002-controlled-delay-decoherence-as-forgetting.md · 2026-09-08 · source ↗ · edit ↗

ADR-0002 — Controlled-delay experiment: physical decoherence as a model of forgetting

  • Status: Proposed
  • Date: 2026-09-06
  • Scope: kannaka_quantum/core.py (the per-minute device guard, run_circuit QASM handling), bench/ (a new decay ledger row), and one paid run on rigetti:rigetti:qpu:cepheus-1-108q via qBraid's direct Rigetti integration.
  • Related: docs/recall-is-amplitude-amplification.md (the correspondence this extends), ADR-0001 (spend discipline), kannaka-memory ADR-0057 (the LLM track, which this explicitly does not accelerate).

Context

qBraid's CTO pointed out (2026-07-20) that the Braket route we have used for every Cepheus run silently drops delay instructions; the direct rigetti:rigetti:qpu:… device routes them into Quil-T so idle time actually elapses. Our bridge refuses that device because it bills per minute (12,000 credits/min ≈ $120/min) and a per-job credit ceiling cannot be enforced from a shot count (core.py ~L373). The headline rate hides the real cost: jobs run tens to a few hundred milliseconds, so a 1,000-shot job is ≈ $0.40, prorated to the microsecond with no minimum.

Kannaka's central empirical finding of the summer was that forgetting is load-bearing (Φ 0.26 → 0.50 by pruning 1,298 memories to 28). The July hardware ledger showed the recall-as-amplification correspondence survives an ideal simulator (50/50) but collapses to 40 % on the chip at 4 qubits, where circuit depth, not idle time, is the loss. What we have never measured is the time axis: a stored amplitude left alone, then recalled.

This is research and story material, not compute. Nothing here speeds up kannaka-brain training or HRM recall at scale.

Decision

  1. Run a two-arm controlled-delay experiment on the direct Rigetti device.
    • Arm A (primary, single qubit): prepare |1⟩ (T1) and |+⟩ (T2\*), insert delay[t], measure. t ∈ {0, 2, 5, 10, 20, 50, 100} µs.
    • Arm B (the Kannaka arm): the 2-qubit recall circuit from the amplitude-amplification writeup with delay[t] between state preparation and the amplification step. The metric is quantum/classical agreement vs delay: an operational forgetting curve for quantum recall.
    • Echo control: Arm A repeated with one x pulse at t/2. Physically this is T2-echo; in Kannaka's vocabulary it is rehearsal mid-interval, and kannaka-crystal already found that dreaming early protects a memory while dreaming late destroys its addressing. The control separates "the medium loses the memory" from "the memory drifts and can be re-phased".
    • 500 shots per point (binomial SE ≈ 2.2 pts). 7 delays × 3 arms = 21 jobs, estimated ≤ $6; hard cap 800 credits ($8) for the whole run.
  2. Pre-register before running, kannaka-crystal style: publish the predicted shape (exponential decay; echo arm slower than free arm; Arm B agreement falling toward chance at 25 % as the 2-qubit state thermalises) as an OpenBotCity artifact with failure conditions first.
  3. Replace the per-minute refusal with a wall-clock ceiling. New max_seconds guard (default 1.0 s) required alongside allow_spend for per-minute devices; credit ceiling = rate/60 × max_seconds; the bridge records billed credits after each job and stops the run if any single job exceeds its ceiling. If qBraid's direct API exposes a pre-submit execution-time estimate or a server-side cap (asked 2026-09-06), use it and drop the client-side estimate. Until answered, the physics bounds the risk: 1,000 shots with a 100 µs delay adds ≈ 0.1 s ≈ $0.20.
  4. Abort conditions, checked first: if delay[100us] and delay[0] produce indistinguishable Arm A distributions, delays are being dropped on this route too and the run stops after those two jobs (≈ $0.80). If Arm B at t = 0 is below 40 % agreement the chip is worse than July's ledger and Arm B is skipped.

Consequences

  • One new ledger row in bench/LEDGER.md (decay curve, fitted T1/T2\*/T2-echo, Arm B agreement vs delay), one writeup, one Ghost Signals segment.
  • The bridge gains a per-minute spend path. It stays opt-in and the free simulator remains the default for every agent surface.
  • run_circuit must pass OpenQASM 3 delay through untouched; a unit test pins that the QASM sent to the direct device still contains it.
  • Not done here: no QUBO/consolidation work, no claim that HRM is quantum, and no change to the recall regression gate. The forgetting curve is a correspondence to be reported, not a mechanism to be asserted.