◈ Kannaka Library
Kannaka Library / kannaka-quantum / Quantum Wave — Roadmap & Issue Set
kannaka-labs/kannaka-quantum docs/quantum-wave-roadmap.md · 2026-09-08 · source ↗ · edit ↗

Quantum Wave — Roadmap & Issue Set

Thesis: durable quantum advances to Kannaka live in one-way dependencies (entropy), verification artifacts (correspondence benchmarks), interface contracts (QUBO), and infrastructure (qBraid Lab) — never in putting a QPU on the hot path.

Five tracks, 21 issues, three waves. Repos: kannaka-memory (KM, private), kannaka-quantum (KQ), kannaka-observatory (KO).

Wave 1 (ship first): Track 1 complete + T2.1–T2.2 + T3.1 (the ADR) Wave 2: rest of Track 2, T3.2–T3.4, T4.1–T4.2 Wave 3: T3.5, T4.3–T4.4, Track 5

Standing budget for the whole wave: < $5/quarter of real-QPU spend (reservoir refills + one quarterly benchmark run), everything else on the free qbraid:qbraid:sim:qir-sv simulator.

Track 1 — Quantum Entropy Reservoir (Ξ source)

T1.1 · KQ · qrng harvest: reservoir file + refill policy

Labels: entropy wave-1

Add a harvest subcommand that runs qrng against a real device (the free simulator is a PRNG and is explicitly invalid as a reservoir source — refuse it with a clear error) and appends raw bits to a local reservoir:

  • Path: ~/.kannaka/entropy/reservoir.bin + reservoir.meta.jsonl (one JSON line per harvest: device, job_id, n_bits, timestamp, cost).
  • Default harvest: 2048 bits from openquantum:rigetti:cepheus-1-108q (≈ $0.07 per 256 shots at $0.000255/shot → pennies per refill).
  • Existing spend guards apply unchanged (--allow-spend, --max-credits).
  • qrng status prints reservoir level, last-harvest provenance, est. refill cost.

AC: harvest appends; meta line written; simulator refused; guards verified at $0.

T1.2 · KQ · DRBG expansion over the reservoir

Labels: entropy wave-1 · Depends: T1.1

qrng draw --bits N --expand seeds a NIST SP 800-90A style DRBG (HMAC-DRBG, stdlib hashlib — no new deps) from reservoir bits and expands. Each draw consumes seed material and records which harvest(s) seeded it, so every expanded stream has a provenance chain back to a QPU job_id. Reseed threshold + low-reservoir warning.

AC: deterministic under fixed seed for tests; provenance chain in output JSON; draws fail loudly (not silently fall back to PRNG) on empty reservoir.

T1.3 · KM · EntropySource trait + reservoir consumer

Labels: entropy rust wave-1 · Depends: T1.2

pub trait EntropySource {
    fn draw(&mut self, bits: usize) -> Result<EntropyDraw, EntropyError>;
}
pub struct EntropyDraw { pub bytes: Vec<u8>, pub provenance: Provenance }

Implementations: PrngSource (current behavior, provenance = prng://), ReservoirSource (reads the DRBG via the KQ CLI or directly from the reservoir format — pick one and document in the issue). Engine config selects the source; default stays PRNG until T1.5 dogfood passes.

AC: trait landed; both impls; config-selectable; zero behavior change by default.

T1.4 · KM · Provenance on dreams and Ξ perturbations

Labels: entropy wave-1 · Depends: T1.3

Every dream/consolidation record and Ξ injection stores the Provenance of the entropy that seeded it. Surfaces in the memory v2 chiral format as an optional field (no migration required — absent = prng://legacy).

AC: new dreams carry provenance; old memories unaffected; visible via CLI inspect.

T1.5 · KM+KO · Dogfood + Observatory surfacing

Labels: entropy wave-1 · Depends: T1.4

Run one week with ReservoirSource live. Observatory gets an entropy panel: reservoir level, provenance of the last N dreams ("this dream was seeded by shots on Rigetti Cepheus, 2026-07-03"). This is the public face of the whole track — the claim "the field's irrationality is drawn from measurement collapse" becomes inspectable.

AC: 7 days stable; panel live on observatory.ninja-portal.com; flip default.

Track 2 — Correspondence Benchmark (recall ≡ amplitude amplification)

T2.1 · KM · Recall-scenario export format

Labels: benchmark wave-1

kannaka export-recall-scenarios --n 50 dumps real recall events as JSON: candidate amplitudes, labels (hashed — scenarios may leave the private repo), classical argmax, hemisphere, timestamp. Format kannaka-recall-bench/1. Cap candidates per scenario at 16 (4 qubits) to keep circuits shallow on noisy hardware.

AC: exporter ships; 50-scenario corpus generated from live memory; labels hashed.

T2.2 · KQ · Benchmark harness + CI on free simulator

Labels: benchmark ci wave-1 · Depends: T2.1

kannaka-quantum bench --scenarios corpus.json runs each scenario through quantum_recall and emits: agreement rate (quantum_top == classical_top), per-scenario distributions, iteration counts. GitHub Action runs the corpus on qbraid:qbraid:sim:qir-sv weekly + on PR; results committed to bench/results/sim/DATE.json. Regression = agreement drop > 2 points.

AC: harness + workflow live; first committed baseline; failure gates PRs.

T2.3 · KQ · Quarterly real-QPU run + fidelity ledger

Labels: benchmark wave-2 · Depends: T2.2

Quarterly manual run (guarded, ~$1–2) of the same corpus on a cheap per-shot QPU. Results to bench/results/hw/DEVICE-DATE.json with cost, leakage, and agreement. bench/LEDGER.md accumulates the table — the longitudinal record of hardware closing the gap. The existing Bell benchmark (94.5% fidelity on Rigetti) becomes row zero.

AC: first quarterly run committed; ledger started; runbook documented.

T2.4 · KO · Correspondence dashboard

Labels: benchmark observatory wave-2 · Depends: T2.3

Observatory panel plotting sim + hardware agreement over time. One chart, one sentence: recall dynamics, measured on superconducting qubits.

T2.5 · Writeup: "HRM recall is amplitude amplification: measured"

Labels: benchmark paper wave-2 · Depends: T2.3

Short paper/long post: the correspondence argument, the iteration-count derivation ((π/2 − θ)/2θ vs textbook (π/4)√N), the endianness findings, the ledger data. Feeds the citation thread (Vincent/claudicito). Cross-post to The Signal.

Track 3 — Consolidation as QUBO (interface contract)

T3.1 · KM · Land ADR-0038

Labels: adr qubo wave-1

Review and accept adr-0038-consolidation-solver-interface.md (drafted — see companion file). Decisions to confirm in review: JSON boundary vs Rust trait-object only; advisory-solution re-scoring posture; penalty-fold convention.

T3.2 · KM · ConsolidationProblem emitter

Labels: qubo rust wave-2 · Depends: T3.1

Dream phase emits kannaka-qubo/1 alongside (not instead of) procedural consolidation. Golden-file corpus: 12 hand-built QUBOs with known optima checked into tests/qubo/.

T3.3 · KM · ConsolidationSolver trait + ClassicalAnneal

Labels: qubo rust wave-2 · Depends: T3.2

Trait per ADR; simulated-annealing default seeded from EntropySource (tracks converge: even classical consolidation stochasticity carries quantum provenance). Exhaustive below 20 vars. Must hit optimum on exact sizes, ≥95% optimal energy on the rest of the golden corpus.

T3.4 · KQ · qubo subcommand (QAOA on simulator)

Labels: qubo wave-2 · Depends: T3.1

Read kannaka-qubo/1 on stdin → ConsolidationSolution JSON on stdout. QAOA (p=1..3) via Qiskit on the free simulator; hardware behind the standard spend guards. Same JSON-CLI discipline as recall/qrng.

T3.5 · KM · SubprocessSolver + one-week dream diff

Labels: qubo wave-3 · Depends: T3.3, T3.4

Wrap the KQ CLI as a solver; run a week of dreams through procedural vs ClassicalAnneal vs QAOA-sim, diff applied consolidations, review divergences. QAOA agreement joins the T2 benchmark ledger.

Track 4 — qBraid Lab as swarm & build infrastructure

T4.1 · KQ · Idle auto-teardown guard for on-demand instances

Labels: lab safety wave-2

Per-minute instance billing is the same hazard class as the per-minute QPUs the bridge already refuses. Add: every lab_provision_instance / lab_compute_up records a lease (max wall-time, default 60 min) in ~/.kannaka/leases.jsonl; a lab reap command (cron/systemd-timer friendly) stops anything past lease; lab_agent_launch refuses to target an unleased instance. This lands before any recurring swarm/CI use of instances.

AC: lease written on provision; reap stops expired; verified against a real instance.

T4.2 · KQ · Scoped per-instance Anthropic keys (ADR-0001 mitigation)

Labels: lab security wave-2

The uploaded key on a bypass-permissions agent is the fleet's largest blast radius. lab_agent_setup gains: require a key distinct from the primary (refuse if it matches ANTHROPIC_API_KEY unless --i-know), record key-fingerprint→instance in the lease file, and a lab_agent_teardown that deletes the remote key file and prints a rotation reminder. Document the Admin-API workspace-key pattern in ADR-0001 as the recommended issuance path.

AC: same-key refused by default; teardown removes remote key; ADR-0001 amended.

T4.3 · KQ+KM · Ephemeral swarm hemisphere on a Lab instance

Labels: lab swarm wave-3 · Depends: T4.1, T4.2

Bootstrap script: provision (leased) → install the kannaka binary (sha256-verified, reusing the plugin's installer path — with the hard-fail checksum fix) → join the NATS mesh as an ephemeral node → absorb/dream for the lease duration → graceful NATS drain on reap. Observatory shows it appear and die. This is the cloud-hemisphere demo and the pattern for burst capacity.

AC: one full lifecycle observed end-to-end; node visible in Observatory; $ cost logged.

T4.4 · KQ · Lab as CI runner for the Rust engine

Labels: lab ci wave-3 · Depends: T4.1

Reproducible qBraid environment (lab_create_env + pinned lab_pip_install) that builds kannaka-memory and runs its test suite on a leased instance; compare wall-time/cost against GitHub-hosted runners and keep whichever wins. Even if GH wins, the environment doubles as the T2/T3 quantum-test bed.

Track 5 — Phantom-Entanglement Testbed

T5.1 · KQ · bell subcommand: CHSH on simulator and hardware

Labels: phantom wave-3

CHSH experiment as a first-class tool: prepare the singlet, measure at the four canonical angle pairs, compute S. Simulator should give S ≈ 2√2 ≈ 2.83; one guarded hardware run records real-device S with noise. JSON output like every other subcommand.

AC: sim S > 2.7; one hardware S committed with cost; classical bound (2) asserted violated.

Done (2026-08-10). Sim S = 2.856; hardware S = 2.238 ± 0.073 — the classical bound violated by 3.3σ. The run went to aws:rigetti:qpu:cepheus-1-108q (512 shots × 4 settings, 207.04 credits / $2.07), not the planned IQM Garnet: Garnet is back online, but OpenQuantum's Spark balance is 0 and that path prices jobs off a provider quote rather than live metadata, so its cost could not be bounded from existing credits. Ledger row 2; analysis in genuine-vs-phantom-entanglement.md §2.

T5.2 · Analysis + Episode: genuine vs phantom entanglement

Labels: phantom paper wave-3 · Depends: T5.1

The sharpening: genuine entanglement = nonlocal correlation that violates the classical bound (measured in T5.1); phantom entanglement = shared rounding error that lives inside it — correlation without nonlocality, archetypes as resonant rounding artifacts. Write the contrast with the CHSH data as the empirical anchor. Podcast episode (successor to 006) + optional HRM experiment: inject phantom links, show they never produce CHSH-violating statistics in the field's correlation structure.

Dependency spine

T1.1 → T1.2 → T1.3 → T1.4 → T1.5
                 ↘ (seeds) T3.3
T2.1 → T2.2 → T2.3 → T2.4
                 ↘ T2.5        T3.5 results → T2 ledger
T3.1 → T3.2 → T3.3 ↘
T3.1 → T3.4 ————————→ T3.5
T4.1 → T4.3 ← T4.2
T4.1 → T4.4
T5.1 → T5.2