quantum-hypothesis-simulation

Quantum Hypothesis Simulation

The quantum-probability layer for YURI's claim/pulse machinery, and tier 3 of the simulation arsenal. It models hypotheses as a superposition in a real-valued Hilbert space (ℝ^N) that only projects (collapses) when evidence is applied — so the ORDER evidence arrives in changes the posterior. That is the one thing a classical order-blind Bayes update structurally cannot represent, and it is the whole point: when P(H | A then B) ≠ P(H | B then A), this is the right instrument.

Engine: _SYSTEM/Scripts/quantum-hypothesis-tracker.mjs (313 lines, 19 exports). Falsification gate: _SYSTEM/Scripts/quantum-vs-bayes-benchmark.mjs. GATE re-verified PASS 2026-06-14 on current code (G1–G4 all true; tracker unit tests 5/5; eval-processing 11/11; validated at a 15M-eval scale). Wired as an instrument of probabilistic-decision-core. Living system — actively being built; the public export API is stable (method-map below intact) but internals + the surrounding arsenal keep moving, so re-read the source before relying on a fine detail.

The simulation arsenal (where this sits) — `faeb5b67` (arsenal) → `15cfc088` (learn-loop closed)

A decision-under-uncertainty pipeline: measure → robustify → commit → learn.

izanagi-simulator (/izanagi) — qualitative 3-branch commit (EV × reversibility × blast). The fast "should I even branch" front.
decision-sim.mjs — robust optimization: robustScore (0.5·mean + 0.5·CVaR), minimaxRegret, pgdWitness (flip-rule), crossEntropyOptimize, infoGapHorizon, multiverse, + halton/sobolish QMC. Reproducible (seeded makeRng). 7/7 green.
quantum-hypothesis-tracker.mjs ← you are here — order-aware evidence (below).
izanagi-bridge.mjs (@capability: izanagi-decision-bridge) — turns measured/MC option values into a robust ruling: CVaR + minimax-regret + PGD with an AUTOMATIC corner-law vertex guard (enumerates paramSpace vertices where an affine/multilinear worst case hides from interior sampling) + per-axis flip thresholds. Demo-validated (runDemo), no unit test yet. This is izanagi made computational.
- Learn-loop edge: izanagiRuling(…, {record:true}) writes the prediction to prediction-ledger.mjs (opt-in, off by default). CLOSED 2026-06-13 (15cfc088): prediction-outcome-resolver.mjs (@capability) re-runs propagation-scan on aged unresolved predictions, scores predicted-vs-observed, records outcomes + populates calibration — idempotent, execFileSync-safe, REPO_ROOT-anchored, live-verified (resolved=1, calibration n=2) + guard test. Remaining edge: a SCHEDULED age-sweep cadence (the resolver runs on-demand; a cron/launchd beat to sweep aged predictions is owner-config).

Reach for izanagi-bridge for "which option do I build / commit under uncertainty"; reach for this quantum tier when the evidence order itself carries information.

Use When (quantum tier specifically)

Evidence is sequential and order-sensitive — a later observation reframes an earlier one (question-order, framing, anchoring, path-dependent diagnosis).
Hypotheses interfere / are non-commuting (the projectors don't commute), so a product-of-likelihoods Bayes update loses information.
You need the Schmidt coupling test — the cross-reference engine's mathematical criterion for whether two subsystems are genuinely coupled vs separable.
You suspect a classical model is silently missing an order-effect and want a falsifiable test of it.

Skip When

Evidence is order-independent and exchangeable → plain Bayes (bayesSequential) is correct and cheaper; don't reach for quantum to look sophisticated.
You want robust optimization over options, not order-effects → that's decision-sim / izanagi-bridge.
The phases need to be complex (full ℂ^N interference) — this engine is ℝ^N only, phases 0 or π. Adequate for real-valued projectors; not a complex-interference model.

Method → question map (quantum tracker)

You want…	Function	What it gives
Posterior after an ORDERED evidence sequence	`hypothesisPosteriors(state, hypotheses, evidence)`	applies each projector in order (collapsing the state), then reads P(H) — order-dependent
The raw sequential measurement	`measureSequential(projectors, psi)`	the order-effect core: project, renormalize, repeat
The falsifiable order-effect signature	`qqEquality(state, P_A, P_B)`	QQ statistic `sAB − sBA`; the quantum model guarantees ≈ 0 — the Wang-Busemeyer test
Are two subsystems genuinely COUPLED	`schmidtDecomposition(psi_AB, m, n)`	Schmidt spectrum (Jacobi SVD) — the cross-ref engine's coupling criterion (separable ⇒ uncoupled)
The classical baseline you must BEAT	`bayesPosterior` / `bayesSequential`	order-blind posterior — the control
Run the whole falsification gate	`runBenchmark()` (benchmark module)	G1–G4 verdict

Plus the ℝ^N primitives (dot/norm/normalize/matVec/matMul/projector/diagonalProjector/measure/stateVector). Pure, dependency-free, owner-gate-free — run as a throwaway harness or wire into a sim test.

Eval-processing — turning a million-eval budget into a decision (`eval-processing.mjs`, NEW)

Reframe (the juicy part): an eval count is a sampling budget, not a result. Needing millions usually means you allocated badly — good design resolves most decisions at thousands; millions are only for rare tails, fine grids, or hi-dim corners. "Processing" collapses N evals → one decision + flip rule + residual CI. The funnel (4 capabilities, 11/11 green as of 15cfc088, all surface via capability-recall); 15cfc088 added residuals — pairedDelta computeUnpaired:false (skip the 2× baseline in hot loops) + confidenceSequence sub-Gaussian fallback (range:null for unbounded data, α-control validated on a Gaussian null):

mkAggregator (streaming-aggregator) — Welford moments + Vitter reservoir, O(1) memory, never store the rows.
pairedDelta (crn-paired-delta) — common-random-numbers → variance-reduced A/B delta CI (the biggest free win; computes the unpaired CI too as proof).
confidenceSequence / sequentialDecide (sequential-stopping) — empirical-Bernstein confidence sequence, peek-valid (α-control empirically validated, not just cited; distinct from quantum measureSequential).
heldOutSplit / kFold / conformalQuantile / inSampleVsHeldout (heldout-split) — kills the in-sample ECE=0 optimism trap.

The CI width + flip location are what made the 16M-eval cap-sim honest — not the count.

The proof discipline (this is the precision, not decoration)

A model that "wins everywhere" is overfitting, not an effect. Gates are domain-blind and two-sided — both halves must hold:

G2 (earns its keep): on real order-effect data (Gallup Clinton/Gore marginals: Clinton 50→57%, Gore 68→60%) quantum RMS must be ≥50% below the best classical static model.
G3 (honesty/control): on synthetic NO-order-effect data quantum must NOT beat classical by more than 0.005 RMS. A win on the control = laundering flexibility as signal → reject.
G1 machinery recovery; G4 QQ residual on the real joint < 0.05.
DATA FLAG: the Clinton/Gore JOINT cells are literature-recalled (Wang & Busemeyer 2013) → OWNER-VERIFY; the marginals are robustly attested and carry G2 on their own.
Corner law (automated downstream): for an affine/multilinear objective the worst case sits at a VERTEX an interior Dirichlet/uniform sampler hits with probability 0 — izanagi-bridge.cornerAwareReadout enumerates vertices so a flip can't hide. See [[feedback-affine-objective-enumerate-corners]].
Promotion path: must beat the classical Bayes baseline on order-sensitive logged sequences before touching any live organ. Until then, advisory.

Boundaries

Output is advisory until verified against live evidence — the falsification gate IS the verification; don't claim a quantum win without the two-sided G2+G3 pass.
ℝ^N real-valued only; don't claim complex-interference behavior it doesn't have.
The arsenal is still being built — value models / bridge constants are hand-encoded (advisory, not measured truth); the learn loop is closed on-demand (prediction-outcome-resolver scores predictions→outcomes), with only a SCHEDULED sweep cadence still owner-config.

Pair with

izanagi-bridge.mjs (@capability: izanagi-decision-bridge) — the measure→robustify→commit step; corner-law-guarded ruling from option values.
decision-sim.mjs — robust optimization tier (CVaR / regret / flip-rule / info-gap / multiverse).
eval-processing.mjs — the funnel that turns a quantum/MC sim's evals into an honest decision+CI.
prediction-outcome-resolver.mjs (@capability) — closes the learn loop: scores recorded predictions against observed outcomes (re-runs propagation-scan), populates calibration. Pairs with izanagi-bridge's {record:true} ledger writes.
probabilistic-decision-core (/pdc) — the EV/calibration discipline the tracker sharpens.
cross-reference-navigation (/xref) — the Schmidt coupling test is the cross-ref engine's coupling criterion; use together when judging whether two mechanisms are truly linked.
izanagi-simulator (/izanagi) — the qualitative 3-branch front.
Recall surface: memory [[ref-simulation-arsenal]] (the 4-tier map + roadmap) · [[ref-capability-first-wiring]].

Session Notes

2026-06-14

session: 101m | peak ctx: 0% | compacts: 0
tools: Bash×455, Read×142, Edit×74, Write×44, WebSearch×31, WebFetch×19, StructuredOutput×17, ToolSearch×8, TodoWrite×5, Workflow×3, AskUserQuestion×2, Skill×2, TaskStop×1
corrections: Base directory for this skill: /Users/marcelspatz/.claude/skills/cross-reference-navigation

Cross-Reference Navigation (XREF)

The GROUND step of the work loop, made reflexive. One question asked a | Base directory for this skill: /Users/marcelspatz/.claude/skills/quantum-hypothesis-simulation