19-EFT.WP.Methods.SynthData v1.0 | Chapter 15 — Use Cases & Reference Implementations

Home ／ Docs-Technical WhitePaper ／ 19-EFT.WP.Methods.SynthData v1.0

Chapter 15 — Use Cases & Reference Implementations

One-line objective: Provide three end-to-end reference paths—tabular privacy synthesis, imaging simulation + learning, and online replay A/B sandbox—delivered and audited with unified P/S/M/I/contracts.

I. Scope & Targets

1. Scope
   • End-to-end exemplars from requirements → design → generation → evaluation → release, covering both offline and streaming modes.
   • Each use case emits manifest.synth.*, contract evaluations, and signatures for replay and audit.
2. Objects
   • D_real, D_syn, SynthSpec, engine, policy.*, contracts.*, TraceID.
   • Quality & risk surfaces: fidelity, utility, privacy, timepath, drift, SLO.
3. Boundaries
   • Vendor-agnostic; interfaces follow I40-* and the unified manifest schema.
   • Constraints: check_dim(expr) passes; delta_form ≤ tol_Tarr; eps_total ≤ eps_budget.

II. Terms & Variables

• Data & distributions: p_data, p_model(x; theta), D_real, D_syn, N_real, N_syn.
• Evaluation: W1, MMD, FID, KID, covg, downstream_metric.
• Privacy: eps, delta, eps_total, MI_risk.
• Time base & arrival: tau_mono, ts, T_arr, gamma(ell), delta_form, Delta_t.
• Runtime: B, K, latency_ms_p99, rho, drop_rate.
• Contracts & manifest: contracts.*, manifest.synth.*, signature.

*III. Axioms P415- **

• P415-1 (Reproducible Experiments): Each use case must be uniquely replayable from {seed, engine/version, SynthSpec}.
• P415-2 (Freeze Gauges Upfront): Lock metrics and thresholds before generation; no “moving the goalposts” prior to release.
• P415-3 (No Leakage): Traceable, disjoint splits for training/validation/test on D_real and evaluation sets.
• P415-4 (Time/Path Consistency): For temporal samples, record both T_arr formulations and verify delta_form.
• P415-5 (Closed-Loop Privacy): Account eps_total end-to-end; on breach, trip the circuit and provide a rollback bundle.
• P415-6 (Contract Gate): assert_synth_contract must pass before freeze_release_synth.

*IV. Minimal Equations S415- **

• S415-1 (Fidelity Distances):
  W1 = ( ∫ | F_real(x) - F_syn(x) | dx )；MMD = || ( 1/N ∑ phi(x_i) - 1/M ∑ phi(x'_j) ) ||^2。
• S415-2 (Utility Preservation):
  utility_gap = | metric(model, D_real) - metric(model, D_syn) |。
• S415-3 (Coverage):
  covg = ( | support(D_syn) ∩ support(D_real) | ) / | support(D_real) |。
• S415-4 (Privacy Accounting):
  eps_total = accountant( { (eps_i, delta_i) } )。
• S415-5 (Dual Arrival Forms):
  T_arr = ( 1 / c_ref ) * ( ∫ n_eff d ell ) and T_arr = ( ∫ ( n_eff / c_ref ) d ell )，delta_form = | … |。
• S415-6 (Runtime Steady State):
  rho = lambda_in / mu_eff < 1，latency ≈ W_q + T_serv(B)。

V. Metrology Flow M40-15 (Three End-to-End Paths)

M40-15a — Tabular Privacy Synthesis (Offline)

• Freeze requirements: target fields, constraints, eps_budget, thresholds {W1, MMD, utility_gap}.
• Schema binding: design_synth_spec; register_schema; validate_dataset.
• Engine fitting: fit_engine(ds=D_real, model=copula|flow, privacy=DP(eps,delta)).
• Sampling: sample(engine, n=N_syn, seed); enforce_constraints (uniqueness/foreign keys/ranges).
• Evaluate & align: measure_fidelity (W1, MMD, covg); measure_privacy (MI_risk); apply balance_distribution if needed.
• Contract gate: assert_synth_contract; sign and freeze_release_synth.
• Persist manifest: manifest.synth with TraceID, seed, eps_total, metrics.*.

M40-15b — Imaging Simulation + Learning (Offline + Downstream)

• Freeze requirements: imaging chain & task (e.g., seg/det), physical constraints, calibration gauges.
• Physics/scene: compose_multimodal; build scene graph G=(V,E) and parameter sweeps (PRNU/DSNU, PSF/MTF per Imaging volume).
• Render & align: generate D_syn(image, label); align_timepath and stamp both T_arr forms.
• Downstream eval: compare downstream_metric on real validation with synthetic-trained models for utility_gap.
• Bias-fix & regenerate: balance_distribution (illumination/pose/scale) until metrics converge.
• Contract & release: assert_synth_contract → freeze_release_synth.
• Manifest & provenance: record spectral/geometric gauges, device calibration citations, manifest.synth.imaging.*.

M40-15c — Online Synthetic Replay A/B Sandbox (Streaming)

• Ingress: stand up stream_synth per policy.runtime, with targets latency_ms_p99 and rho_max.
• Randomization: stratified randomization and exposure control for incoming requests (see CrossStats Ch.8).
• Replay & injection: sample(engine, n, condition) to inject synthetic events; timepath_hardening to bound delta_form.
• Experiment readout: run_ab_test(stream, metric, alpha_spending); detect_drift continuously.
• Accounting & circuit breaking: privacy_accountant; on breach, circuit_open and degrade.
• Audit & release: emit_runtime_manifest; periodically freeze bundles and reports.

VI. Contracts & Assertions

1. C40-1501 Tabular Privacy Synthesis
   • W1 ≤ W1_max, MMD ≤ mmd_max, covg ≥ covg_min;
   • utility_gap ≤ gap_max; eps_total ≤ eps_budget; unique(pk) and foreign_key all pass.
2. C40-1502 Imaging Simulation + Learning
   • utility_gap ≤ gap_max; FID|KID ≤ fid_max;
   • Optical & metrology gauges consistent (see Imaging Ch.4/5); delta_form ≤ tol_Tarr.
3. C40-1503 Online Replay A/B Sandbox
   • latency_ms_p99 ≤ target_p99; rho ≤ rho_max; drop_rate ≤ tol_drop;
   • alpha_spent ≤ alpha_budget; eps_total ≤ eps_budget.

VII. Implementation Bindings I40-15*

• Tabular Privacy Synthesis
  I40-31 design_synth_spec → I40-41 fit_engine → I40-52 sample → I40-61 enforce_constraints → I40-121 measure_fidelity → I40-101 measure_privacy → I40-131 assert_synth_contract → I40-141 freeze_release_synth。
• Imaging Simulation + Learning
  I40-61 compose_multimodal → I40-81 align_timepath → I40-121 measure_fidelity(FID|KID) → I40-111 balance_distribution → I40-131 assert_synth_contract → I40-141 freeze_release_synth。
• Online A/B Sandbox
  I40-141 stream_synth → I40-151 privacy_accountant → I40-171 detect_drift → I40-181 run_ab_test → I40-191 emit_runtime_manifest。
• Invariants: idempotent(TraceID, seed), rho < 1, delta_form ≤ tol_Tarr, eps_total ≤ eps_budget.

VIII. Cross-References

• EFT.WP.Methods.Cleaning v1.0: Chapters 10/11 (compliance/freeze, streaming & backpressure) for release and runtime conventions.
• EFT.WP.Methods.Imaging v1.0: Chapters 4/5/9/11 (metrology, PSF/MTF, geometry, HDR) for simulation gauges.
• EFT.WP.Methods.CrossStats v1.0: Chapters 7/8/14 (drift, A/B, SLO) for monitoring & audit.
• EFT.WP.Core.DataSpec v1.0: manifests and primary/foreign key contracts.

IX. Quality SLIs & Risk Control

1. SLIs
   • Offline: W1, MMD, FID|KID, utility_gap, covg.
   • Streaming: TS.sli.rho, TS.sli.latency_p99, TS.sli.drop_rate, TS.sli.delta_form, TS.sli.eps_total.
2. Risk & rollback
   • Fidelity breach: trigger reweight/resample; if needed, reduce complexity or roll back to a prior engine.
   • Privacy breach: open the circuit; switch to low-sensitivity templates or halt experiments.
   • Arrival breach: run timepath_hardening and rate-limit time-sensitive modalities.
   • Runtime breach: rate_limit, auto_batcher, degrade(profile), cross-zone migration.

Summary

• This chapter binds P40x-/S40x-/M40-/I40-/C40-* into reproducible practice through three reference paths.
• The resulting manifest.synth.* and audit reports can be plugged directly into catalogs and compliance modules.
• Contract gates and rollback strategies ensure an engineering-optimal balance across fidelity, utility, privacy, arrival time, and SLOs.