19-EFT.WP.Methods.SynthData v1.0 | Appendix B — Contract Library & Strategy Cards (C40 Full Set)

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

Appendix B — Contract Library & Strategy Cards (C40 Full Set)

I. Scope & How to Use

• This appendix consolidates the C40-* contracts and runtime strategy cards that serve as quality and compliance gates across the full lifecycle of synthetic data—training, evaluation, release, and streaming operations.
• Each contract is persisted as a quadruple — rule expression + threshold source + severity + action — and is mapped to manifest.synth.contracts.*.
• Cross-references follow the fixed form: “see EFT.WP.Core. / Methods.Cleaning / Methods.Imaging / Methods.CrossStats Chapter x S/P/M/I…”.

II. Contract Fields & Naming

• Core keys
  contracts[].id, name, expr, tol, severity ∈ {info, warn, block}, window, evidence, action_plan.
• Evidence & metrics
  metrics.* (e.g., W1, MMD, FID, KID, covg, utility_gap, MI_risk), u(metric) (uncertainty), alpha/power.
• Time base & arrival
  tau_mono, ts, T_arr_form1, T_arr_form2, delta_form, offset/skew/J.
• Privacy budget
  eps_total, delta_total, eps_budget, delta_budget.

III. Threshold Conventions & Generation Methods

• Reference baselines & adaptive thresholds
  Resampled intervals from the reference set D_ref: CI_1−alpha(metric) = bootstrap(metric; B); release threshold tol_metric = q_{1−alpha} + margin.
• Bayesian posterior thresholds
  p(metric ≤ tol | D) ≥ 1 − beta; posterior quantiles drive pass/fail; beta books the type-II risk budget.
• Differential privacy budget
  Composed accounting: (eps_total, delta_total) = privacy_accountant(events); the contract is eps_total ≤ eps_budget ∧ delta_total ≤ delta_budget.
• Dual arrival-time formulations
  delta_form = | ( 1 / c_ref ) * ( ∫ n_eff d ell ) − ( ∫ ( n_eff / c_ref ) d ell ) | ≤ tol_Tarr.

IV. Contract Library (Rules & Notes)

1. C40-01x Schema & Referential Integrity
   • C40-011 unique_pk: unique(pk) = true.
   • C40-012 fk_integrity: foreign_key(parent.pk, child.fk) = true.
   • C40-013 non_null: null_rate(field) ≤ tol_null.
   • C40-014 dim_check: check_dim( y − f(x) ) = pass.
2. C40-02x Time Base & Arrival Consistency
   • C40-021 mono_time: non_decreasing(tau_mono) = true.
   • C40-022 arrival_dual_forms: delta_form ≤ tol_Tarr.
   • C40-023 sync_jitter: |offset| ≤ tol_off ∧ |skew| ≤ tol_skew ∧ J ≤ tol_J.
3. C40-12x Fidelity (Distributional Closeness)
   • C40-121 wasserstein: W1(real, syn) ≤ tol_W1 (declare geometry/feature metric).
   • C40-122 mmd: MMD_k(real, syn) ≤ tol_MMD (declare kernel & bandwidth).
   • C40-123 fid_kid: FID ≤ tol_FID ∧ KID ≤ tol_KID (declare embedding net & layer).
   • C40-124 coverage: covg(real|syn) ≥ covg_min (explicit support-set definition).
4. C40-13x Utility (Downstream Non-Inferiority)
   • C40-131 non_inferiority: utility_gap = metric_syn − metric_real ≥ −tol_gap.
   • C40-132 power_min: power(design, effect_min, alpha) ≥ power_min (see CrossStats Ch.8).
   • C40-133 calibration_consistency: ECE_syn ≤ tol_ECE or |bias_syn| ≤ tol_bias.
5. C40-14x Privacy & Security
   • C40-141 dp_budget: eps_total ≤ eps_budget ∧ delta_total ≤ delta_budget.
   • C40-142 membership_risk: MI_risk ≤ tol_MI (declare attack families & eval set).
   • C40-143 k_anon: k_anonymity(QID) ≥ k_min or l_diversity ≥ l_min (tabular).
   • C40-144 linkability: linkability_score ≤ tol_link.
6. C40-15x Bias, Fairness & Representativeness
   • C40-151 group_repr: | p_syn(g) − p_ref(g) | ≤ tol_repr for all groups g.
   • C40-152 psi_drift: psi(ref, syn) ≤ tol_psi (population stability index).
   • C40-153 threshold_parity: | t_syn(g) − t_syn(g') | ≤ tol_tpar (when a published threshold must be cross-group consistent).
7. C40-16x Multimodal Coherence & Physical/Geometric Constraints
   • C40-161 cross_modal_coherence: coherence_rules(bundle) = pass.
   • C40-162 physical_laws: e.g., mass_balance_resid ≤ tol_res, energy_conserve_resid ≤ tol_res.
   • C40-163 geometry_consistency: reproj_err ≤ tol_px (see Imaging Ch.9).
8. C40-17x Watermark & Provenance
   • C40-171 watermark_detect: TPR ≥ tpr_min @ FPR ≤ fpr_max (declare detector & threshold).
   • C40-172 provenance_hash: hash_sha256(blob) = manifest.signature.blob_hash.
9. C40-18x Runtime SLOs (Streaming)
   • C40-181 stability: rho = lambda_in / mu_eff < 1.
   • C40-182 latency_p99: latency_ms_p99 ≤ target_p99.
   • C40-183 drop_rate: drop_rate ≤ tol_drop.
   • C40-184 drift_guard: W1_cur ≤ tol_W1_run ∧ psi_cur ≤ tol_psi_run (windowed Delta_t).

V. Strategy Card Template (for any C40-*)

1. card.id: SC40-xxx.
2. intent: summary of the contract’s purpose and business impact.
3. scope: data domain/modality/window Delta_t/channels.
4. inputs: required metrics, reference sets, priors or models.
5. test: the contract expression expr and threshold source (bootstrap | posterior | policy_fixed).
6. tiers: info / warn / block with tiered thresholds or posterior lower bounds.
7. actions (automated response map):
   • info: log, visualize, lake ingestion.
   • warn: resample/reweight/parameter rollback, de-weighted release.
   • block: circuit-break, roll back to prior version, queue human review.
8. runtime_hooks: whether to link with stream_synth backpressure, alpha_spending, privacy_accountant.
9. logging: TraceID, evidence snapshots, seed, version signature.

VI. Sample Strategy Cards (Selected)

1. SC40-121 (FID/KID Fidelity Gate)
   • test: FID ≤ tol_FID ∧ KID ≤ tol_KID, with tol_* from bootstrap(FID,KID; B) quantile q_{1−alpha}.
   • tiers: warn when tol_warn < metric ≤ tol_block; block when metric > tol_block.
   • actions: warn → halve frozen thresholds and resample once; block → roll back to Engine.version−1.
2. SC40-141 (DP Budget Guard)
   • test: eps_total ≤ eps_budget ∧ delta_total ≤ delta_budget.
   • actions: warn → pause non-essential sampling; block → trip training & sampling, initiate budget approval.
3. SC40-022 (Dual Arrival-Form Consistency)
   • test: delta_form ≤ tol_Tarr.
   • actions: warn → increase integration steps and re-align; block → quarantine samples & path segments, mark non-releasable.
4. SC40-151 (Group Representativeness)
   • test: | p_syn(g) − p_ref(g) | ≤ tol_repr for all g.
   • actions: warn → balance_distribution(reweight); block → reject release and emit gap analysis.
5. SC40-182 (P99 Latency Constraint)
   • test: latency_ms_p99 ≤ target_p99.
   • actions: warn → tighten rate_limit and enable auto_batcher; block → switch to read-only plus offline queues.

VII. Composite Contracts & Decision Logic

• Default gate: pass = ∧_{family} pass_family, where family ∈ {schema, timepath, fidelity, utility, privacy, bias, physics, watermark, runtime}.
• Priority order (high → low): privacy > runtime.safety > timepath > schema > physics > fidelity > utility > bias.
• Conflict resolution: if utility passes but fidelity is slightly short, allow de-weighted release only; if privacy fails, it must block.

VIII. Manifest Mapping & Persistence

• manifest.synth.contracts.*: per-contract id, expr, tol, severity, result, evidence_ref.
• manifest.synth.metrics.*: corresponding metrics with uncertainty u(metric) and declared computation gauges.
• manifest.synth.runtime.*: windowed Delta_t of rho, latency_ms_p99, drop_rate, drift.
• manifest.signature: hash_sha256(blob), signature, TraceID, seed.

IX. Implementation Notes & Automation Hooks

• Unified evaluation entry: assert_synth_contract(ds_syn, rules) executes a selected subset from this appendix’s contract library.
• Runtime hooks: integrate with stream_synth, detect_drift, privacy_accountant, emit_runtime_manifest.
• Failure tiers: warn triggers online remediation; block triggers version rollback and a human review queue.

X. Cross-References

• Schema/manifests & dimensions: EFT.WP.Core.DataSpec v1.0 and Methods.Cleaning v1.0 Chapters 4/10.
• Arrival time & paths: Methods.Cleaning v1.0 Chapters 6/12.
• Imaging constraints & geometric coherence: Methods.Imaging v1.0 Chapters 9/14.
• Statistical thresholds, power & drift: Methods.CrossStats v1.0 Chapters 5/6/7/8/14.