15-EFT.WP.Methods.Falsification v1.0 | Chapter 10: Compliance Templates & the Audit Trail

Home ／ Docs-Technical WhitePaper ／ 15-EFT.WP.Methods.Falsification v1.0

Chapter 10: Compliance Templates & the Audit Trail

I. Scope & Objectives

• Define the compliance deliverable templates, audit trail, and forensic interfaces for falsification work, covering Pre-Registration (PreReg), the Evidence.bundle, the environment lock EnvLock with signatures *.sig, and long-horizon archiving & replay forensics.
  The goal is to demonstrate — with a minimal yet sufficient evidence set — that under a locked environment and declared gates a valid falsification campaign was executed, and that the evidence is traceable, verifiable, replayable, and non-repudiable.
• Interfaces and flows remain consistent with Chapter 7 (statistical tests), Chapter 8 (uncertainty), and Chapter 9 (online gating): all statistical conclusions are bound to alpha/beta/power and the online alpha_spend(t); risk measures and gating decisions are fixed by the shared time base ts = alpha + beta * tau_mono and EnvLock.

II. Terms & Symbols

1. Compliance cards & evidence objects
   • PreReg.card: pre-registration card containing H0/H1, thresholds, sample size, alpha/beta, the alpha-spending scheme, and AnalysisPlan.hash.
   • Evidence.bundle: minimal evidence set with fingerprint; see S52-49.
   • AuditTrail: the event chain with incremental hashes and signatures; see S52-46/47.
   • EnvLock: environment lock pinning Graph.sig, ParamCard.sig, InferPipelineCard.sig, rng.seed, rng_family, and anchor.
   • canon_json(•): canonical serialization used for replayable signing.
2. Additional notation
   • Compliance confidence: conf_compliance = 1 - P( noncompliance | evidence ).
   • Evidence score: score_pkg (aggregate of metrics from Chapters 5/7/8/9).
   • Audit hash: h_k; signature: sig_k; key fingerprint: key_id.
   • Retention policy: T_retention; access policy: ACL; replay window: T_replay.

III. Postulates & Minimal Equations

• P51-18 (Pre-registration locking postulate)
  Any acceptable falsification conclusion must have pre-registered H0/H1, thresholds, and AnalysisPlan.hash, and be bound to a unique anchor and EnvLock.
• P51-19 (Replayable evidence postulate)
  Under the same EnvLock and PreReg.card, replay with the same golden_set_hash and rng.seed must yield equidistributed conclusions; deviations trigger E_NONDETERMINISM or E_ENV_MISMATCH.
• P51-20 (Tamper-proof audit chain postulate)
  The audit trail forms a time-ordered hash chain; any insertion, deletion, or modification changes the tail hash/signature and triggers E_AUDIT_CHAIN_BROKEN.
• P51-21 (Minimal sufficient evidence postulate)
  If the Evidence.bundle meets the power, error-control, and coverage gates, no additional experiments are required for a compliance conclusion.
• S52-43 (Evidence score synthesis)
  score_pkg = w_acc * score_acc + w_cov * cov_spec + w_mut * kill_rate + w_cal * ( 1 - ECE ) + w_rob * ( 1 - delta_offon ),
  where Σ w_* = 1 and component sources follow Chapters 5/7/8/9.
• S52-44 (Acceptance & rejection gates)
  Accept if:
  ( score_pkg ≥ tau_accept ) ∧ ( power ≥ power_min ) ∧ ( FDR ≤ q_star ) ∧ ( TS.error ≤ tau_error ).
  Reject if:
  ( score_pkg < tau_reject ) ∨ ( power < power_min ) ∨ ( FDR > q_star ).
• S52-45 (Compliance-confidence estimate)
  conf_compliance = 1 - P( noncompliance | Evidence.bundle ) ≈ 1 - p_value_adj,
  where p_value_adj is the multiple-testing adjusted global p-value or a posterior risk upper bound.
• S52-46 (Audit hash chain)
• h_0 = anchor ;
• h_k = hash( h_{k-1} || canon_json(event_k) ), k = 1..N

Verification: recomputed h_k must match the archived value for any prefix k.

• S52-47 (Verifiable signatures)
• payload_k = canon_json(event_k) || h_{k-1} ;
• sig_k = Sign( sk, payload_k ) ;
• Verify( pk, payload_k, sig_k ) = True ;
• key_id = hash( pk )
• S52-48 (Retention & forensic conditions)
  Compliance objects must satisfy now - ts_event ≤ T_retention, and principals in the allowed set A (per ACL) must be able to Verify and (partially) Reproduce. Audit queries should be O(log N) or better.
• S52-49 (Minimal evidence bundle E_min)
• E_min = {
• PreReg.card, TestPlan.hash, DataSpec.hash, golden_set_hash,
• CEReport, CoverageReport, AttackReport,
• power_analysis.json, multitest_correction.json,
• gate_audit.log, decisions.parquet, lr_trace.csv, ewma.csv,
• EnvLock, Graph.sig, ParamCard.sig, InferPipelineCard.sig
• }
• S52-50 (Compliance deltas & regression gating)
  delta_claim = ( score_claim - score_obs );
  if delta_claim < -tau_regress, then gate_release must return block and the AuditTrail must record the cause.

IV. Data & Manifest Conventions

• PreReg.card fields
  {project_id, anchor, H0, H1, metrics:[{name, direction, tau_accept, tau_reject}], alpha, beta, power_min, alpha_spending:scheme, AnalysisPlan.hash, sample_size_plan, randomization:{rng_family, rng.seed}, DataSpec.hash, golden_set_hash, owner, reviewer, ts_register, EnvLock}.
• Evidence.bundle layout
• evidence/metrics/*.json
• evidence/coverage/coverage_online.csv
• evidence/attacks/AttackReport
• stats/power_analysis.json
• stats/multitest_correction.json
• gate/gate_audit.log
• gate/decisions.parquet
• stream/lr_trace.csv
• stream/ewma.csv
• env/*.sig
• cards/*.card
• fingerprint
• MANIFEST.json (with the full list of hash(•))
• AuditTrail event schema
  {ts, actor, action, object_ref, digest, h_prev, h_curr, key_id, sig},
  with action ∈ {preregister, start, stop, attach, compute, sign, verify, gate_pass, gate_hold, gate_block, rollback}.

V. Algorithms & Implementation Bindings

1. Interface prototypes (added to I50-*)
   • I50-30 preregister(spec:dict) -> PreReg.card
   • I50-31 build_evidence(inputs:dict, env:dict) -> Evidence.bundle
   • I50-32 sign_and_lock(bundle:any, sk:any, EnvLock:any) -> {bundle_signed:any, Graph.sig, ParamCard.sig, InferPipelineCard.sig}
   • I50-33 verify_chain(bundle:any, pk:any) -> {ok:bool, broken_at:int}
   • I50-34 archive_store(bundle:any, policy:dict) -> {uri:str, ttl:float}
   • I50-35 forensic_replay(bundle:any, EnvLock:any, options:dict) -> ReproduceReport
   • I50-36 audit_export(bundle:any, format:str) -> bytes
2. Key exceptions (extended)
   E_SIGNATURE_MISMATCH, E_AUDIT_CHAIN_BROKEN, E_RETENTION_VIOLATED, E_POLICY_DENIED, E_REPLAY_DIVERGED, E_MANIFEST_INCOMPLETE.
3. Audit-record generation
   For every event_k, write to the AuditTrail and apply S52-46/47. I50-33 performs full or segment verification of the hash chain and signatures.

VI. Metrology Flows & Run Diagram

1. Mx-70 Pre-registration & locking
   • preregister writes PreReg.card, computes AnalysisPlan.hash, DataSpec.hash, golden_set_hash.
   • Create the initial anchor and EnvLock; start the AuditTrail.
2. Mx-71 Evidence construction
   • Execute the test matrix and adversarial flows; compile CoverageReport, AttackReport, power_analysis.json, multitest_correction.json.
   • Compute score_pkg, conf_compliance; produce MANIFEST.json and the fingerprint.
3. Mx-72 Signing & publication
   • Compute hash(•) for MANIFEST.json and all artifacts; build the hash chain h_k.
   • sign_and_lock generates *.sig bound to EnvLock; publish a read-only uri.
4. Mx-73 Archiving & forensics
   • archive_store persists the bundle with T_retention and ACL.
   • Periodically run verify_chain; on demand, run forensic_replay and compare ReproduceReport.
5. Mx-74 Gating integration
   gate_release(evidence:dict, policy:dict) -> GateDecision per S52-44/50; log to gate_audit.log and append signature.

VII. Verification & Test Matrix

1. Integrity & non-repudiation
   • Tamper scenario: mutate a row in decisions.parquet; verify_chain must return ok = False with broken_at = k*.
   • Key rotation: after key_id changes, legacy evidence remains verifiable.
2. Replayability & consistency
   • Replay under the same EnvLock: require ReproduceReport.delta ≤ tau_replay.
   • Cross-site replay (Chapter 11): delta_dev ≤ tau_equiv.
3. Statistical compliance
   • power ≥ power_min and FDR ≤ q_star; online spending satisfies Σ alpha_i ≤ alpha_total.
   • conf_compliance ≥ conf_min; on boundary sets, kill_rate ≥ tau_kill.
4. Archival & forensics
   Randomly sample p% of bundles; Verify and Reproduce must pass; retrieval latency ≤ target T_lookup.

VIII. Cross-References & Dependencies

• Closes the evidence loop with Chapter 7 (SPRT, FDR, alpha-spending), Chapter 8 (conf_compliance, ECE/NLL), and Chapter 9 (GateDecision, gate_audit.log).
• Depends on Core.DataSpec (schemas & fields), Core.Metrology (metric semantics), and Core.Threads (execution & routing).

IX. Risks, Limitations & Open Questions

• Risks & limitations
  Divergence between pre-registration and execution may go under-detected; cross-domain replay can suffer unmodeled hardware drift; long-retention raises key-rotation and crypto-aging risks; over-templating may suppress necessary exploratory analysis.
• Open questions
  Joint optimization of online FDR and archival cost; verifiable federated auditing across organizations; cross-language standards for canon_json; balancing EnvLock with privacy compliance (de-identification vs. forensic utility).

X. Deliverables & Versioning

1. Deliverables
   PreReg.card, Evidence.bundle (with MANIFEST.json and fingerprint), AuditTrail.log, *.sig, EnvLock, ReproduceReport, Compliance.summary.
2. Versioning policy
   • Evidence refresh without changing the analysis plan → patch.
   • Threshold or statistical-scheme changes → minor (re-registration required).
   • Environment or graph-structure changes → major (new EnvLock and full replay).
   • Any schema changes require updating Graph.sig and MANIFEST.schema.hash, and registration in Appendix C.