14-EFT.WP.Methods.Inference v1.0 | Chapter 1: Positioning, Scope & Inference Levels

Home ／ Docs-Technical WhitePaper ／ 14-EFT.WP.Methods.Inference v1.0

Chapter 1: Positioning, Scope & Inference Levels

I. Scope & Objectives

1. Define the object domain of Inference for this volume: centered on trained parameters theta with a fixed inference graph Graph(theta), covering offline batch inference, online streaming inference, and hybrid edge/central deployments. Training and labeling are out of scope, but inference must remain strictly aligned with the train–validation conventions.
2. Intended readers: model and platform engineers, metrology and compliance owners, operations and release owners, and third-party verifiers.
3. Deliverables & pass criteria:
   • Deliverables: InferPipelineCard, ParamCard, BenchReport, ConsistencyReport, CalibReport, ReleaseBundle.
   • Unified gate (see “Inference Levels” below): decide via gate.inf. Core metrics include accuracy, calibration, consistency, and SLO. Example composite gate:
     gate.inf: pass iff ( acc >= acc_min ) ∧ ( ECE <= tau_ece ) ∧ ( delta_offon <= tau_offon ) ∧ ( TS.latency_p95 <= tau_lat ) ∧ ( TS.error <= tau_err ).

II. Terms & Symbols

• Probability & risk: x, y, theta, y_hat, p(y | x, theta), p(theta), R_emp, R_exp, L(y, y_hat).
• Calibration & uncertainty: ECE, MCE, NLL, mean, var, quantile(q).
• Consistency & scoring: delta_offon = ( norm( y_hat_off - y_hat_on ) / norm( y_hat_off ) ), R_infer = 1 - delta_offon, score_inf.
• Runtime & observability: TS.latency, TS.thrpt, TS.error, TS.sla_violation, budget.cpu/gpu/mem/power.
• Environment & traceability: EnvLock, hash(•), fingerprint, anchor, seed.
• Conflict-name enforcement & reuse conventions: T_fil denotes Tension only; T_trans denotes transmission coefficient only; n is number density, n_eff is effective refractive index. Any expression with division or integrals must use parentheses and declare path gamma(ell) and measure d ell. Arrival-time two-form convention remains:
  T_arr = ( 1 / c_ref ) * ( ∫ n_eff d ell ) and T_arr = ( ∫ ( n_eff / c_ref ) d ell ).

III. Postulates & Minimal Equations

• P41-1 Equivalence of inference (fixed graph & locked environment)
  y_hat_off ≈ y_hat_on, with discrepancy measured by delta_offon and gated.Under EnvLock and a fixed Graph(theta), for identical input x and identical seed, offline and online inference—after time-base alignment—yield equivalent output distributions:
• P41-2 Time-base alignment
  Inference-time mapping ts = alpha + beta * tau_mono. All online metrics must be mapped to a common base before comparison.
• S42-1 Risk minimization & error decomposition
  R_emp = ( 1/N ) * Σ_{i=1..N} L( y_i, y_hat_i ); R_exp = E_{p(x,y)}[ L( y, y_hat ) ]. During inference, report R_emp and relative drift
  delta_R = ( R_emp_new - R_emp_ref ) / R_emp_ref.
• S42-2 Calibration consistency
  ECE = Σ_{b=1..B} ( n_b / N ) * | conf_b - acc_b |; declare binning method, B, and the metrology window.
• S42-3 Consistency & composite score
  score_inf = w_acc * acc + w_cal * ( 1 - ECE ) + w_cons * R_infer + w_lat * ( 1 - norm_lat ) + w_err * ( 1 - TS.error )
  Weights w_* and normalization are published with the BenchReport.R_infer = 1 - delta_offon. Example composite score:

IV. Data & Manifest Conventions

1. Minimal I/O field set
   • Input: id, ts, features, source, window = [t0, t1], hash(features).
   • Output: y_hat, uncertainty = {mean,var,quantile}, ts_out, routing, fingerprint(Graph,theta).
2. Traceability & compliance
   Each inference must persist Provenance: EnvLock, Graph(theta) fingerprint, ParamCard digest, seed, rng_family, and hash(•) linking to the data-lake entry.
3. Feature-consistency contract
   Training and inference must share the same normalization operators and pass check_dim(expr) for dimensionality. If operators change, mark a compatibility flag and migration window in ChangeLog.

V. Algorithms & Implementation Bindings

1. I40-* prototypes (anchored in this chapter)
   • I40-1 build_inference_graph(spec:dict) -> Graph
   • I40-2 load_artifacts(anchor:str) -> Runtime
   • I40-3 run_inference(rt:Runtime, inputs:any, opts:dict) -> outputs:any
   • I40-10 compare_offline_online(off:any, on:any, policy:dict) -> ConsistencyReport
2. Idempotency & exception contract
   With identical EnvLock, anchor, seed, and input set, repeated I40-3 calls must be idempotent. If nondeterministic operators exist, call with opts: { nondet_guard:true } to enable deterministic paths or report E_NONDETERMINISM.

VI. Metrology Flow & Run Diagram

• Mx-41 Build & lock
  Construct Graph(theta) → create EnvLock → persist InferPipelineCard and ParamCard.
• Mx-42 Offline evaluation
  On a benchmark set, produce BenchReport: record acc, R_emp, ECE, resource usage, and shadow TS.* metrics.
• Mx-43 Online canary
  Parallel dual-run; compute delta_offon; produce ConsistencyReport; trigger gate.inf.
• Mx-44 Release & archive
  Upon passing gates, sign and release; generate ReleaseBundle; archive fingerprints and ChangeLog entries.

VII. Verification & Test Matrix

• Minimum required cases
  Nominal samples; boundary samples (missingness, outliers, extreme distributions); latency-peak windows; resource-constrained windows.
• Consistency & calibration
  delta_offon <= tau_offon; ECE <= tau_ece; NLL <= tau_nll; declare confidence levels and metrology windows.
• SLO & availability
  TS.latency_p95 <= tau_lat; TS.error <= tau_err. Violations trigger degrade & rollback.
• Physical/spectral consistency (where applicable)
  For sequence/signal tasks, verify var( x ) ≈ ( ∫ S_xx(f) df ) and report delta_psd.

VIII. Cross-References & Dependencies

• Conventions & terminology: see Core.Terms, Core.DataSpec, Core.Metrology.
• Concurrency & observability: see Core.Threads for TS.* fields.
• Reproducibility & release: see EFT.WP.Methods.Repro, Chapters 3 (postulates), 8 (scoring), 9 (run & release), 12 (acceptance), 13 (change policy).

IX. Risks, Limitations & Open Questions

• Nondeterminism sources: GPU atomics, parallel reductions, randomized operators; replace with deterministic paths or report E_NONDETERMINISM.
• Distribution shift: online p(x) drifting from training; continuously monitor_drift with alert thresholds.
• Feature bias & convention drift: differences between offline replay and online real-time streams in windowing/imputation.
• Quantization & precision switching: fp32/fp16/int8 stability and saturation; validate separately across device tiers.
• Fairness & compliance: if human attributes are involved, report calibration and error by sensitive dimensions (metrics remain in English symbols).

X. Deliverables & Versioning

• Release manifest
  InferPipelineCard, ParamCard, BenchReport, ConsistencyReport, CalibReport, ReleaseBundle (including fingerprint(Graph,theta,EnvLock) and hash(dataset)).
• Channels & versions
  canary/stable/LTS channels with semantic versioning. Each change enters ChangeLog with delta_spec, compatibility flag, and migration window.
• Acceptance gate & archiving
  Use gate.inf as the unified pass gate. On pass, create a long-term evidentiary bundle with EnvLock, anchor, seed, TS.* summary, and a signature chain.

Inference Levels

1. L0 Sanity — Baseline connectivity
   • Goal: interfaces connect, shapes and dimensions correct.
   • Gate: TS.error == 0, R_emp computable.
2. L1 Offline Deterministic — Stable offline under EnvLock
   Gate: acc >= acc_min, ECE <= tau_ece_off, resource budgets satisfied.
3. L2 Online Parity — Offline–online equivalence
   Gate: delta_offon <= tau_offon, TS.latency_p95 <= tau_lat.
4. L3 Cross-Site/Device — Multi-site/hardware consistency
   Gate: per-site delta_offon <= tau_offon_site; after quantization switch, acc_drop <= tau_drop.
5. L4 Regulated Traceable — Auditable & replayable
   Gate: complete ReleaseBundle with signature chain; satisfy audit sampling and replay time-base alignment confidence for alpha, beta.

Chapter Anchors (for memory)

• Postulates: P41-1, P41-2.
• Minimal equations: S42-1, S42-2, S42-3.
• Metrology flows: Mx-41, Mx-42, Mx-43, Mx-44.
• Bindings: I40-1, I40-2, I40-3, I40-10.
• Key metrics: delta_offon, R_infer, ECE, NLL, TS.latency, TS.error, score_inf, gate.inf.