14-EFT.WP.Methods.Inference v1.0 | Chapter 2: Terms, Boundaries & the Object Model

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

Chapter 2: Terms, Boundaries & the Object Model

I. Scope & Objectives

1. Define a unified glossary and symbol set for the Inference domain. Establish a five-layer object model with boundary contracts across Env / Data / Model / Runtime / Metrics, so teams, sites, and devices communicate without ambiguity.
2. Target deliverables: the terminology list, conflict-name and convention constraints, the object model with interface boundaries, cross-volume field mappings, a minimal exception set, and metrology anchors. Keep alignment with Chapter 1’s gate.inf and Chapter 3’s P41-*, S42-*.
3. Pass criteria:
   • All interface inputs and outputs are expressed in English and plain text. Inline quantities are wrapped in backticks.
   • Any expression containing division or an integral uses parentheses and declares the path gamma(ell) and the measure d ell.
   • Before any cross-environment or cross-device comparison, map metrics to the common time base ts = alpha + beta * tau_mono.

II. Terms & Symbols

1. Basics and probabilities
   • x (input), y (target), theta (model parameters), y_hat (prediction), p(y | x, theta) (likelihood), p(theta | D) (posterior).
   • Loss and risk: L(y, y_hat), R_emp = ( 1/N ) * Σ L( y_i, y_hat_i ), R_exp = E_{p(x,y)}[ L( y, y_hat ) ].
2. Time bases and windows
   tau_mono (monotonic internal time), ts (published time), alpha, beta (alignment coefficients), window = [t0, t1], lookahead = 0 (online constraint).
3. Environment and traceability
   EnvLock (environment lock fingerprint), hash(•) (content hash), fingerprint (artifact fingerprint), anchor (load anchor).
4. Consistency and calibration
   delta_offon = ( norm( y_hat_off - y_hat_on ) / norm( y_hat_off ) ), R_infer = 1 - delta_offon, ECE, MCE, NLL.
5. Observability and resources
   TS.latency, TS.thrpt, TS.error, budget.cpu/gpu/mem, budget.power.
6. Conflict-name enforcement (cross-volume)
   • T_fil denotes Tension only. T_trans denotes transmission coefficient only. n is number density, n_eff is effective refractive index. Mixing is prohibited.
   • Arrival-time two-form convention is inherited:
     T_arr = ( 1 / c_ref ) * ( ∫ n_eff d ell ) and T_arr = ( ∫ ( n_eff / c_ref ) d ell ).

III. Object Model Overview

1. Five layers and their main channels
   • Env layer: EnvLock supplies a replayable, verifiable execution environment for the inference graph and runtime.
   • Data layer: Dataset/Stream unifies windows and conventions, then emits standardized features features(x).
   • Model layer: Graph(theta) with operators that prefer determinism and a declared quantization policy.
   • Runtime layer: Runtime covering load orchestration, concurrency and batching, device and precision management.
   • Metrics layer: MetricsBus collects accuracy/ECE/NLL and TS.*, then emits BenchReport and ConsistencyReport.
2. Key edges and functions (bound to I40-*)
   • build_inference_graph (I40-1), load_artifacts (I40-2), run_inference (I40-3), compare_offline_online (I40-10).
   • Calibration and drift: calibrate (I40-5), monitor_drift (I40-7).

IV. Env Layer: Environment Domain & Locking

1. Minimal composition fields
   • OS/Kernel/LibC/Compiler/BLAS/FFT, drivers and firmware, cuda/cudnn or equivalent acceleration stacks, device{model,sm,driver}.
   • Floating-point environment and FPU flags, quantization qformat (e.g., fp16, int8), random sources rng_family and seed.
   • Time zone and time-base coefficients {alpha, beta}, threading and concurrency policy TS.policy.
2. Contracts and exceptions
   • Under the same EnvLock, I40-3 is idempotent for identical {anchor, x, seed}. Drift raises E_ENV_DRIFT.
   • Any cross-site comparison must first run align_timebase, otherwise raise E_TIMEBASE_SKEW.

V. Data Layer: Inputs, Features & Windows

1. Minimal manifest and conventions
   • Input record: {id, ts, source, window=[t0,t1], features, hash(features)}.
   • Feature engineering must pass check_dim(expr). Standardization and missing-value strategies are frozen in ParamCard.
2. Windowing and streaming constraints
   Online lookahead = 0. Offline replay must replicate the production window and the same alignment function ts = alpha + beta * tau_mono.
3. Traceability and de-identification
   Data-lake entries are linked via hash(•) and fingerprint. De-identified fields are explicitly marked in DataSpec.

VI. Model Layer: Operators & Graph

• Model definition
  Graph(theta) includes the operator set, topology, precision policy, and dynamic-axis specification. opset_version and the device-availability matrix are pre-release gates.
• Numerics and determinism
  Constrain overflow/underflow and condition numbers. Any nondeterministic operator must provide a masked path or deterministic replacement. If masking is infeasible, runs must label E_NONDETERMINISM and fail gate.inf.
• Quantization and compatibility
  Precision switching fp32/fp16/int8 must ship with an accuracy_drop assessment and a tolerance window. Cross-device equivalence testing is required.

VII. Runtime Layer: Execution & Concurrency

• Composition
  Runtime = {scheduler, batcher, device_pool, memory_manager, logger, metrics_sink}.
• Concurrency and observability
  Threading semantics and TS.* follow Core.Threads. Record TS.latency_p50/p95/p99, TS.thrpt, TS.error, TS.sla_violation.
• Idempotency and side effects
  For the same {EnvLock, anchor, seed, x}, run_inference has no harmful side effects. If caching is used, the cache key includes that quadruple.

VIII. Metrics Layer: Measurement & Alerts

1. Offline evaluation
   Emit BenchReport: acc, R_emp, ECE, NLL, resource usage, and shadow TS.*.
2. Online parity and drift
   • Emit ConsistencyReport: delta_offon, R_infer = 1 - delta_offon, with the trigger delta_offon <= tau_offon.
   • Drift monitoring uses distribution comparison or spectral conventions. For sequence tasks, verify
     var( x ) ≈ ( ∫ S_xx(f) df ) and publish delta_psd with window U_w/ENBW.
3. Alert matrix
   Calibration degradation ECE > tau_ece. Latency breach TS.latency_p95 > tau_lat. Error-rate breach TS.error > tau_err. Each maps to a degrade or rollback path.

IX. Boundaries & Interface Contracts

1. Boundary definitions
   • Env ↔ Model: I40-1/2 read EnvLock only, runtime mutation is prohibited.
   • Data ↔ Runtime: inputs must carry {ts, window}. Runtime is responsible for align_timebase.
   • Model ↔ Runtime: device and precision are decided by Runtime, without violating determinism constraints of Graph(theta).
   • Runtime ↔ Metrics: metric collection is asynchronous, yet consistency reports align with the same inference batches.
2. Minimal parameters for key interfaces
   • I40-1 build_inference_graph(spec:dict): must include opset_version, precision_policy.
   • I40-3 run_inference(rt, inputs, opts): inputs include {features, ts, window}; opts may include {seed, nondet_guard:true, batch:int}.
   • I40-10 compare_offline_online(off, on, policy): policy includes match_key, window_tolerance, thresholds.
3. Exceptions (minimal set)
   E_MODEL_MISMATCH, E_SCHEMA_MISMATCH, E_TIMEBASE_SKEW, E_NONDETERMINISM, E_CALIBRATION_FAIL, E_DRIFT_DETECTED, E_RESOURCE_EXCEEDED, E_PRECISION_LOSS.

X. Cross-Volume Field Mapping & Convention Inheritance

• With EFT.WP.Methods.Repro
  Inherit EnvLock, hash(•), fingerprint, TS.*, release channels canary/stable/LTS, acceptance and archival flows. Consistency delta_offon is the inference-domain counterpart of delta_rep.
• With Core. volumes
  Core.DataSpec provides DataRecord and FeatureSpec. Core.Threads provides TS.* fields and concurrency semantics. Core.Metrology provides S_xx(f), U_w, ENBW, and spectral-consistency conventions.
• Unified numbering
  This volume uses postulates P41-*, minimal equations S42-*, metrology flows Mx-4*, implementation bindings I40-*. Cross-volume citations follow: “see the companion white paper Energy Filament, Chapter x, S/P/M/I…”.

XI. Risks, Limitations & Open Questions

• Nondeterminism sources
  GPU atomic ordering, parallel reductions, data races. Replace with deterministic paths or raise E_NONDETERMINISM.
• Dynamic shapes and batch scheduling
  Adaptive batching can alter numeric paths; verify effects on acc/ECE within the BenchReport.
• Feature drift and convention changes
  If training vs. inference standardization and missing-value policies diverge, expect E_SCHEMA_MISMATCH and abnormal delta_offon.
• Quantization switches
  fp32 → int8 tolerance is task-dependent; recommend stand-alone equivalence checks across device tiers.

XII. Numbering Anchors (for memory)

• Postulates: P41-1 (equivalence of inference), P41-2 (time-base alignment).
• Minimal equations: S42-1 (risk minimization), S42-2 (calibration consistency).
• Metrology flows: Mx-41 build-and-lock, Mx-42 offline evaluation, Mx-43 online canary, Mx-44 release and archive.
• Implementation bindings: I40-1, I40-2, I40-3, I40-5, I40-7, I40-10.
• Key metrics and gates: delta_offon, R_infer, ECE, NLL, TS.latency, TS.error, gate.inf.