The Architecture of Proof Glossary: Precise Definitions | AI Governance

Architecture

A C D E L M R S

AUC (Area Under the Curve)

Statistical Artifact Stage 1

/ˌeɪ.juːˈsiː/

A standard metric for model accuracy and separability. In the Architecture of Proof, AUC is considered a "Stage 1" metric: it provides a population-level aggregate that can mask individual-level logic failures.

Key Distinction: AUC measures performance (what happened); Causal Traces measure logic (why it happened).
Governance Action: Useful for lab validation but insufficient for Tier 4 Autonomy.

Accumulated Context Graph

Context Architecture Stage 3-4

/əˈkjuː.mjʊ.leɪ.tɪd ˈkɒn.tekst ɡrɑːf/

A structured entity-relationship topology that records and maintains conversational, situational, and factual state across agent execution boundaries without stochastic context decay.

Key Distinction: Vector embeddings retrieve similarity; an Accumulated Context Graph provides deterministic relational traversal.
Governance Action: Prevents multi-turn hallucination loops in long-horizon autonomous agent workflows.

Architectural Invariance Coefficient (AIC)

Systems Defensibility Stage 4

/ˌɑː.kɪ.tek.tʃər.əl ɪnˈveə.ri.əns kəʊ.ɪˈfɪʃ.ənt/

A formal stability metric quantifying whether end-to-end system outputs remain provably invariant across model checkpoint updates, prompt variations, and vendor migration events.

Key Distinction: Benchmark evaluations measure raw model scores; AIC measures enterprise operational repeatability independent of underlying weights.
Governance Action: Acts as a mandatory gating threshold before swapping foundation model backends in production.

Attention Dilution Tax

Token Economics Stage 2-3

/əˈten.ʃən daɪˈluː.ʃən tæks/

The non-linear decline in agent precision and reasoning fidelity that occurs when expanding context windows with unstructured tokens rather than structured, isolated chunks.

Key Distinction: Token expense scales linearly; Attention Dilution Tax scales exponentially in missed constraints and logical errors.
Governance Action: Enforces context limits and mandates chunk isolation prior to context injection.

Calibration Curve

Model Integrity Stage 2

/ˌkæl.ɪˈbreɪ.ʃən kɜːv/

A visualization that maps predicted probabilities to actual frequencies. A well-calibrated model ensures that when it says there is an 80% chance of fraud, fraud occurs exactly 80% of the time in the long run.

Key Distinction: Accuracy is about being "right"; Calibration is about being "honest" with uncertainty.
Governance Action: A prerequisite for Evidence-Based Promotion into Tier 2+ autonomy.

Causal Drift

Model Integrity Stage 4

/ˈkɔː.zəl drɪft/

The "silent killer" of autonomous systems. Causal drift occurs when the logical relationship between an input and an output shifts, even if the data distributions (covariates) remain stable.

Key Distinction: Unlike Covariate Drift (data changes), Causal Drift is a change in the rules of the world.
Governance Action: Triggers an automated [Tier Downgrade](/control-tiers).
Read More: The AI Maturity Model

Causal Trace

Audit & Forensic Stage 4

/ˈkɔː.zəl treɪs/

A "black box flight recorder" for AI logic. Unlike a standard log file that shows what happened, a Causal Trace captures the step-by-step logical path from data ingestion to final output, including which specific business rules or model layers carried the most weight.

Key Distinction: Standard logs are observational; Causal Traces are forensic.
Governance Action: Provides the 4-minute root cause diagnosis required for [Stage 4 Maturity](/maturity-score#stage-4-forensic-observability).

Control Tiers

System Autonomy Stages 1-4

/kənˈtrəʊl tɪərz/

A four-level hierarchy that dictates the operational boundary of an AI system, ranging from Tier 1 (Observe) to Tier 4 (Full Autonomy).

Tier 1 (Observe): AI monitors and logs only.
Tier 2 (Advise): AI suggests an action; human must approve.
Tier 3 (Act & Notify): AI acts autonomously but alerts human instantly.
Tier 4 (Full Autonomy): AI acts within guardrails; human reviews via periodic audit.
Governance Action: Requires Evidence-Based Promotion to move between tiers.
Read More: Control Tiers for AI-Enabled Processes

Cost per Successful Resolution (CPSR)

AI Economics Stage 3-4

/kɒst pɜː səkˈses.fʊl ˌrez.əˈluː.ʃən/

The true unit cost of an autonomous task completion, factoring initial inference, retry loops, verification checks, and downstream human escalation expenses.

Key Distinction: Cost-per-token measures vendor API fees; CPSR measures true operational business economics.
Governance Action: Replaces raw token pricing in executive ROI models and capital allocation reviews.

Deterministic State Graph

Autonomy Architecture Stage 4

/dɪˌtɜː.mɪˈnɪs.tɪk steɪt ɡrɑːf/

A finite-state execution substrate that restricts probabilistic agent actions to mathematically validated state transitions, eliminating open-ended or unverified execution paths.

Key Distinction: Prompts offer soft guidance; Deterministic State Graphs mathematically enforce allowable execution transitions.
Governance Action: Mandatory architectural foundation for Tier 4 autonomy and automated emergency circuit breakers.

Drift (Data/Model)

Model Monitoring Stage 2

/drɪft/

The decay of model performance over time as production data diverges from initial training distributions.

Key Distinction: Standard Drift is a change in covariates (data patterns); Causal Drift is a change in the underlying logic or rules of the world.
Governance Action: Triggers a mandatory performance review and potential [Tier Downgrade](/control-tiers).

Evidence Budget

Audit & Forensic Stage 3-4

/ˈev.ɪ.dəns ˈbʌdʒ.ɪt/

The dedicated allocation of compute, storage, and latency earmarked to produce replayable audit artifacts, cryptographic signatures, and causal logs for every autonomous decision.

Key Distinction: Traditional logging stores debug text after errors occur; an Evidence Budget guarantees primary system resources for defensibility.
Governance Action: Mandatory operational parameter audited quarterly during AI platform governance reviews.

Evidence-Based Promotion

AI Governance Gate

/ˈev.ɪ.dəns beɪst prəˈməʊ.ʃən/

In the Architecture of Proof framework, Evidence-Based Promotion is the formal governance bridge between an experimental pilot and high-stakes production autonomy. Rather than promoting an AI system based on static accuracy scores or arbitrary project deadlines, this protocol requires a system to earn its promotion into higher Control Tiers by demonstrating a dense, replayable audit trail.

A system is ready for promotion only when its Proof Infrastructure: including its Benford Perimeter for data integrity and its Causal Trace for logic transparency: is robust enough to handle the inevitable edge cases where probabilistic models fail. By anchoring promotion in defensibility rather than intuition, organizations transform AI from an opaque liability into an auditable business asset.

LIME

Post-Hoc Explainability Stage 1-2

/laɪm/

Local Interpretable Model-agnostic Explanations. A post-hoc technique that fits a surrogate linear model around a single prediction to provide local intuition.

Key Distinction: LIME is an approximation of local behavior; Causal Traces are deterministic records of actual logic.
Governance Action: Useful for basic transparency but insufficient for high-fidelity Evidence-Based Promotion.

Marginal Accuracy per Token (MAPT)

Token Economics Stage 3-4

/ˈmɑː.dʒɪ.nəl ˈæk.jʊ.rə.si pɜː ˈtəʊ.kən/

The differential gain in task accuracy obtained by expanding prompt context or retrieval depth relative to incremental token expense and latency overhead.

Key Distinction: Context length measures raw data volume; MAPT measures analytical leverage per dollar invested.
Governance Action: Sets empirical ceilings on prompt expansion and prioritizes GraphRAG chunk optimization.

Marginal Frontier Alpha (MFA)

Portfolio Strategy Stage 3-4

/ˈmɑː.dʒɪ.nəl ˈfrʌn.tɪə ˈæl.fə/

The measurable operational gain delivered by upgrading to a next-generation frontier model after deducting higher token tariffs, latency penalties, and migration friction.

Key Distinction: Vendor benchmarks highlight raw laboratory capabilities; MFA quantifies net realized enterprise productivity.
Governance Action: Establishes the formal gating criteria for approving foundation model version upgrades.

Replayability

Audit & Forensic Stage 2+

/ˌriː.pleɪ.əˈbɪl.ə.ti/

The technical capability to reconstruct the exact state of an AI system at the millisecond a specific decision was made. This includes the model version, precise input data (feature vector), external API responses, and the specific rules or weights that fired.

Key Distinction: Explainability is an intuition; Replayability is evidence.
Governance Action: Essential for the [AI Incident Golden Hour](/ai-incident-response-plan) and successful board-level audits.
Read More: AI Audit Trails: Replayable AI

SHAP

Post-Hoc Explainability Stage 2

/ʃæp/

Shapley Additive Explanations. A game-theoretic approach to explaining model outputs by assigning an importance value to each feature based on its contribution to the final prediction.

Key Distinction: SHAP is a post-hoc calculation that often introduces forensic drag; Causal Traces are in-situ captures available instantly.
Governance Action: Better for aggregate auditing than individual incident Replayability.

Stable Reasoning

Model Integrity Stage 3-4

/ˈsteɪ.bəl ˈriː.zən.ɪŋ/

The convergence of probabilistic model patterns into deterministic-like consistency. Stable Reasoning occurs when an AI system produces non-divergent logical paths (Causal Traces) for similar inputs, eliminating the stochastic variance that typically characterizes unmanaged LLMs.

Key Distinction: Probabilities are fluid and drifting; Stable Reasoning is a frozen logic path promoted from consistent model behavior.
Governance Action: A prerequisite for Tier 4 Autonomy and the elimination of forensic drag in root cause analysis.

How Do PMs Enforce Vocabulary Invariants Across Product, Engineering, and Risk?

When Product, Engineering, and Risk teams collaborate on AI systems, imprecise language creates hidden liabilities: engineers say "guardrail" when they mean a prompt string; risk officers hear "deterministic" and assume 100% legal certainty. Use this alignment matrix in cross-functional PRD reviews to bind terms to active engineering invariants.

Architectural Term	The Loose Casual Misconception	The Engineering Failure Mode	Deterministic Invariant Standard	Operational PRD Release Gate
Control Tiers	"We have human oversight in our product"	Rubber-stamp human review after irreversible API execution	Hard architectural bounding: Tier 0 (Suggest), Tier 1 (Sandbox), Tier 2 (Breakers), Tier 3 (Air-Gap)	System lacks execution hooks to execute Tier 2+ mutations without verified promotion
Causal Trace	"We log all model inputs and outputs in Datadog"	Observational log dump that cannot explain why the model branched	Immutable Decision State Vector capturing inputs, prompt hashes, seeds, tool outputs, and line spans	Root cause diagnosis completed within $\le 4\text{ minutes}$ during incident audits
Benford Perimeter	"Standard data validation checks"	Ingesting poisoned or synthetic tabular data into financial extraction models	Algorithmic first-digit frequency distribution tests across numeric corpora	Statistical deviation ($p < 0.01$) triggers automated data ingestion quarantine
Stable Reasoning	"The model gave the same answer twice in testing"	Stochastic divergence under real-world prompt permutations	High semantic invariance score ($AIC > 0.95$) across adversarial golden test suites	Output distribution remains invariant across model minor checkpoint updates
Evidence Budget	"Standard application observability logging"	Logging dropped to reduce cloud storage bills during production crunches	Earmarked compute, latency, and immutable storage dedicated to evidentiary proof	Every state mutation commits a signed cryptographic hash to an append-only ledger

What Survives an Adversarial Stress-Test of the Governance Glossary?

1. What is the strongest argument against this approach?

The Objection: Standardizing a rigid architectural lexicon creates semantic pedantry and cognitive friction for engineering teams accustomed to industry-standard terminology. Enforcing bespoke terminology—such as Causal Traces, Architectural Invariance Coefficients, and Control Tiers—risks alienating developers and slowing integration with mainstream open-source frameworks.

The Practical Resolution: Map every proprietary architectural definition directly to corresponding industry standards and RFCs. Correlate Causal Traces to OpenTelemetry semantic conventions, bind Control Tiers to NIST AI RMF profiles, and integrate term definitions directly into linter rules and IDE language server plugins. This provides immediate developer ergonomic support while preserving precise systems semantics.

2. Under what circumstances would this fail?

The Failure Mechanism: A shared vocabulary fails when terms degrade into superficial compliance buzzwords divorced from technical implementations. If teams label conventional prompt templates as accumulated context graphs or call standard cosine similarity stable reasoning without enforcing underlying architectural invariants, the glossary provides an illusion of rigor while operational failure risks remain unmitigated.

The Practical Resolution: Couple every glossary definition to automated architectural assertion tests. If an engineering squad claims a component implements a Deterministic Perimeter, CI/CD test suites must verify sub-millisecond execution times and deterministic drop rates on synthetic payloads. Definitions must function as compiled interface specifications rather than decorative documentation.

3. What would a skeptical customer, client, or colleague object to?

The Skeptical Pushback: Enterprise enterprise stakeholders and compliance reviewers will object that defining terms does not prevent models from failing in production. They will argue that having a precise definition of Causal Drift provides zero operational value if the system cannot automatically mitigate the drift in real time.

The Practical Resolution: Explicitly pair each vocabulary definition with mandatory operational triggers and recovery protocols. In the Architecture of Proof lexicon, every term contains an explicit Governance Action invariant: defining Causal Drift automatically binds the system to automated tier demotion and counterfactual replay within the four-minute RCA Golden Hour, transforming static terminology into active runtime safeguards.

4. Which alternative deserves more consideration?

The Counter-Architecture: Relying exclusively on existing external regulatory glossaries—such as ISO/IEC 42001, EU AI Act statutory classifications, and NIST AI RMF taxonomies—rather than defining a specialized engineering vocabulary.

The Practical Resolution: Establish a bidirectional semantic crosswalk. Use statutory and regulatory definitions for external compliance reporting and legal contracts, while using the Architecture of Proof technical vocabulary for internal runtime systems, state machine transitions, and database schema fields where statutory language lacks the bit-level granularity required for deterministic software execution.

5. What am I assuming that may not be true?

The Latent Vulnerability: We assume that engineering, legal, and executive stakeholders will maintain disciplined cross-functional fidelity to technical definitions over multi-year software lifecycles without semantic drift or organizational divergence.

The Practical Resolution: Embed architectural vocabulary into machine-readable schema repositories, protocol buffers, and automated documentation generators. By treating the glossary as an active data dictionary maintained under version control with automated PR validation, teams eliminate semantic ambiguity across cross-functional boundaries.

One-Line Synthesis

Defensible AI governance requires an unambiguous, mathematically grounded systems vocabulary that binds theoretical safety concepts directly to enforceable software invariants.