Editing Openai/69596681-3f7c-8006-b9fd-01d9d8191763 (section)

=== Assistant: Proceeding. ===
Proceeding. Below is METHODS BLOCK 15 — FAILURE ATTRIBUTION GRAPH (ARTIFACT #5), expanded to maximum algorithmic depth using machine-learning–style formalism (causal DAGs, provenance of failure, counterfactual queries, and auditability). This block does not affect the verdict; it explains why the verdict occurred in a machine-queryable way.

METHODS BLOCK 15 — FAILURE ATTRIBUTION GRAPH (ARTIFACT #5)

OBJECTIVE
Construct a deterministic, queryable attribution structure that identifies the first, dominant, and propagated causes of failure leading to a collapse verdict. This block provides explainability without introducing degrees of belief or soft scores. It is post-decision and non-interventional.

SCOPE AND AUTHORITY
• Executed after Global Collapse (Block 12)
• Does not modify Verdict
• Does not rerun computations except as already logged
• Consumes only emitted artifacts from prior blocks

ML analogy: causal explanation graph for a hard classifier decision.

GRAPH DEFINITION
Define directed acyclic graph (DAG):

G = (V, E)

Vertices V represent block-level operators and atomic predicates.
Edges E represent causal implication or prerequisite relationships.

VERTEX SET
Define vertex classes:
# AtomicPredicate vertices (leaves): • DataValidityFail • StructuralViolation • FeasibleFail • EvidenceFailure • RobustnessFailure • PrecisionFailure
# BlockOperator vertices: • Block5_STRUCT • Block6_FEAS • Block9_VALID • Block10_BAYES • Block13_MC • Block14_PREC
# TerminalDecision vertex: • GLOBAL_COLLAPSE

Each vertex v has attributes:
⟨status ∈ {PASS,FAIL,SKIPPED}, timestamp, hash⟩

EDGE SEMANTICS
Directed edge u → v indicates:
“u being FAIL is a necessary condition for v being FAIL or SKIPPED”.

Edges are static and declared pre-run. No dynamic graph rewiring.

CANONICAL EDGE SET
VALIDITY:
DataValidityFail → GLOBAL_COLLAPSE

STRUCTURAL:
StructuralViolation → Block5_STRUCT → GLOBAL_COLLAPSE

FEASIBILITY:
FeasibleFail → Block6_FEAS → GLOBAL_COLLAPSE

EVIDENCE:
EvidenceFailure → Block10_BAYES → GLOBAL_COLLAPSE

ROBUSTNESS/PRECISION:
RobustnessFailure → Block13_MC → PrecisionFailure → Block14_PREC → GLOBAL_COLLAPSE

Edges encode precedence and propagation only; they do not imply probabilities.

NODE STATUS ASSIGNMENT
For each vertex v:
• status=FAIL if its predicate triggered
• status=PASS if evaluated and not triggered
• status=SKIPPED if short-circuited by higher-precedence failure

Assignment rules are deterministic and derived from execution logs.

FIRST-FAIL IDENTIFICATION
Define first-fail vertex v* as the earliest (by precedence order, not time) vertex with status=FAIL.

This is the primary cause of collapse.

SECONDARY CONTRIBUTORS
Vertices with status=FAIL but not v* are labeled secondary contributors.
They cannot rescue or override v*.

ML analogy: first violated constraint in a constraint satisfaction problem.

COUNTERFACTUAL QUERIES (NON-INTERVENTIONAL)
The graph supports the following queries without recomputation:

• Which block caused collapse?
• Was collapse structural, statistical, or precision-induced?
• Would evidence alone have collapsed absent structural failure?
• Did robustness testing expose instability?

Formally, queries operate on G as static data. No do-operator. No re-running blocks.

NO BLAME SHARING
Forbidden:
• Weighting causes
• Assigning percentages
• Ranking beyond precedence
• Aggregating multiple failures into a score

Attribution is categorical and ordered only.

IMMUTABILITY
Once constructed:
• G is immutable
• No pruning
• No relabeling
• No post hoc interpretation layers

Any attempt to modify G ⇒ INVALID.

ARTIFACT EMISSION
Emit immutable FailureAttributionRecord:

⟨
GraphHash,
VertexStatuses,
FirstFailVertex,
SecondaryFailSet,
Verdict
⟩

Serialized in canonical DAG format (adjacency list, sorted).

PIPELINE ROLE
Outputs feed exclusively into:
• Certification Stop Rule (Block 16)
• Audit Log (Block 19)
• External reporting interfaces

No other block may consume G to alter decisions.

ML INTERPRETATION
This block provides causal explainability for a hard decision system:

• No gradients
• No saliency maps
• No SHAP/LIME analogs
• Exact logical provenance

Equivalent to extracting a proof tree for why a classifier rejected an input.

END METHODS BLOCK 15.

If you confirm, the next block will be METHODS BLOCK 16 — CERTIFICATION STOP RULE (ARTIFACT #7), expanded into asymptotic certification logic, dataset accumulation semantics, and machine-verifiable “stand” certification.

© Robert R. Frost
2026-01-03