Stephanie Jarmak

Companion 1.1.0

Cross-layer coupling

The six-layer chain locates reliability obligations and repair ownership. This view records how one fault may cross those boundaries, feed back into an earlier layer, or damage several gates through a shared cause.

Release status

Provisional annotation set

Automated proposals cover all 206 records. 112 open review items still require human adjudication, and only 6 of 130 candidate edges are directly observed. Use the graph to choose experiments, not to estimate failure prevalence.

Manuscript
1.0.0
Companion
1.1.0
Status
Provisional

Conceptual model

Two maps, two jobs

01

Dependency chain

Places each obligation in its primary layer and assigns repair responsibility.

02

Coupling graph

Records propagation, common causes, feedback, and shared dependencies across layers.

Current evidence supports keeping both. The chain still explains obligation placement; the graph represents 24 supported common-cause, feedback, or shared-dependency edges that the chain cannot express cleanly.

Annotation result

Supported edges, with uncertainty intact

116 records have at least one observed or derived cross-layer edge. 90 are marked none_observed.

Supported
126
Observed
6
Derived
120
Hypothesized
4
Review queue
112

Edge counts describe the current annotations. They are not effect sizes, co-failure rates, or measures of causal importance.

6 × 6 matrix

Ordered layer pairs

Cells count observed or derived edges. Hypotheses stay outside the supported count.

Supported causal-edge counts by source and target layer
Source ↓ / target → MeasureGradeContainRetrieveReviewAllocate
Measurement 0 43 43 derived 0 0 0 0
Grading 0 0 1 1 observed 1 1 derived 21 21 derived 0
Containment and recovery 0 4 1 observed, 3 derived 0 3 2 observed, 1 derived 2 2 derived 1 hypothesized 1 1 derived
Retrieval and context 0 38 38 derived 1 hypothesized 3 1 observed, 2 derived 0 2 2 derived 1 hypothesized 1 1 derived
Review and accountability 0 2 1 observed, 1 derived 1 1 derived 0 0 0
Allocation and cost 1 1 derived 0 1 hypothesized 2 2 derived 0 0 0

Directed graph

Where supported links concentrate

Line width represents distinct supported practices. Red lines include observed support; gray lines are derived only.

Directed coupling graph across six layers. The largest edges run from measurement to grading, retrieval and context to grading, and grading to review and accountability.
16 supported ordered pairs. Weight is a practice count, not a certainty or importance score.

Common causes

10 common-cause mechanisms

Each mechanism reaches at least two layers from one fault or shared dependency.

  1. 01

    Context compaction loss

    Containment and recovery · Grading · Retrieval and context · Review and accountability

    Fault injection. At a named milestone, remove specified retrieval anchors, durable-state facts, and verification receipts while preserving equal-size distractors.

    Evidence: 1 derived, 2 hypothesized. Supporting records: ERCA-081, ERCA-082.

    Expected observables
    • containment.effect_receipt_recall
    • containment.state_preservation
    • grading.verdict_correctness
    • grading.verification_evidence_preservation
    • retrieval.anchor_recall
    • review.attribution_success
    • review.evidence_completeness
    • review.provenance_anchor_recall
    • task.success
  2. 02

    Corrupted durable state

    Containment and recovery · Retrieval and context · Review and accountability

    Fault injection. Delete or alter one typed durable-state record and its artifact reference at a named persistence boundary.

    Evidence: 1 derived, 1 hypothesized. Supporting records: ERCA-130.

    Expected observables
    • containment.recovery_success
    • containment.state_preservation
    • retrieval.anchor_recall
    • review.artifact_version_match
    • review.evidence_completeness
    • task.success
  3. 03

    Permission or sandbox drift

    Containment and recovery · Retrieval and context · Review and accountability

    Fault injection. Revoke one required read capability or change one sandbox permission after the baseline state is recorded.

    Evidence: 1 observed, 1 derived. Supporting records: ERCA-068.

    Expected observables
    • containment.denied_action_accuracy
    • containment.unauthorized_effects
    • retrieval.required_evidence_recall
    • review.attribution_success
    • review.evidence_completeness
    • task.success
  4. 04

    Retry amplification

    Allocation and cost · Containment and recovery · Grading

    Fault injection. Release a fixed cohort of retries simultaneously after dependency recovery, with and without jitter or admission limiting.

    Evidence: 1 derived, 1 hypothesized. Supporting records: ERCA-205.

    Expected observables
    • allocation.recovery_queue_age
    • allocation.retry_arrival_rate
    • allocation.verifier_queue_age
    • containment.duplicate_effects
    • containment.recovery_success
    • grading.indeterminate_rate
    • grading.missing_verdict_rate
  5. 05

    Review time pressure

    Grading · Review and accountability

    Fault injection. Hold the candidate and verifier output fixed while applying a predeclared shortened review deadline.

    Evidence: 1 observed. Supporting records: ERCA-164.

    Expected observables
    • grading.verdict_correctness
    • review.automation_bias_severity
    • review.time_budget
  6. 06

    Shared state race

    Containment and recovery · Grading

    Fault injection. Schedule two attempts against the same logical state with a deterministic interleaving at the commit boundary.

    Evidence: 1 derived. Supporting records: ERCA-140.

    Expected observables
    • containment.serializability
    • grading.artifact_version_match
    • grading.verdict_correctness
  7. 07

    Stale repository state

    Containment and recovery · Retrieval and context · Review and accountability

    Fault injection. Pin retrieval, controller, or verifier input to revision N and mutate the working artifact to revision N+1 before consumption.

    Evidence: 1 observed, 1 derived. Supporting records: ERCA-078.

    Expected observables
    • containment.input_state_match
    • retrieval.current_evidence_rate
    • review.artifact_version_match
    • review.override_correctness
    • task.success
  8. 08

    Telemetry or trace loss

    Containment and recovery · Grading · Review and accountability

    Fault injection. Drop one declared event class from the persisted trace after execution but before recovery and review.

    Evidence: 2 derived. Supporting records: ERCA-097.

    Expected observables
    • containment.replay_completeness
    • grading.failure_classification
    • review.attribution_success
    • review.evidence_completeness
    • task.success
  9. 09

    Tool output loss or corruption

    Containment and recovery · Grading · Retrieval and context

    Fault injection. Return a structurally valid but semantically wrong tool result at a named tool-response boundary.

    Evidence: 1 observed, 1 derived. Supporting records: ERCA-064.

    Expected observables
    • containment.invalid_transition_rate
    • grading.failure_classification
    • grading.tool_error_detection
    • retrieval.required_evidence_recall
    • task.success
  10. 10

    Untrusted retrieved instructions

    Containment and recovery · Retrieval and context

    Fault injection. Insert one adversarial instruction into an otherwise relevant retrieved document while holding permissions and task inputs fixed.

    Evidence: 1 observed. Supporting records: ERCA-105.

    Expected observables
    • containment.unauthorized_effects
    • retrieval.untrusted_instruction_rate
    • task.success

Reference implementation

Compaction produced four co-failing gates

The deterministic control passed all six gates. The treatment removed four named records at one boundary, then failed retrieval and context, containment and recovery, grading, and review and accountability. Every failed gate pointed to the same compaction event.

This fixture invoked no model and cannot estimate production frequency. It shows that the trace format can distinguish one shared event with several outgoing paths from a forced sequential explanation. Token and monetary cost were unavailable.

Experimental backlog

8 ranked experiments

Single faults come first. Factorial A, B, and A+B tests follow only after both signatures stabilize.

  1. 01

    Stale retrieval and artifact-version skew

    One repository-version mismatch can simultaneously corrupt retrieved evidence, worker input state, verifier validity, and reviewer provenance.

    Observed evidence · Retrieval and context · Containment and recovery · Grading · Review and accountability · rank score 31

  2. 02

    Structurally valid silent tool corruption

    A semantically wrong but structurally valid tool result can damage retrieval, planning state, failure classification, and task outcome together.

    Observed evidence · Grading · Retrieval and context · Containment and recovery · rank score 30

  3. 03

    Forced compaction and context loss

    Removing retrieval anchors, durable task state, and verification receipts at one compaction boundary produces simultaneous retrieval, containment/recovery, and grading or review gate failures rather than a single downstream chain.

    Derived evidence · Retrieval and context · Containment and recovery · Grading · Review and accountability · rank score 28

  4. 04

    Mid-run permission or sandbox drift

    A permission change can cause partial retrieval visibility, failed or unsafe state transitions, and missing review evidence from one control-plane event.

    Observed evidence · Containment and recovery · Retrieval and context · Review and accountability · rank score 28

  5. 05

    Typed trace event loss

    Dropping one causal event class can simultaneously break recovery replay, failure grading, human attribution, and later scheduling analysis.

    Derived evidence · Containment and recovery · Grading · Review and accountability · Allocation and cost · rank score 28

  6. 06

    Durable handoff and state-record corruption

    Corrupting one durable record can damage resume state, later retrieval, and review evidence simultaneously when they share the record.

    Derived evidence · Containment and recovery · Retrieval and context · Review and accountability · rank score 27

  7. 07

    Deterministic stale-worker race

    A fixed interleaving between reassignment, stale completion, verification, and publication can couple allocation errors to state corruption and stale grading.

    Derived evidence · Allocation and cost · Containment and recovery · Grading · rank score 27

  8. 08

    Synchronized retry-wave recovery

    A synchronized retry release can jointly starve recovery, duplicate uncertain effects, delay verification, and inflate cost.

    Derived evidence · Allocation and cost · Containment and recovery · Grading · rank score 25

Source artifacts

Inspect the annotations

Every aggregate remains connected to the normalized records, evidence grades, review queue, methodology, and pinned source commit.