Revisions – ML-Draft-018

Comparing Revision 00 (original) with Revision 01. Struck-through text was removed; highlighted text was added.

1. Removed

   # DP14 – Epistemic Integrity (V2)

2. Added

   # DP14 – Trust and Transparency

3. Added

   *Epistemic Integrity (V2) — the conditions under which participants can form reliable beliefs about system behavior*

4. Added

   <!-- dp-local-version: 1.0 | standardized: 2026-07-27 -->

5. ## 1. Purpose of This Draft

6. 88 unchanged paragraphs

7. Failure: - signal neglect, performative transparency

8. Added

   ## 9. Evaluation Criteria

9. Added

   DP14 is testable only if epistemic integrity is measured by **outcomes** rather than by the presence of disclosure surfaces. A system with extensive transparency UI can still score poorly if participants routinely form false beliefs from it.

10. Added

    ### 9.1 Fidelity measures

11. Added

    - **explanation–decision match rate:** share of explanations that survive counterfactual or ablation testing - **consistency rate:** similar inputs produce materially similar explanations - **evidence binding coverage:** share of high-impact claims linked to retrievable DP15 artifacts - **uncertainty accuracy:** stated confidence tracks observed correctness

12. Added

    ### 9.2 Comprehension measures

13. Added

    - **task-based comprehension:** participants can correctly state what happened, why, and how certain it is - **misinterpretation rate:** frequency of confident but incorrect inferences drawn from a surface - **overload indicators:** abandonment, non-use, or skimming of disclosure surfaces - **accessibility parity:** comprehension holds across languages, literacy levels, and assistive technologies

14. Added

    ### 9.3 Contestability measures

15. Added

    - **dispute intake latency:** time from report to acknowledgment with a tracked identifier - **correction latency:** time from valid dispute to published correction with diff - **overturn rate:** share of disputes resulting in changed signals, explanations, or classifications - **repeat-dispute rate:** recurrence of the same defect after correction

16. Added

    ### 9.4 Durability measures

17. Added

    - **signal degradation rate across hops:** loss of context when artifacts move between systems (DP7) - **history completeness:** share of signals with reconstructable version history - **drift detection time:** interval between a signal diverging from reality and detection - **spoofing detection rate:** identification of forged or inflated trust indicators

18. Added

    ### 9.5 Thresholds and gating

19. Added

    Measurement must connect to consequences, or it becomes another transparency surface.

20. Added

    - high-risk zones carry higher fidelity and evidence-binding floors - exceeding mismatch or misinterpretation thresholds triggers review (Section 6) - repeated failures downgrade confidence ratings or restrict features that depend on the affected signals - metrics have named owners, publication cadence, and methodology disclosure

21. Added

    Failure modes:

22. Added

    - **metric theater**, where favorable indicators are published while defects persist - **survivorship measurement**, where only participants who persisted are measured - **unowned metrics**, where results exist but no actor is responsible for acting on them

23. Rewritten

    ## 9.10. Non-Goals and Explicit Boundaries

24. DP14 does not require full disclosure of all internals.

25. It explicitly disallows: - explainability theater - selective disclosure that misleads - UI patterns that distort meaning - trust signals without provenance

26. Principle: > Systems may simplify, but must not mislead.

27. Rewritten

    ## 10.11. Minimum Alignment (Non-Normative)

28. A DP14-aligned system MUST: - bind explanations to verifiable evidence (DP15) - provide sufficient context to avoid misinterpretation - preserve signals across systems with degradation notices - enable contestability and correction - maintain history of signals and explanations

29. Failure modes: - deceptive legibility - unverifiable claims - silent changes over time

30. Systems lacking evidence binding, contestability, or memory SHOULD NOT be considered aligned.

31. Rewritten

    ## 11.12. Open Questions and Future Work

32. DP14 requires operational answers to questions that determine whether transparency produces **reliable understanding** rather than noise or manipulation.

33. Rewritten

    ### 11.112.1 Explanation fidelity (provable faithfulness)

34. Define measurable standards for whether an explanation is **causally linked** to the decision pathway.

35. - Methods: counterfactual tests, feature ablations, rule tracing, policy matching - Requirement: explanations MUST fail when the underlying decision changes - Open problem: standardizing faithfulness across models (symbolic, statistical, hybrid)

36. Rewritten

    ### 11.212.2 Bounded completeness (anti-misleading thresholds)

37. Determine the **minimum context set** required to avoid misleading users.

38. - Define “misleading by omission” thresholds per use case - Tiered disclosure: summary → details → raw evidence - Role-based views: participant, auditor, steward

39. Rewritten

    ### 11.312.3 Transparency vs. security (safe disclosure envelopes)

40. Formalize disclosure envelopes that prevent: - leaking exploit thresholds - enabling evasion of safeguards

41. while still exposing: - policy classes - decision categories - uncertainty and limits

42. Rewritten

    ### 11.412.4 AI explanation standards (self vs. external explanation)

43. Separate: - **system-generated explanations** (prone to self-justification) - **independent verification layers** (overlay auditors)

44. Define when external corroboration is required for high-impact decisions.

45. Rewritten

    ### 11.512.5 Cross-system preservation (loss models)

46. Specify loss models for transparency signals across systems: - what fields must persist - what degradation is acceptable - how to signal loss to users

47. Rewritten

    ### 11.612.6 Measuring epistemic reliability (outcomes, not intent)

48. Define metrics such as: - explanation consistency rate - dispute overturn rate - correction latency - signal degradation rate across hops - user comprehension accuracy (task-based)

49. Rewritten

    ### 11.712.7 Governance of transparency (who sets the bar)

50. Define: - who can raise transparency requirements in high-risk zones - how disputes over “misleading” are adjudicated - escalation from local to cross-system governance

51. Rewritten

    ## 12.13. Relationship to Other Desirable Properties (Operational Binding)

52. DP14 converts other DPs into **perceivable and actionable reality**.

53. - **DP15 (Evidence):** DP15 provides proofs; DP14 defines how proofs are surfaced, summarized, and validated by users. Missing DP14 → evidence exists but is unusable. - **DP16 (Commitments):** Roadmap claims must be presented with uncertainty, funding state, and change history. Missing DP14 → commitments appear firmer than they are. - **DP17 (Finance):** Incentive disclosures must be legible and tied to behavior. Missing DP14 → hidden extraction persists behind complex reporting. - **DP8 (Governance):** Decisions require visible rationale and contest paths. Missing DP14 → governance legitimacy degrades. - **DP12 (AI):** Model scope, limits, and policy must be visible at interaction time. Missing DP14 → AI over-claim and misinterpretation. - **DP4 (Data):** Collection, inference, and sharing must be explained at the point of impact. Missing DP14 → consent is uninformed. - **DP20 (Ownership):** Rights and surplus flows must be legible. Missing DP14 → “ownership theater”.

54. Constraint: > No DP can claim alignment if its guarantees are not **legible, bounded, and contestable** at the interface.

55. Rewritten

    ## 13.14. Foresight and Failure Design (Epistemic Incident Model)

56. Treat transparency failures as **epistemic incidents** with lifecycle management.

57. 4 unchanged paragraphs

58. ### Retrospective - root cause (policy, model, UI, incentives) - prevention changes (tests, thresholds, UI fixes)

59. Rewritten

    ### Learning loops - update conformance tests (Section 14)15) - adjust thresholds and disclosure tiers

60. Rewritten

    ## 14.15. Path Toward ML-RFC (Conformance & Testing)

61. DP14 must be testable.

62. Rewritten

    ### 14.115.1 Conformance suites - **Fidelity tests:** explanation vs. decision pathway - **Consistency tests:** similar inputs → similar explanations - **Deception tests:** selective disclosure, UI framing, AI narrative traps - **Propagation tests:** export/import with degradation signaling

63. Rewritten

    ### 14.215.2 Reference implementations - overlay panels with summary + drill-down evidence - standardized explanation cards with confidence + provenance links - dispute/appeal widgets bound to items

64. Rewritten

    ### 14.315.3 Data and artifacts - explanation schemas (fields, types, confidence) - provenance links (DP15) required for high-impact claims - change logs for explanations (versioned)

65. Rewritten

    ### 14.415.4 Governance procedures - SLA for dispute response and correction - thresholds for mandatory external verification (high risk) - zone-based escalation paths (DP8)

66. Rewritten

    ### 14.515.5 Promotion criteria - ≥ target fidelity score across scenarios - measurable reduction in misleading incidents - verified cross-system preservation with explicit degradation - functioning dispute → correction → retrospective loop

67. Rewritten

    ## 15.16. Closing Orientation (Operational Standard)

68. DP14 sets the **operational standard for belief formation** on the Meta-Layer.