DBaD White Paper v3

Key Highlights

• Structured decision trace with local, systemic, and effective state.
• Lifecycle governance across evaluation, verification, restoration, and audit.
• Dependency and provenance reasoning instead of isolated point-in-time scoring.
• Three confirmed protocol flaws from red-team discovery.
• Runtime Enforcement Layer v1 with narrow trust-inheritance controls.
• Deterministic validation endpoint for the current trace rule set.
• Proof-backed examples linked to stored canonical traces.
• Multi-model peer review findings and a public logic-review reporting flow.
• Known limits documented instead of hidden behind confidence theater.

Why It Matters

Most evaluation systems produce a score or recommendation for a single action and stop there. DBaD is aimed at the harder problem: how decisions should remain governable when lineage, handoffs, verification, and audit all matter.

The white paper focuses on trust inheritance failures. It shows how apparently clean continuations can still fail when verifier independence breaks, actor continuity breaks, or lineage risk rises materially over time.

Need the shorter public explainer? Read DBaD Explained →

Downloads

• DBaD White Paper v3 (PDF)
• DBaD White Paper v3 (Markdown)
• Papers & artifacts index

Supporting links: methodology · API docs · demo

Current Public Runtime

• The evaluator creates and stores traces.
• Stored traces can be inspected through the public trace detail surface.
• Deterministic validation can be run against the current core rule set.
• Verification and transition actions mutate versioned traces rather than replacing them silently.
• Traces now include manually updated outcome_status and append-only outcome_history. Outcome status records what was later observed, not whether the decision was correct or safe.
• Traces now include escalation_closure and escalation_closure_history when escalations are resolved. Trust-positive continuation requires closure where escalation context applies, but closure does not prove the original decision was correct.
• The v2.2 integrity stack is now implemented in deterministic runtime form: blind spots, expected outcome, state-transition evidence, optional evidence hashing, and completeness attestation all now exist as stored trace fields or transition subfields.
• These layers record claims, references, responses, and statuses. They do not prove truth, completeness, correctness, or safety.
• A first DecencyMeter advisory demo is now live at /decencymeter/demo using audited trace trc_20260428181140_42396240. It computes a transparent score with explicit subjective weights, does not store scores, does not mutate traces, and is not DBaD validation.

What Changed After Peer Review

• DBaD’s confirmed public boundary is structure, not truth.
• The v2.2 deterministic integrity stack now covers evidence, scope, expectation, outcome, resolution, and completeness-claim recording.
• That implementation still does not make DBaD a truth engine or solve cross-trace completeness.

The public runtime-audited example is available on the DBaD v2.2 Runtime Demo page.

Confirmed Flaws

• Verifier Independence: multi-party clearance can fail if reciprocal or cartel-like approval relationships are treated as independent.
• Actor Continuity: silent handoff can look like ordinary continuation unless continuity is checked at runtime.
• Trust Trajectory: a high-risk action can borrow credibility from a clean-looking lineage path unless trajectory degradation is checked.

Runtime Enforcement Layer v1

The first runtime layer is intentionally narrow. It adds three checks at the points where trust inheritance can fail: verifier independence, actor continuity, and lineage risk trajectory.

Good history can reduce the severity of the first response, but it does not remove re-validation friction. The goal is minimal enforcement that blocks the known failure classes without forcing a full schema redesign.

The public system now also exposes deterministic trace validation, proof-backed examples, and a logic-review challenge surface without claiming that a valid trace proves correctness or safety.