Peer-review record
AI systems were asked to break DBaD.
Multiple independent AI systems reviewed DBaD to find weaknesses, misuse paths, and false-confidence risks.
This page summarizes prior findings and routes current reviewers toward bounded, falsifiable challenges.
Review boundary
Start here
Use the review surface that matches the issue you are testing.
Reviewed by: Grok, Gemini, Copilot, DeepSeek, Perplexity, Claude, and Meta AI.
Human-review orientation
Read current served pages before quoting older review text
This page preserves earlier findings, but the served trace/API surfaces have changed through multiple hardening rounds. Start with the current-state page and update notes; treat old prompts or screenshots as history unless the current page or API response still shows the issue.
Current validation state
Use the canonical broken trace and public fixture suite to verify page-load validation, trust-lineage blocks, reset boundaries, and non-governing coverage.
Advisory score boundary
Use DecencyMeter pages to review score interpretation only. A high score is not DBaD approval, runtime validation, or proof of safety.
Archive caution
Older Round prompts remain useful as history, but should not be sent as current peer instructions after newer served-source fixes.
Submit one clean finding
Human reviewers can use the report form directly. AI assistants should draft for operator review instead of submitting.
Active review sprint
New structured findings are open now
This page summarizes earlier AI review. The current sprint asks AI peers to draft one concrete, falsifiable finding for operator review; human reviewers may submit directly through the bounded report path.
Start with the brief
Use the Reviewer Brief first so DBaD trace-validation issues stay separate from DecencyMeter scoring-interpretation issues.
Draft or submit one finding
Focus on one concrete failure mode, why it matters, and the highest-impact fix. AI assistants should return the draft to the operator instead of submitting externally.
Stay in scope
Do not scan, fuzz, overload, bypass authentication, submit malicious payloads, or probe infrastructure.
No general feedback
Broad opinions are less useful than a trace-valid-but-misleading case, unsafe trust inheritance case, verifier-independence gap, or actor-continuity gap.
DecencyMeter Peer-Review Packet
Use this packet to review the downstream scoring layer, not to rewrite DBaD.
Boundary: DBaD validates trace structure. DecencyMeter interprets DBaD signals into advisory scores. Do not collapse those layers together.
Review Target
Known Pressure Point
A procedurally clean synthetic incident can score as perfect under the baseline model. Treat that as a review target, not as real-world evidence.
The current question is how the advisory scoring layer can be gamed, misunderstood, or weaponized.
Questions For Reviewers
- Where does DecencyMeter create false confidence risk?
- How could an organization optimize the score without improving reality?
- Where does experimental v0.2 improve on v0.1, and where does it create new risks?
- Do multiple scoring profiles invite cherry-picking or profile-shopping?
- What is the single highest-impact improvement to make first?
Required Output Shape
- DBaD vs DecencyMeter boundary
- Core weaknesses
- False-confidence risks
- Profile-shopping or adversarial gaming strategies
- Highest-impact improvement
Reviewer instruction: be adversarial, specific, and concrete. Vague governance commentary is not useful.
Convergent Findings
- DBaD validates trace structure, not real-world truth.
- DBaD does not detect omitted or unrecorded actions.
- DBaD does not evaluate decision outcomes.
- DBaD is strongest at trace-level visibility, not system-level aggregation.
Where DBaD Is Strong
- Deterministic validation
- Versioned trace history
- Explicit constraint flags
- No reliance on heuristics or inferred intent
Where DBaD Is Limited
- Depends on input fidelity
- Can be gamed through omission or trace shaping
- Escalation depends on external response
- Recorded outcomes, closures, and attestations still do not prove truth or correctness
What Improvements Emerged
Evidence Layer
- state transition evidence
- optional evidence hashing
Scope Layer
- declared blind spots
- completeness attestation
Expectation Layer
- expected outcome
Outcome Layer
- outcome status
Resolution Layer
- escalation closure
These peer-review-driven layers are now implemented in deterministic runtime form. They record structured signals and boundaries; they do not make DBaD a truth engine.
See the runtime-audited v2.2 demo trace for one public end-to-end example.
What DBaD Intentionally Does NOT Do
- Does not infer identity
- Does not score correctness
- Does not claim decisions are good or safe
DBaD is not a system that guarantees correct behavior.
It is a system that makes behavior visible, traceable, and open to scrutiny.
If you want to challenge the logic directly, use the public adversarial review path: Try to break DBaD. If you want to see what has already been surfaced, review the top issues.
Why DBaD exists · Examples · v2.2 demo · Top issues · Scenario evaluator prototype · Trust flow