starshard-ai Adversarial Reviewer Panel
Purpose
Turn "ask reviewers" into a constraint-driven review loop, not a chorus of interesting opinions.
Use reviewers to find what a theory, design, or decision forbids, predicts, breaks, or operationally changes.
Default Execution Model
The valuable unit is the reviewer persona prompt bound to a constraint, not a specific model provider. Use any strong models (GPT / Claude / Gemini / local models) as interchangeable executors. If a given reviewer bridge is unstable, do not block the mainline; create a local copy-pasteable reviewer packet and run a deterministic synthesis, or queue the packet for any available strong executor.
Default order:
- Design 3-5 persona reviewers from constraint classes.
- Write each persona as an operational prompt: what it forbids, what it tests, what concrete action it would change.
- Launch model reviewers only when the bridge is healthy or the risk class justifies waiting.
- If bridge launch fails or is slow, record
Review sanity-check: unavailable/pending, keep reversible implementation moving, and preserve the persona packet as the review artifact.
Default Panel
Pick 3-5 reviewers by constraint class, not by vibe:
- Physics / measurement / falsifiability: overclaim check; operational tests.
- Practice / ethics / relationships: concrete behavior, role boundaries, ritual, trust.
- Public reality / social legitimacy: how others can reject, misunderstand, exit, or audit.
- Control systems / safety: feedback loops, brakes, error amplification, rollback.
- HCI / product: what the user actually perceives, what reduces friction.
- Embodied / contemplative / apophatic: for consciousness, interface, spirituality-adjacent, or body-insight questions, check whether the agent is mistaking language, logic, or model-compressed concepts for direct realization or lived reality.
- Physics / metaphysics bridge: when a topic compares physics with metaphysics, do not use "metaphysics" as a dismissal. Treat them as potentially different epistemic interfaces: third-person measurement/reproducibility vs. first-person embodied realization. The task is to map correspondences and boundaries without borrowing authority across domains.
Named personas are encouraged as mnemonic handles, but each must carry a constraint. Examples: Planck = measurement/falsifiability, Confucius = lived conduct/ritual, Arendt = plurality/public reality, Dijkstra = complexity and invariant discipline, Shannon = channel/noise/coding limits, Grace Hopper = operational debuggability and user tooling.
Reviewer Prompt Contract
Every reviewer receives a compact packet with:
- decision question or theory claim,
- source facts/artifact paths,
- the constraint class they own,
- explicit instruction to avoid generic praise.
Require this output:
- Strongest insight.
- Main overclaim / self-delusion / failure mode.
- One concrete design or action implication.
- One disconfirming observation or anti-signal.
- How the reviewer process itself should be improved.
Integration Contract
The synthesizing agent must synthesize, not paste reviewer chatter. Surface only:
- the shared conclusion,
- material disagreements,
- action/design implication,
- falsifiable test or anti-signal,
- next owner gate if any.
If reviewers agree too easily, add a hostile constraint reviewer or ask:
What would make this beautiful theory false, harmful, or useless tomorrow?
For embodied/contemplative topics, also ask:
What is being lost because this review is made of language? Where might the
model be confusing a sayable map for the lived territory?
For physics/metaphysics bridge topics, also ask:
Which claim belongs to third-person measurement, which belongs to first-person
transformation, and what would make them two projections of one underlying
structure rather than a category error?
Meta-Optimization Loop
After each important review, update the panel recipe:
- Which reviewer caught a real risk?
- Which reviewer produced decorative language?
- Which missing constraint would have changed the action?
- Did the review produce an implementation constraint, not just ontology?
- Did it reduce effort or merely add discourse?
Promote durable reviewer lessons to your persistent memory when they affect future routing.
Hard Boundaries
- Do not outsource the critical path to reviewers when deterministic low-risk implementation can continue.
- Do not let reviewers replace owner gates for money, identity, public commitment, privacy, account auth, or irreversible actions.
- Do not present reviewer personas as authorities; present their constraints.
provenance: maker: Starshard homepage: https://github.com/starshard-ai source: https://github.com/starshard-ai/reviewer-wheels license: MIT version: 0.1.0 contact: https://github.com/starshard-ai/reviewer-wheels/issues
About the maker: Starshard builds open agent skills; you can find the source and report issues at https://github.com/starshard-ai. License: MIT.