The dogfood loop keeps three lanes separate: NeuralWikis can explore private governed agent-memory and exchange patterns; LLMWikis publishes public builder guidance after review; UAIX remains canonical for UAI-1, AI Memory, Project Handoff, schemas, validators, and conformance language.
Portable lesson
You do not need these exact project names to use the loop. Any organization can separate private runtime evidence, public handbook guidance, and external standards authority so each claim lands in the lane that can actually support it.
Lane Responsibilities
| Lane | Owns | Must not claim | Output to next lane |
|---|---|---|---|
| NeuralWikis-style private lane | Private agent-memory experiments, scoped exchange records, evidence capture, review queues, and dogfood observations. | Public handbook authority, public write access, public MATM storage, certification, or UAIX conformance. | Sanitized promotion candidates with source identity, risk notes, and target route. |
| LlmWikis public handbook lane | How-to guidance, reference pages, templates, setup wizard prompts, starter files, examples, and support-boundary language. | Access to private traces, live Memory Curator operation, public MCP/API runtime, endorsement, or validator authority. | Public routes, starter artifacts, discovery files, and release notes. |
| UAIX canonical standards lane | UAI-1, AI Memory, Project Handoff, schema, registry, validator, and conformance vocabulary. | LlmWikis ownership or NeuralWikis runtime behavior unless explicitly documented by UAIX. | Canonical links and terminology boundaries. |
| NeuroWikis human education lane | Human-facing education and conceptual explanation where used. | Agent exchange runtime authority or LlmWikis handbook status. | Learning context only. |
Dogfood Review Loop
- Observe. A private agent-memory or handoff run exposes a repeated problem: ambiguity, duplicate memory, unsafe promotion, missing review, stale source, or tool idempotency failure.
- Record. The private lane stores a durable copy, source hash, run/task/tool IDs, source scope, sensitivity, and reviewer owner.
- Generalize. A maintainer turns the specific incident into a general rule, template, schema field, wizard question, or negative fixture.
- Gate. Promotion gates check source possession, privacy, evidence hash, lane review, public-safe rewrite, and non-claim language.
- Publish. LlmWikis updates the relevant public route, starter artifact, wizard preset, llms.txt route map, and release note.
- Feedback. Future private runs use the public guidance as the review checklist, producing new evidence only when the guidance misses a real case.
Accepted Public Outputs
Guidance pageFor durable workflows such as private-memory promotion, public-safe rewrite, or supersession.
Reference matrixFor allowed values, gate logic, claim-shape rules, and agent behavior.
Wizard presetFor repeatable planning packets that remain browser-only and proposal-oriented.
Case studyFor synthetic or redacted examples that teach a failure mode without exposing the private run.
Dogfood Anti-Patterns
- Publishing a private trace because it is useful instead of because it is safe, reviewed, and generalized.
- Letting a private runtime feature become public copy that reads like a public product claim.
- Copying UAIX words such as conformance, validator, official schema, or certification into LlmWikis pages without canonical UAIX backing.
- Flattening NeuralWikis, NeuroWikis, LLMWikis, AIWikis, and UAIX into one authority lane.