DevRel Operating Model

Updated 2026-07-07

The operating model is the set of roles, steps, and schedules that keep this wiki accurate. Use this page when you need to decide who owns a piece of incoming work, how to check it, and when it is ready to publish.

Incoming work is a page, a recipe, a benchmark, a source update, or an ecosystem entry. Every kind of work follows the same path here: it gets an owner, a reviewer, a source check, and an acceptance test before it goes live.

Know the roles

Each role below owns one kind of decision. One person can hold more than one role, but every published page needs a named page owner and a named reviewer. For example, a new llama.cpp recipe has a recipe author who writes it, a runtime validator who repeats the setup, and a reviewer who signs off before the maintainer publishes it.

RoleOwnsMain outputs
DevRel maintainerWiki roadmap, page set, publication decisionsAccepted backlog, published pages, contributor guide
Docs librarianSource quality, claim tracking, link hygieneClaim/source matrix, source status, quarantined claims list
ResearcherDiscovery work and source collectionResearch notes, official source fallbacks, open questions
Recipe authorLocal runtime and integration guidesRunnable recipes with acceptance tests
Benchmark ownerTest schema, task suite, result qualityBenchmark template, recorded runs, regression notes
Page ownerOne page or one ecosystem entryDrafts, refresh dates, issue triage for that page
Layer ownerOne supply chain layerRouting decisions, entry consistency, open gaps
Runtime validatorRuntime checks for local recipesRepeated setup results, device notes, failed paths
ReviewerPractical check before publicationSource review, runnable check, clarity review
Ecosystem liaisonUpstream wording and validationPrismML or project confirmations, roadmap links, approved wording

A claim is quarantined when we remove it from reader pages because it has no primary source. It stays on the quarantined list until someone finds an official source for it or retires it.

Take in new work

Every new piece of work goes through the same eight steps before anyone starts writing:

  1. Capture the request in the backlog or issue tracker.
  2. Assign the work to one supply chain layer from the ecosystem map.
  3. Assign the work to one public wiki page.
  4. Classify the source state as primary, discovery only, mixed, or unsourced.
  5. Name the page owner, reviewer, and source owner.
  6. Define the output as a doc, recipe, benchmark artifact, or source update.
  7. Define the acceptance test before work starts.
  8. Reject or quarantine any product claim that has no primary source.

Record the answers on an intake card. A filled card for a recipe request looks like "Title: Bonsai with Ollama. Work type: recipe. Acceptance test: a reviewer can pull the model and get one completion." Copy this template:

### Intake card

- Title:
- Work type: page | recipe | benchmark | source update | ecosystem entry
- Target page:
- Supply chain layer:
- Contributor:
- Page owner:
- Reviewer:
- Source owner:
- Source state: primary | discovery only | mixed | unsourced
- Expected output:
- Acceptance test:
- Open questions:
- Publish blocker:

Route requests to the right place

Match the request to its topic and send it to the page that owns that topic. Each route also names the proof the work needs before it can publish.

If the request is aboutRoute it toRequired proof
PrismML model claims, file size, license, device support, or qualityThe PrismML claim/source matrixOfficial PrismML post, repo, model card, paper, or app page
llama.cpp, MLX, Ollama, LM Studio, WebGPU, or iPhone pathsRuntime recipe backlog, e.g. Bonsai with llama.cppA command or documented manual path that a reviewer can repeat
Quality, latency, memory, energy, or offline behaviorBenchmark harness pagesBenchmark schema and recorded run details
Architecture, KV cache, kernels, or training methodsExplainers under Technical GuidesPrimary docs, papers, or official project docs
X/Twitter signal or third-party amplificationResearch notes firstDurable official source before publication. Claims from X stay labeled "discovery, needs primary source"

Move pages through the lifecycle

A page moves through nine states from proposal to publication. Each state has one exit rule, and the person named in the rule decides when the page moves forward. For example, a draft recipe cannot reach Runnable checked until someone repeats its setup on a real machine.

StateExit rule
ProposedIntake card has owner, reviewer, source state, output, and acceptance test
AcceptedMaintainer confirms the page belongs in the wiki architecture
DraftPage owner writes the first version and links all sources
Source checkedDocs librarian verifies claims and marks gaps
Runnable checkedRecipe author or reviewer repeats the setup, prompt, benchmark, or manual flow
ReviewedReviewer checks clarity, audience fit, source status, and acceptance test
PublishedDevRel maintainer approves and assigns refresh date
Refresh duePage owner checks links, source status, and current runtime behavior
Retired or quarantinedMaintainer removes stale or unsupported public claims

Recipes and the Runnable checked gate

A recipe that no one has run stays labeled "not verified" and must ship with an acceptance test a reviewer can repeat, e.g. "A reviewer can start a local OpenAI-compatible server and run one prompt." It cannot pass Runnable checked on the author's word.

Assign owners to pages

Every page area has a default owner, a required reviewer, and an event that triggers a refresh. When you take on a page, check this table to see who reviews your work and what event means the page needs another look.

PagePrimary ownerRequired reviewerRefresh trigger
OrientationDevRel maintainerDocs librarianThesis or map changes
Small-model state of the artResearcherDocs librarianNew model family, benchmark, or architecture source
Runtime and KV cacheRuntime page ownerRecipe author or benchmark ownerRuntime release or recipe change
Training and alignmentTraining page ownerBenchmark ownerNew fine-tuning or alignment path
PrismML integration guidePrismML liaisonDocs librarianPrismML release, model card change, or app update
Benchmark harnessBenchmark ownerDevRel maintainerNew metric, task suite, or published result
Contribution guideDevRel maintainerExternal contributor reviewerContributor friction or workflow change
Runtime recipesRecipe authorRuntime validatorRuntime, model file, device, or command change
Ecosystem entriesPage ownerLayer ownerSource change or owner change

Every maintained page records its ownership in the page frontmatter or the issue tracker. Record the page owner, reviewer, source owner, supply chain layer, last source check, last runnable check, next refresh date, and current status. Readers never see this record, so keep it out of the page body.

Keep pages current

Run these checks on a schedule so pages do not go stale between releases.

CadenceTask
WeeklyTriage new intake cards and assign owners
WeeklyReview source hygiene blockers and quarantined claims
Every release or model updateRefresh PrismML pages and runtime recipes
MonthlyCheck priority backlog items and retire stale work
Before publicationComplete the acceptance test and ownership record

Write and review to the standard

Apply these rules before writing or reviewing any page:

  1. Name the reader by role, not by vague interest.
  2. State the reader's task, e.g. run, compare, or verify.
  3. Classify the topic under exactly one supply chain layer in the frontmatter, and keep that classification out of the reader prose.
  4. Convert social or search discoveries into source tasks before publishing claims.
  5. Prefer runnable proof and benchmark logs over broad assertions.
  6. State uncertainty in plain reader language, e.g. "These commands come from the upstream docs; we have not yet run them on hardware."
  7. End every page with a next step the reader can click.

Next steps