Assembly Map
Collate the open-source AI ecosystem into one evidence-gated lab.
The product is not a pile of tools. It is one artifact graph where learning, research rooms, open models, data, evals, discussions, funding, and AI agents all leave inspectable records.
Assembly spineOne artifact graph, many open-source layers
Learning, research, agents, releases, and funding all write into the same evidence trail.
Unifying rule
The artifact graph is the platform spine.
Every claim has a source, artifact, reviewer, status, and caveat.
Every model release says whether it is open source, open weight, open recipe, private, or embargoed.
Every discussion, agent run, evaluation, bounty, and experiment links back to a content object.
Every public promise is either live, planned, blocked, or explicitly out of scope.
First walking skeleton
Attention to Serving now has a candidate eval artifact.
This is the shape to repeat: a room route, a structured eval card, a deterministic local witness, a result artifact, caveats, openness boundary, model/data/eval reference card, and review-ready learning, RFC, discussion, decision, credit, funding, and model-program trails.
kv_cache_gib · 3 source anchors
Open roomMHA 64GiB · GQA 16GiB · MQA 2GiB
scripts/lab-os/attention-serving-kv-eval-card.mjs6 stages · predict, witness, caveat, contribution · learner test pending.
5 blocked learning claims10 stages · 12 artifacts · first room attention serving · GPT Pro unavailable in current codex session.
6 promotion gatesTransformer Lab · 8 stations · 3 live local stages · 6 room artifacts · 2 missing room artifacts · GPT Pro unavailable in current codex session.
8 blocked overclaimsTransformer Lab · 8 stations · 8 claims · 11 source anchors · AI review unavailable.
5 blocked source overclaims7 stages · 6 learner outcomes · learner test pending.
5 blocked learning claimsTransformer Lab · 5 packets · 5 tracks · signup not open.
5 blocked review claims7 rules · discussions candidate · chat not installed.
6 blocked venue claimsTransformer Lab · 7 categories · public credit closed · ledger local artifact only.
6 blocked credit claimskv-memory · gqa-changes-h-kv · supported as local memory equation only.
human pending · GPT Pro not performed in current sessionTransformer Lab · 2 model refs · 2 data refs · 3 eval refs.
6 blocked model/data claimsDataset not selected · card missing · Croissant not created.
9 blocked data claimsExecution not run · metrics missing · samples missing.
9 blocked run claims5 questions · 4 reviewer kinds · GPT Pro not performed in current session.
6 blocked review claims7 stages · 6 runtime boundaries · GPT Pro unavailable in current codex session.
10 blocked live claims75 artifacts · 244 relationships · export index only no export.
7 blocked package claims59 candidates · export locally generated · publication not published · release blocked.
12 metadata mappingsTarget 10 reviewers · signup not open · cohort not started.
6 review packets6 acceptance gates before promotion.
.github/ISSUE_TEMPLATE/lab-os-eval-card.yml29/29 checks passed; GPT Pro and maintainer review remain explicit gates.
prompts/lab-os/attention-serving-eval-card-review.mdsource-grounded · local-witness · candidate-eval; model release no; dataset release no.
unavailable GPT Pro review1 model ref · 1 dataset ref · 2 eval refs.
4 blocked overclaimsDataset not selected · card missing · Croissant not created.
8 blocked data claims6 stages · model not started · dataset not started · release no model or dataset release.
8 blocked release claims8 precedents · 7 lanes · 2 room applications.
6 blocked program claimsExecution not run · metrics missing · samples missing.
8 blocked run claims5 expected artifacts · 6 acceptance gates · agent proposal only.
not opened issuePublic discussions candidate · chat not installed · 6 routing rules.
agent safety gated4 blocked claims; human review pending.
accepted6 credit categories · authorship blocked · merge blocked.
local artifact onlyPublic ask closed · bounty closed · compute grant closed.
$0 cash budgetPublic ask closed · fiscal host not selected · sponsor tiers closed.
6 blocked funding claims8 routes compared · public ask closed · compute credit not submitted.
10 source anchors8 lanes · GPT Pro not connected · outreach not started.
8 narrow asksProduct layers
Every layer writes into the same evidence trail.
The Claude session made the important correction: the graph is not another feature. It is the contract that lets parallel work combine instead of drift apart.
Learning Atlas
Concept notebooks, source panels, demos, prediction checkpoints, and code witnesses give learners the object before asking them to contribute.
Writes: concept, equation, source, demo, code witnessResearch Rooms
Each room has one concrete research object, a ladder of readiness, and a boundary between what is proven, toy evidence, and planned work.
Writes: room, claim, ladder stage, discussion anchorRFC -> Issue Pipeline
Ideas become scoped hypotheses with prior work, data, compute, risks, expected artifacts, and acceptance gates before anyone asks an agent or contributor to work.
Writes: rfc, issue, task, acceptance criteriaExperiment Runner
Start with deterministic local scripts and logged CPU runs; promote to GPU jobs only when the claim, data, eval, and budget are ready.
Writes: run, config, log, output, limitationOpen Model Foundry
Small-model finetuning, distillation, post-training, model cards, dataset cards, eval results, and release packaging live behind the same evidence gates.
Writes: model, dataset, recipe, card, checkpointOpen Review Layer
Human reviewers and AI reviewers inspect source support, license boundaries, benchmark contamination, safety, and reproducibility before promotion.
Writes: review, caveat, decision, reviewer creditPublic Funding Ledger
Bounties, compute credits, grants, and sponsorships map to visible work packages instead of a vague donation pool.
Writes: bounty, sponsor, expense, compute grantMission Control
The platform tracks its own roadmap, agent runs, costs, PRs, decisions, coverage, and blocked work as a private room that can later become public read-only.
Writes: agent run, PR, cost, deploy, roadmap statusOpen-source assembly
Use existing ecosystems as components, not as vague inspiration.
Continuous Function should orchestrate the best public building blocks through a versioned catalog: 10 layers, 80 official source anchors, and an explicit GPT Pro review gate marked unavailable in current codex session.
Openness Standard
Every release, model card, eval card, and room badge reads from this boundary.
Open Model Foundry
Research rooms can point to model families without copying artifacts or hiding license caveats.
Dataset And Lineage Layer
A learner can trace a claim from a concept page to the dataset or data statement that supports or limits it.
Evaluation Harness
Every research room can graduate from toy witness to candidate eval to externally comparable benchmark.
Run Artifact Store
A model or eval claim can point to the exact run artifact instead of relying on screenshots or chat summaries.
Research Workflow And Review Layer
A public contributor sees the exact path from question to evidence to implementation without needing private chat context.
Learning Path Layer
Continuous Function stays first a place to learn, and only then asks learners to become reviewers or builders.
Contribution Ladder
Contribution begins as review and source correction, then grows toward scoped PRs, eval maintenance, and room stewardship.
Funding And Governance Ledger
Funding becomes a graph object: what was funded, by whom if public, what it produced, and what remains unfunded.
AI-Assisted Review And Agent Runs
Agent output becomes a reviewable lab record instead of invisible magic.
Artifact index
Six questions make the evidence graph navigable.
The index is now a public-safe contract with 75 entries, 244 relationships, and 7 standards anchors. A local JSON package candidate is locally generated with 59 checksummed artifacts; publication remains not published. The metadata crosswalk has 12 rows for future RO-Crate, BagIt, SPDX, citation, CodeMeta, and DataCite outputs, with no rich export generated.
Can I understand the room, make a prediction, inspect a source, and see the caveat?
21 linked artifacts · W3C PROV-O provenance · package candidate locally generated · crosswalk crosswalk ready no rich export.
What source, artifact, review state, and blocked claim sits behind this statement?
39 linked artifacts · W3C PROV-O provenance · package candidate locally generated · crosswalk crosswalk ready no rich export.
What is the smallest useful review or correction I can offer, and what credit does it grant?
36 linked artifacts · W3C PROV-O provenance · package candidate locally generated · crosswalk crosswalk ready no rich export.
What must exist before a model-backed eval, data release, benchmark, or Open Source AI claim?
32 linked artifacts · W3C PROV-O provenance · package candidate locally generated · crosswalk crosswalk ready no rich export.
Which private, security, runtime, and funding gates remain before public collaboration?
51 linked artifacts · W3C PROV-O provenance · package candidate locally generated · crosswalk crosswalk ready no rich export.
Can this evidence graph become a portable research object package without overclaiming release readiness?
40 linked artifacts · W3C PROV-O provenance · package candidate locally generated · crosswalk crosswalk ready no rich export.
packet ready review not run
Given the generated local JSON artifact package and metadata crosswalk, which one rich export should Continuous Function implement next: RO-Crate, BagIt, SPDX, CITATION.cff, CodeMeta, DataCite, or no rich export yet? Options: 7 · source anchors 9 · rich exports created 0 · review not run.
59 SHA-512 checksum entries.
The Lab OS has a generated local JSON artifact package candidate and metadata crosswalk, but it is not published, archived, citable, release-approved, or a RO-Crate, BagIt bundle, SPDX bill of materials, CITATION.cff, codemeta.json, DataCite metadata, DOI, archive deposit, or release packet.
Implement RO-Crate first
Implement BagIt first
Implement SPDX first
Implement CITATION.cff first
GPT Pro allowed · Oracle allowed · public issue blocked.
Agent execution blocked.
RO-Crate, BagIt, SPDX, CITATION.cff, codemeta.json, DataCite metadata, DOI, archive deposit, release packet, or public package output has been created.
The generated local JSON artifact package is published, citable, archived, release-approved, DOI-backed, or externally reviewed.
A maintainer, GPT Pro, Oracle, human reviewer, GitHub reviewer, OpenReview venue, advisor, funder, learner, or public reviewer has selected the next rich export.
Self-building lab
Make Building Continuous Function the first internal research room.
Mission Control should not be a separate admin dashboard. It should be a gated view over the same graph: TODOs, agent runs, PRs, coverage, costs, decisions, deploys, and blocked work.
cohesion spine ready; review not run
The learner/research cohesion spine connects learning, source grounding, contribution, model/data/eval, review, activation, funding, package, and release boundaries, but no learner test, live runtime, public intake, model/data/eval run, public funding, package publication, GPT Pro review, Oracle review, human pedagogy review, accessibility review, or maintainer approval exists.
8 lanes · runtime blocked · intake closed · model run blocked · package blockedWhat am I looking at, and why does it matter?
Name the object, the narrow question, and the current status before continuing.The route is proven intuitive, accessible, effortless, or learner-tested.Run a learner or pedagogy review note over the route rail and record maintainer disposition.Can I make a prediction and inspect a local witness?
Predict before reveal, compare with witness output, and record what the witness does not prove.Live notebook, browser Python, learner telemetry, or runtime-generated evidence is enabled.Choose one reviewed local-witness prototype and pass accessibility, security, learner, and maintainer review before live runtime claims.What source supports this claim, and what caveat travels with it?
Match one public claim to one source anchor and one limitation.All claims are expert-reviewed, GPT Pro-reviewed, Oracle-reviewed, or maintainer-approved.Capture real review output as an external review note before changing source-to-claim status.Would model/data/eval work actually teach something new here?
Identify the missing model, data, eval, run, review, and release evidence before proposing work.A model, dataset, eval task config, run, benchmark, or Open Source AI release exists.Run one real review of the Transformer Lab eval-task candidate or boundary packet before opening any model/data/eval RFC.What is the smallest useful thing I can improve?
Pick one friction note, source correction, local witness reproduction, caveat rewrite, review question, or proposal question.Public issue submission, public discussion, public credit, agent execution, or model/data/eval proposal intake is open.Privately dry-run one friction note and one source correction before enabling any active public intake route.Who reviewed this, and what credit or authority follows?
Separate feedback, accepted artifact change, public credit, role, funding, and authority.External reviewers have joined, reviewed, approved, earned public credit, or gained authority.Run one real review packet in GPT Pro, Oracle, human, GitHub, or maintainer context and store it as a review note.Which parts are live, private, funded, or still contract-only?
Check whether auth, read model, discussions, agents, funding, and public release are live or blocked.Live auth, live collaboration, public funding, bounties, public agents, or public community venues are active.Choose one owner-approved live activation slice and pass security, accessibility, and maintainer review before enabling runtime behavior.Can this become a portable public research object?
Distinguish local JSON package candidate from RO-Crate, BagIt, SPDX, citation metadata, DOI, archive, and release approval.A public package, DOI, archive, RO-Crate, BagIt bundle, SPDX SBOM, citation release, or release packet is published.Review the generated JSON package candidate before choosing any rich export or public release wording.static route rail implemented; review not run
A static learner route rail is implemented on /lab-os from the learner-path map. It improves orientation but remains review-not-run, non-interactive, not learner-tested, not accessibility-reviewed, and does not open public intake, runtime, model/data/eval, funding, release, or Open Source AI claims.
- 01Object
Name the equation, mechanism, or route object first.
This does not prove the route is learner-tested or easy to understand. - 02Predict
Ask for prediction before witness, reveal, or answer copy.
This does not prove benchmark, production, hosted-demo, or model quality. - 03Source
Pair source support with caveat and blocked overclaim.
This does not prove endorsement, peer review, external review, or Open Source AI readiness. - 04Eval boundary
Show no-model, no-dataset, no-task-config, no-run, and no-release states as useful boundaries.
This does not authorize a task config, run command, metric, benchmark, hosted demo, or funding ask. - 05Contribute
Offer narrow friction, source correction, reproduction, caveat rewrite, or review-packet question paths.
This does not grant authorship, maintainer status, merge authority, funding entitlement, or agent budget. - 06Review
Route GPT Pro, Oracle, human, maintainer, GitHub, and OpenReview-style feedback through review notes.
This does not mean GPT Pro, Oracle, human, public, maintainer, or security review has completed. - 07Package
Show local JSON package, checksum manifest, metadata crosswalk, and export decision state.
This does not publish RO-Crate, BagIt, SPDX, CITATION.cff, CodeMeta, DataCite, DOI, archive, or release packet outputs.
static chooser implemented; intake not open
A static contribution chooser is implemented on /lab-os with six narrow proposal-only actions. It does not open public intake, issue submission, public credit, agents, model/data/eval work, funding, release, or authority claims.
Name the route, object, confusing moment, and what would have helped you move one step forward.
proposal-only learner friction noteprivate dry run only · credit candidate: triage, pedagogy-reviewThis does not prove learner testing, accessibility, pedagogy quality, public intake, public credit, or maintainer approval.Name the claim, source anchor, mismatch, and safer caveat or replacement source.
proposal-only source correctionprivate dry run only · credit candidate: source-review, writing-reviewThis does not prove endorsement, peer review, GPT Pro review, Oracle review, Open Source AI readiness, or release approval.Record the local witness, command or artifact, observed output, environment, and mismatch from expectation.
proposal-only reproduction noteprivate dry run only · credit candidate: validation, softwareThis does not prove benchmark validity, model quality, production serving, hosted demo readiness, or run approval.Rewrite one over-broad claim as a source-backed statement with the missing boundary visible.
proposal-only caveat rewriteprivate dry run only · credit candidate: caveat-review, writing-reviewThis does not authorize public copy changes, review completion, funding asks, model/data/eval work, or release language.Ask one specific question that a GPT Pro, Oracle, human, GitHub, or OpenReview-style reviewer should answer.
proposal-only review-packet questionprivate dry run only · credit candidate: eval-review, pedagogy-reviewThis does not mean GPT Pro, Oracle, human, maintainer, GitHub, or OpenReview-style review has run or endorsed the work.Describe the task scenario, needed data context, eval boundary, and why no run should happen yet.
proposal-only model/data/eval RFC sketchproposal only not public intake · credit candidate: eval-review, validationThis does not approve a model, dataset, metric, task config, run command, benchmark, hosted demo, funding ask, release, or Open Source AI state.evidence ledger ready; runtime not live
A public-safe activation evidence ledger is ready on /lab-os. It is runtime-not-live, separates current contract evidence from missing runtime evidence, and blocks live collaboration, agent, funding, and review-complete claims.
7 rows · live auth blocked · agents not installedlib/accountMemoryAuth.ts resolves dev-owner, Clerk-mirror-required, missing-session, and invalid-dev-owner states.
No production Clerk middleware/provider route protection is verified for this Lab OS surface.Users can sign in and collaborate with account-backed Mission Control.Anonymous fetch receives scrubbed public packet; browser-authenticated owner fetch receives private snapshot with an app-owned users.id.lib/missionControlReadModel.ts defines a read-only port for research_threads, research_comments, ai_runs, and evidence_refs.
No Drizzle/Neon adapter is connected to the read-model port.Mission Control reads live collaboration rows.Private Mission Control shows readModel.status ready with owner-scoped row counts and recent rows from the four collaboration tables.components/discussion/ResearchReadingRoom.tsx provides object-attached discussion UI scaffolding.
No GET/POST research thread API is live.Public discussion or in-app research comments are live.A learner opens an object-attached thread, a reviewer comments with evidence refs, and a maintainer promotes the thread to a tracked issue or review artifact.agent-run-journal.schema.json defines the run record shape.
No Codex, Claude, Copilot, GPT Pro, Oracle, or GitHub agent workflow is installed for this repo.Issues can trigger repo-writing agents or AI reviewers.Approved task creates a draft PR or review note, CI runs, human review completes, and Mission Control records the run.contribution-credit.schema.json defines artifact-scoped credit.
No first-10 reviewer cohort has completed review.External reviewers have joined, reviewed, or received public credit.At least one accepted review artifact has evidence refs, opt-in credit, blocked claims, and maintainer disposition.responses/open-research-lab-os/mission-control-public-summary.json is generated as a scrubbed public candidate.
No owner sign-off on public release is recorded.Mission Control is a live public operating dashboard.Public Lab OS shows only scrubbed statuses, public artifacts, and issue links with no private rows or participant data.funding-item.schema.json defines scoped funding items.
No donation page, bounty board, compute grant, or fiscal host is open.Continuous Function is accepting public funds, bounties, compute grants, or paid review work.A public funding item names exact artifact scope, budget, acceptance gates, caveats, and credit policy.proof runbook ready; runtime not live
A source-grounded activation proof runbook is ready. It names eight proof lanes from auth to package provenance, but every live activation claim remains blocked until runtime evidence, owner review, external/human review where needed, and maintainer disposition exist.
8 lanes · live auth blocked · DB adapter blocked · agent runner blockedClerkProvider or equivalent auth entrypoint is installed for the Pages Router surface.
Owner configures Clerk keys, allowed origins, and production user mirror policy.3 prepared artifactsMission Control has live owner-only authentication and private collaboration rows.Read-only Drizzle adapter queries owner-scoped research_threads, research_comments, ai_runs, and evidence_refs.
Owner approves database URL, read role, migration state, and row-scope policy.3 prepared artifactsThe Lab OS dashboard reads live database collaboration state.GET and POST routes require objectKey, owner or org scope, visibility, source refs, and moderation state.
Owner decides whether the first venue is in-app threads, GitHub Discussions, or both.3 prepared artifactsPublic discussions, chat, forums, or object-thread writes are open.One private reviewer completes a bounded review packet against a named artifact.
Owner invites the first reviewer and approves public credit wording.3 prepared artifactsThe first-10 reviewer cohort, public signup, or public credit ledger is live.Executable labels are maintainer-only and public submissions stay at agent:proposal.
Owner verifies branch protection, workflow permissions, fork approval, secret handling, and one separate live-runner PR.4 prepared artifactsCodex, Claude, Copilot, GPT Pro, or any public issue-triggered agent runner is installed.Visibility scrub checks pass for credentials, private reviewer data, participant data, embargoed work, raw prompts, and private transcripts.
Owner approves public publication and verifies repository settings outside this JSON artifact.3 prepared artifactsMission Control is a published public live operating dashboard.One funding item names artifact scope, budget, acceptance criteria, review state, credit rules, and caveats.
Owner approves payment platform, fiscal boundary, tax/legal responsibility, and public wording.3 prepared artifactsPublic funding, bounties, compute grants, paid review, or sponsor-visible asks are open.Every included artifact is indexed, public-safe, and checksum-covered.
Maintainer approves package wording and any rich export or public release packet.4 prepared artifactsA public package, rich export, DOI, archive deposit, release packet, or citable Lab OS bundle is published.assembly board ready; no run
The model/data/eval assembly board makes the source-intake, model-card, dataset-card, data-information, eval-config, run-manifest, review, package, and release gates visible, but no model, dataset, eval task config, run, benchmark, package publication, Open Source AI claim, GPT Pro review, Oracle review, human review, or maintainer approval exists.
7 lanes · model blocked · data blocked · run not run · Open Source AI blockedcontent/research-rooms/transformer-lab/source-intake-eval-task-candidate-20260708.json records an eval-task candidate.
No source import, task YAML, prompt template, metric, dataset fixture, model, sample output, aggregate metric, run command, or maintainer decision exists.Transformer Lab has an accepted eval task or task config.Record GPT Pro, Oracle, human, or maintainer review output as a review note, then decide whether a proposal-only RFC can be opened.content/research-rooms/transformer-lab/model-data-reference-card.json separates no-model state from future placeholders.
No base model, checkpoint, adapter, model card, parameter pointer, intended-use approval, license review, or evaluation result exists.A Transformer Lab model has been selected, trained, evaluated, served, recommended, or released.Keep the no-model decision visible until a maintainer-approved model proposal names card fields, license, intended use, limitations, and run evidence.content/research-rooms/transformer-lab/data-information-statement-template.json is a no-dataset template.
No dataset, dataset card, Croissant metadata, split, processing note, contamination review, privacy review, or redistribution review exists.Transformer Lab has selected, used, published, or documented a dataset.Open a data-information proposal only after maintainer approval, then fill provenance, license, processing, contamination, privacy, and dataset-card fields.The source-intake candidate identifies task-config patterns before any model run.
No eval config, metric definition, sample schema, dataset fixture, prompt function, scoring rubric, contamination caveat, or safety caveat exists.Transformer Lab has a benchmark, metric, leaderboard, or reproducible eval task.Review the candidate and draft a proposal-only task-config RFC before writing any config or public metric copy.content/research-rooms/transformer-lab/run-artifact-manifest-template.json is ready as a no-run manifest.
No execution, training, evaluation, output, metric, artifact store, log bundle, hardware statement, or reproducible command exists.A Transformer Lab run has executed or produced model outputs, metrics, logs, or artifacts.Only after maintainer-approved RFC and review, execute a narrow run and fill the manifest with real evidence.content/research-rooms/transformer-lab/model-data-eval-boundary-review-packet-20260708.json is ready for review.
No GPT Pro, Oracle, human expert, GitHub reviewer, OpenReview-style reviewer, or maintainer has completed the review.Transformer Lab model/data/eval boundaries have been externally reviewed or approved.Run one real review context, scrub the output, store it as a review note, and record maintainer disposition.The Lab OS artifact package boundary and generated JSON candidate can reference public-safe artifacts.
No RO-Crate, BagIt, SPDX, CITATION.cff, CodeMeta, DataCite, DOI, archive, release packet, public package publication, or maintainer release approval exists.Transformer Lab or Continuous Function has a published model/data/eval release package.Review the local package candidate and decide whether a future export should become RO-Crate, BagIt, SPDX, citation metadata, or no rich export yet.track status, owner, current artifact, next gate
task spec, tool, files touched, diff link, cost, outcome, reviewer
issue link, branch, CI state, reviewer decision, merge state
provider, run class, budget cap, actual spend, funding item
room ladder coverage, claim review coverage, route QA, mobile QA
question, options, sources, decision, caveat, reviewer
review ask, reviewer kind, source packet, output artifact, privacy boundary, credit opt-in, maintainer disposition
question, channel, summary, linked artifact, promotion gate
starting question, prediction, witness, feedback, caveat, contribution bridge
auth gate, read model, discussion persistence, agent runner, reviewer credit, public release, funding gate
review tracks, review packets, intake form, opt-in credit, conduct boundary, maintainer decision
model candidate, dataset card, data information, recipe, run artifact, eval config, model card, release decision
artifact path, layer, source state, standards profile, links, blocked claims, next gate
export profile, included artifacts, metadata state, checksum state, publication gate, blocked release claims
Building Continuous Function should dogfood the lab model as the first internal research room.
264/264 checks passed; visibility starts private and agent work stays behind maintainer approval, draft PRs, CI, and human review.
Publish only after the scrub checklist passes and a maintainer signs off.
path map ready; review not run
The Lab OS has a learner-path operating map that connects object, prediction, source, model/data/eval boundary, contribution, review, and package states. It remains review-not-run and does not prove learner testing, external review, public intake, live runtime, funding, model/data/eval execution, release, or Open Source AI readiness. The route rail has 7 stages and 10 source anchors.
Orient to the object
Predict and witness
Ground the claim in sources
Use a compact path rail: Object, Predict, Source, Eval boundary, Contribute, Review, Package.
Every claim card should show source support, local witness, caveat, review state, and next allowed action.
Offer small, separate actions for friction note, source correction, reproduction note, caveat rewrite, and review-packet question.
attention-serving: KV cache memory equation
transformer-lab: KV memory as the bridge from attention to long-context and serving pressure
This map has been learner-tested, reviewed by GPT Pro, reviewed by Oracle, reviewed by a human pedagogy expert, or approved by a maintainer.
Continuous Function is proven intuitive, effortless, immersive, world-class, learner-tested, accessible, or educationally effective.
The map opens public contribution intake, reviewer credit, GitHub issues, discussions, OpenReview venues, funding, bounties, compute grants, or agent execution.
Review queue
Run the smallest grounded review before opening larger promises.
Collate the current GPT Pro, Oracle, human, security, learner, eval, funding, GitHub, and OpenReview-style review packets into one ranked queue so advice becomes recorded review artifacts instead of scattered chat or broad platform claims. Current state: queue ready gpt pro review recorded; GPT Pro not connected; Oracle not connected; machine source review recorded.
eval design review
Smallest current gate before any eval RFC, task config, model, dataset, run, metric, or public Transformer Lab claim changes.
Machine source-review note and GPT Pro external review note recorded; GPT Pro recommends revise-candidate, while Oracle, human review, and maintainer disposition remain pending.content/research-rooms/transformer-lab/source-intake-eval-task-review-packet-20260708.jsonrun one scoped review packetPrefer the smallest packet that unlocks a concrete next artifact while preserving no-run, no-funding, no-public-outreach, and no-agent-execution boundaries.
Every completed GPT Pro, Oracle, human, GitHub, OpenReview-style, or advisor review must become an external-review-note with artifact refs, cited sources where applicable, accepted/rejected suggestions, caveats, and maintainer disposition before promotion.
gpt pro review recorded revise before rfc
Turn the current rank-one review queue item into concrete recorded critique and maintainer-disposition input without claiming approval, eval RFC, task config, model, dataset, metric, run, benchmark, funding ask, release, or Open Source AI state. Current state: GPT Pro complete · Oracle not run · human not run · maintainer pending · machine source recorded.
Review the Transformer Lab eval-task source-intake candidate as proposal-only work. Use these inputs only: content/research-rooms/transformer-lab/source-intake-eval-task-review-packet-20260708.json, content/research-rooms/transformer-lab/source-intake-eval-task-candidate-20260708.json, content/research-rooms/building-continuous-function/review-queue-20260708.json, content/research-rooms/building-continuous-function/external-review-note.schema.json, content/research-rooms/building-continuous-function/external-review-note-template-20260708.json, content/research-rooms/transformer-lab/model-data-reference-card.json, content/research-rooms/transformer-lab/data-information-statement-template.json, and content/research-rooms/transformer-lab/run-artifact-manifest-template.json. Decide accept-candidate, revise-candidate, or reject-candidate. Answer the five review questions: source-fit, minimum-task-evidence, model-data-run-boundary, review-output-route, and public-copy-risk. Cite artifact paths or official source anchors for each finding. Return an external-review-note-shaped output with decision, question-by-question findings, source caveats, blocked claims that remain blocked, suggested next artifact or no-change rationale, and maintainer disposition needed before promotion. Do not claim that GPT Pro, Oracle, a human evaluator, GitHub reviewer, OpenReview venue, advisor, funder, or maintainer has approved anything unless a named artifact proves it. Do not claim an eval RFC, task config, prompt template, metric, dataset, model, run, benchmark, hosted demo, funding ask, release, or Open Source AI state exists.machine source review · revise candidate
5 findings · 1 high · disposition pending
revise: source fit revise proposal only
revise: minimum task evidence needs revision
Rank-one review queue item
Transformer Lab eval-task review packet
Transformer Lab eval-task source-intake candidate
External review note schema
decision: accept-candidate, revise-candidate, or reject-candidate
answers to source-fit, minimum-task-evidence, model-data-run-boundary, review-output-route, and public-copy-risk
artifact references or official source anchors for every material finding
The workbench has a machine source-review note and GPT Pro external review note, but it has not been run through Oracle, a human evaluator, GitHub reviewer, OpenReview venue, advisor, funder, or maintainer.
The GPT Pro external review recommends revise-candidate; no Oracle review note, human review note, GitHub review, OpenReview venue review, maintainer disposition, public issue, public outreach, public review venue, reviewer credit, compensation, or funding route has been opened by this workbench.
No eval RFC, task config, prompt template, metric, dataset fixture, model selection, sample output, aggregate metric, benchmark, run command, artifact store, hosted demo, release, or Open Source AI claim is approved.
gpt pro review recorded maintainer pending
Package and track the Transformer Lab eval-task candidate for external review without claiming that GPT Pro critique is maintainer disposition, eval RFC approval, task config readiness, model/data/eval readiness, benchmark, funding, release, or Open Source AI state. GPT Pro complete · Oracle not run · human pending · external review note recorded.
Review the Transformer Lab eval-task source-intake candidate as proposal-only work. Use only the attached public-safe artifacts: source-intake-eval-task-review-packet, source-intake-eval-task-candidate, source-intake-eval-task-machine-source-review-note, eval-task-maintainer-disposition-packet, external-review-note.schema, rank-one-review-workbench, external-review-routing-boundary, release-evidence-ladder, Transformer Lab model/data/eval boundary packet, and the regenerated Mission Control public summary. Decide one of accept-candidate, revise-candidate, reject-candidate, or defer-external-review. Answer source-fit, minimum-task-evidence, model-data-run-boundary, review-output-route, and public-copy-risk. For every material finding, cite a local artifact path or official source anchor. Return JSON shaped like cf-lab-os-external-review-note-v1 with target, provenance.usedArtifacts, reviewer.kind, inputPacket, findings, maintainerDisposition.state='pending', privacy flags, blockedClaims, sourceAnchors, and nextGate. Do not claim GPT Pro, Oracle, human, GitHub, OpenReview, advisor, funder, learner, or maintainer approval unless a named artifact proves it. Do not authorize an eval RFC, task config, model, dataset, metric, run, benchmark, public issue, funding ask, release, Open Source AI claim, or agent execution.GPT Pro: review recorded 2026 07 08
Oracle: available only if user provides session
Human evaluator: pending
Eval-task review packet
Eval-task source-intake candidate
Machine source-review note
schemaVersion: cf-lab-os-external-review-note-v1
reviewKind: gpt-pro, oracle, human-expert, maintainer, or other allowed reviewer kind
target artifact id, path, route, review question, and blocked promotion
Remove browser session state, cookies, credentials, billing details, API tokens, account identifiers, and private contact details.
Do not include raw local Claude transcript contents or raw GPT Pro/Oracle chat unless they are scrubbed and converted into a review note.
Do not publish reviewer identity, affiliation, endorsement, conflict details, or public credit unless the reviewer opts in and a maintainer approves.
template ready review not run
Should the Transformer Lab eval-task source-intake candidate be accepted, revised, rejected, or deferred as proposal-only input before any RFC, task config, model/data/eval run, public issue, funding ask, release, or Open Source AI claim? Placeholder findings: 5 · real external review not run · maintainer disposition pending.
source fit placeholder
minimum task evidence placeholder
model data run boundary placeholder
eval-task-external-review-handoff-20260708.json
source-intake-eval-task-review-packet-20260708.json
source-intake-eval-task-candidate-20260708.json
GitHub Pull Request Reviews
OpenReview Notes
OpenReview Invitations
GPT Pro, Oracle, a human evaluator, GitHub reviewer, OpenReview venue, advisor, funder, learner, or maintainer has reviewed, approved, endorsed, funded, or accepted the eval-task candidate.
This template is a completed external review note, maintainer disposition, proposal RFC, task config, benchmark, run artifact, public issue, funding ask, package release, or Open Source AI claim.
lm-evaluation-harness, LightEval, or any eval framework has been imported, selected, configured, executed, or added as a dependency.
gpt pro review recorded revise before rfc
Smallest current gate before any eval RFC, task config, model, dataset, run, metric, or public Transformer Lab claim changes.
transformer-lab-source-intake-eval-task-review-packet-20260708packet ready learner test not run
The platform goal is a world-class learning experience, but learner-tested and world-class claims are blocked until a walkthrough produces recorded friction and accepted fixes.
building-continuous-function-learner-experience-review-packet-20260708packet ready review not run
Boundary review should precede any model/data/eval RFC so the second room cannot drift into model, dataset, benchmark, or Open Source AI overclaiming.
transformer-lab-model-data-eval-boundary-review-packet-20260708review packet ready runner not installed
Agent execution is higher blast radius than artifact review; keep it queued behind concrete review packets and owner-side GitHub governance checks.
building-continuous-function-agent-runner-security-review-packet-20260708route comparison ready no funding open
Funding should attach only after scoped artifacts, review state, budget boundaries, and public wording gates exist.
building-continuous-function-funding-route-comparison-review-packet-20260708Run the rank-1 eval-task candidate review packet and record a maintainer disposition before any eval RFC.
Run one learner walkthrough or GPT Pro/Oracle pedagogy critique against /lab-os/ and record the first friction.
Review the no-model, no-dataset, and no-run boundary before any model/data/eval proposal.
Keep all live agent runner work blocked until GitHub governance and security review pass.
Keep funding review-credit-only until a specific artifact, route, budget, and maintainer decision exists.
The review queue records one machine source-review note and one GPT Pro external review note for the rank-one packet, but does not approve, reject, or promote any review packet.
GPT Pro, Oracle, Claude, human reviewers, GitHub reviewers, OpenReview venues, advisors, funders, and maintainers have not approved Continuous Function through this queue.
No public outreach, public review venue, expert/advisor campaign, public issue review queue, or public learner cohort is open from this queue.
Build order
Build the walking skeleton before deep social features.
1. Make the artifact graph visible
Expose what already exists: content objects, claim checks, sources, caveats, route state, and room ladder status.
A public graph-backed room status page.2. Run one flagship room end to end
Use Attention to Serving as the first complete loop: concept route, witness, script, experiment, eval sketch, review, and contribution issue.
One honest public research artifact.3. Add Mission Control privately
Track TODOs, agent runs, costs, PRs, coverage, and blocked work as the platform building itself.
Owner-only operating room.4. Invite the first reviewers
Use the Join page to ask for narrow critique, not broad commitment. Credit every useful review as an artifact contribution.
First-10 reviewer pilot.5. Move agent work onto issues
Convert task specs into GitHub issues, then route approved labels to Codex/Claude-style runs that open draft PRs for human review.
Task -> agent run -> PR -> review trail.6. Fund only concrete work
Publish sponsorship asks after there are real rooms, bounties, eval queues, and compute needs to fund.
Transparent funding ledger.Discipline
Avoid the traps that make open labs feel empty or unserious.
Do not launch empty social infrastructure before there are people asking to contribute.
Do not imply a model is open source unless data information, code, and parameters meet the release standard.
Do not build GPU orchestration before local reproducible scripts and eval gates exist.
Do not let public issue text directly trigger repo-writing agents without maintainer approval.