Loading evidence and metric snapshots
Upstream is an evidence ledger for open source AI accelerator support. It helps a researcher answer a hard question: what public evidence connects a named capability in a repository to tests, CI, a release, a compatible benchmark, a reviewed decision, and a cited report?
The problem is not a lack of GitHub events. The problem is that pull requests, workflow runs, releases, benchmark results, and analyst decisions live in different systems and can be easy to connect incorrectly. Upstream keeps those links typed, versioned, and inspectable. A missing link remains missing. It never becomes a failure score.
This identifies the research snapshot used by the public pages. It is not the same as the current collector checkpoint or page-read time.
Pollers and signed webhooks gather GitHub pull requests, workflow attempts, releases, repository state, and benchmark submissions. The original observations stay retained.
The data plane deduplicates provider records, versions identities, and keeps missing or conflicting facts explicit. Collection does not create a support claim.
Related material changes enter an analyst queue. Cases can be grouped for review while each source observation and membership remains traceable.
A chain connects a capability to implementation, tests, pull requests, CI, releases, benchmarks, an analyst decision, and a report. Unproven stages and links stay visible as gaps.
Public research, exports, and reproducibility packets pin a cutoff and version set so a later collection run cannot silently rewrite a cited result.
Upstream shows three related times. A collection checkpoint says when the collector last checked a provider resource. Source activity says when the newest qualifying public event happened. A research cutoff says which source activity was admitted when the public snapshot was frozen. Collection can be current while source activity is old or a published snapshot is several days old.
A stale source observation means its age exceeded the policy pinned to that snapshot. A stale chain comes from evidence or stage freshness. An unproven stage is first a completeness gap. It can also leave chain freshness missing when no freshness proof exists, but it does not make the chain stale. Those causes stay separate where the data supports it.
Upstream makes controls look like controls and leaves evidence as text. You should not need to click a card, count, or status label to discover whether it does something.
Choose the group that matches what you need to do. All 24 stable journeys have complete software paths. Each card separately shows who can access the path and whether current live evidence, a named human action, or operating proof is still open. The blue guide on the destination then shows the active journey, ordered steps, and next useful action.
Start here if the product, evidence model, or weekly dashboard is new to you.
New user
Understand the problem, evidence pipeline, freshness clocks, roles, and routes.
Boundary: The walkthrough explains the system. It does not replace source inspection.
Open Start hereResearch reader
Scan contribution, CI, parity, benchmark, coverage, and freshness separately.
Boundary: The dashboard is a set of signals, not one health score.
Open DashboardFind, filter, compare, and share published ecosystem findings.
Ecosystem researcher
Compare contribution volume, response time, merge time, and aged backlog by repository.
Boundary: Pinned summaries cover the metric window. PostgreSQL detail rows are sanitized pull request facts frozen with that snapshot. Evidence and chain links appear only when a separate public relationship exists.
Open ActivityMaintainer
Compare pass rate, exclusions, duration, queue time, reruns, flakes, and attempts by repository.
Boundary: Pinned summaries cover the metric window. PostgreSQL detail rows are sanitized workflow run and job facts frozen with that snapshot. Evidence and chain links appear only when a separate public relationship exists.
Open ActivityPerformance researcher
Read compatible InferenceX results with method and source context.
Boundary: Throughput alone does not prove capability or release parity.
Open BenchmarksCoverage reviewer
Separate collected, mapped, missing, stale, and unavailable states.
Boundary: Missing evidence does not mean unsupported capability.
Open CoverageResearcher
Filter by ecosystem, repository, capability, workflow, status, freshness, and snapshot.
Boundary: Filters operate on published chains, not the broader collection cohort.
Open ResearchResearcher
Copy a canonical URL for the same filters and frozen snapshot.
Boundary: A latest-data link can resolve a different snapshot later.
Open ResearchChange reviewer
Separate source, cohort, taxonomy, method, and freshness changes.
Boundary: The view does not compute a numeric delta across incompatible methods.
Open ChangesAuditor
Trace one exact methodology version to every frozen snapshot and public report that used it.
Boundary: The index covers public snapshots and public reports in this read model. Private drafts are excluded.
Open ChangesTrace a claim to its sources, reproduce its inputs, or use the public API.
Accelerator specialist
Trace implementation, tests, pull requests, CI, releases, and benchmarks.
Boundary: The live SDPA chain remains partial until its missing links are proven.
Open ResearchAuditor
Follow a claim through a chain to sanitized evidence and source hashes.
Boundary: Private payloads and analyst notes are not public evidence.
Open ResearchReport reader
Audit a frozen publication, correction history, and cited versions.
Boundary: The first named-human live report has not been published.
Open ReportsAPI consumer
Read versioned JSON, OpenAPI, ETags, and snapshot metadata.
Boundary: Legacy fixture-only resources return a typed unavailable response in production.
Open Developer guideReproducer
Download a signed packet with pinned versions, public evidence, and definitions.
Boundary: The packet reproduces published inputs. It does not establish authorship or approval.
Open ResearchData consumer
Download the visible research rows as JSON or CSV with matching filters.
Boundary: Exports contain public projection fields only.
Open ResearchSubmit proposed definitions, registry changes, or benchmark evidence for review.
Taxonomy maintainer with analyst mediation
Propose and review effective-dated capability or taxonomy definitions.
Boundary: Submission does not change approved state and currently requires an analyst identity.
Open New proposalProducer with analyst mediation
Prepare a signed upload envelope and review its durable intake state.
Boundary: No live producer bundle has been accepted and submission currently requires an analyst identity.
Open Benchmark intakeMaintainer with analyst mediation
Submit a repository or mapping proposal without mutating approved state.
Boundary: Submission currently requires an allowlisted analyst identity.
Open New proposalUse the protected analyst workflow to resolve cases and publish cited research.
Analyst
Record a narrow, versioned decision with a reason and audit receipt.
Boundary: Representative analyst validation remains open.
Open Analyst inboxAnalyst
Claim, renew, release, defer, resolve, merge, or split evidence cases.
Boundary: The large historical backlog still needs an approved compaction plan.
Open Analyst inboxReport editor
Draft, freeze, publish, correct, and retract cited research.
Boundary: A named analyst must write and approve the narrative.
Open Report workflowKeep collection, freshness, access, recovery, and database state healthy.
Collection operator
Inspect checkpoints and request a full repository sync with a reason.
Boundary: A sync gathers source facts. It does not rewrite a frozen research snapshot.
Open Collection controlsReliability operator
Monitor queues, freshness, recovery, database state, and deploy identity.
Boundary: The seven-day operating proof remains open.
Open OperationsThis repository was built as a direct technical response to the SemiAnalysis Full Stack Engineer, AI Accelerator Open Source Ecosystem role and its challenge. The software covers every named capability in the current posting. It does not yet prove the research-team outcome because the first named-human report, real producer bundle, broader live evidence, and operating pilot still require external or human work.
| Posting asks for | What is implemented |
|---|---|
| Next.js, React, TypeScript | Implemented in the web app with App Router, Server Components, route handlers, and Server Actions. |
| PostgreSQL schema design | Implemented with versioned history, partitions, constraints, indexes, public projections, and audited commands. |
| GitHub API pipelines | Implemented with scheduled REST and GraphQL collection, signed webhook intake, checkpoints, retries, and deduplication. |
| Pull requests and CI health | Implemented with repository comparison, contribution timing and backlog, CI pass rate, exclusions, duration, queue time, reruns, flakes, and optional public evidence links. |
| Test parity and accelerator support | Implemented with reviewed implementation, pull request, test mapping, artifact policy, execution, release, and benchmark relationships across CUDA, ROCm, TPU, and Trainium. |
| Performance benchmarks | Implemented with signed intake, validator receipts, raw-sample statistics, compatibility checks, public comparisons across all four ecosystems, and analyst review. |
| Weekly technical reports | Implemented with draft discovery, versioning, citation, correction, retraction, signed packets, and a public Atom feed. A named-human live report is still open. |