Provenance Record

Strong 776-word News piece on OpenAI's Symphony release (Apr 28, 2026). Structure (Overview / What We Know / How the Workflow Hangs Together / What We Don't Know) is solid. The article goes deep into spec mechanics (state machine, polling defaults, GraphQL integration, lifecycle) without becoming a regurgitation of the README. The 'What We Don't Know' section properly bounds the 500% claim as OpenAI's own internal metric.

All 5 source snapshots read; every claim verified verbatim against the underlying material. • source-0.html.gz (GitHub repo openai/symphony): Confirms repo description 'Symphony turns project work into isolated, autonomous implementation runs, allowing teams to manage work instead of supervising coding agents' (verbatim); 'low-key engineering preview for testing in trusted environments' (verbatim); Apache-2.0 license; 21.5k stars (article says 'more than 21,000'); 1.9k forks (article says 'over 1,900'); Elixir as primary language (95.5%). • source-1.html.gz (SPEC.md): Confirms 'long-running automation service that continuously reads work from an issue tracker (Linear in this specification version), creates an isolated workspace for each issue, and runs a coding agent session for that issue inside the workspace' (verbatim); poll interval default 30000 ms (= 30 seconds); all five phases (PreparingWorkspace, BuildingPrompt, LaunchingAgentProcess, StreamingTurn, Finishing); all five terminal states (Succeeded, Failed, TimedOut, Stalled, CanceledByReconciliation); stall_timeout_ms default 300000 (= 5 minutes); exponential backoff retry; LINEAR_API_KEY env var; api.linear.app/graphql endpoint; bash -lc launch model; WORKFLOW.md YAML front matter + prompt body structure. • source-2.html.gz (elixir/README.md): Confirms App Server mode launch; linear_graphql tool exposed; Phoenix stack with LiveView dashboard, JSON API at /api/v1/*, Bandit HTTP server; max_turns default 20. • source-3.html.gz (InfoWorld): Confirms verbatim 'an open-source specification for turning issue trackers such as Linear into control planes for Codex coding agents'; 500% landed-PRs increase 'in the first three weeks' on 'some internal teams'; verbatim list 'monitor issue states, restart agents that crash or stall, manage per-issue workspaces, watch CI, rebase changes, resolve conflicts, and shepherd pull requests toward review'. • source-4.html.gz (Gizmodo): Confirms verbatim 'every open task gets an agent, agents run continuously, and humans review the results'. Every inline link points to a source containing the specific claim. No misattribution. No fabricated quantitative claims.

All factual claims verified. The article correctly handles the 500% figure as OpenAI's internal claim (in 'What We Don't Know' section) rather than treating it as an independent benchmark. Spec details (defaults, state machine, GraphQL plumbing) match SPEC.md exactly.

APPROVE_WITH_CORRECTIONS. High-quality, technically detailed coverage with every claim sourced and verified verbatim. Verdict driven by the github.com (3 entries — primary-source repo links) and infoworld.com (1 entry) allowlist warnings. Article is publishable as-is.