Content Quality: A genuine Analysis: it compares six agent-assisted rewrites (Bun, GitHub's Copilot runtime, OpenAI's Habitat, Theo Browne's TypeScript port, Cloudflare's tsrs, DHH's Campfire benchmark; recounted from the body, six confirmed) along three axes (who supplies the figures, which baselines are contested, what verification is described) and separates reported facts (Background, Current situation) from labelled analysis (Implications, introduced with 'They are not predictions'). 1,486 words with link URLs stripped (range 800-2000).
Source Verification: Snapshots read from disk (source-0..5.html.gz, all HTTP 200, suspicious_patterns null for all six), then a live curl re-fetch of each URL confirmed the key figures still present. source-0 (The New Stack, Sawers, Sep 17): 800,000+ lines of Rust, 128 pull requests, Toub 'AI agents wrote most of the code', Bun 535,496 lines of Zig, Sumner with Claude Code agents, eleven days, all six platforms, Habitat (OpenAI, Codex and GPT-5.5, 'handling 95% of our production requests', six times less CPU, 15 times less memory), 'individual case studies, reported by the companies with the most to gain', 'inherently self-interested', 'shouldn't be read as a Claude-versus-Copilot benchmark': all verbatim/faithful. source-2 (The Register, Jackson, Sep 18): ~$120,000 tokens, ~3 weeks developer time, 430,000 TS lines to 800,000 Rust, 7.55 vs 120 lifecycles per second (15.9x), 'over 135 releases' across 14.5 weeks, 'dozens of regressions', Toub quotes, Crossman at RustConf, Kelley 'unreviewed slop' (parenthetical attribution in the source), 'Anthropic-owned' Bun: all confirmed. The 135 releases vs 128 pull requests difference is real and the article's 'different units' framing is fair (The Register itself says roughly 1.3 port pull requests opened per day). source-3 (Register, Oct 7): 36,260 / 241 / 722 / 3,860 requests per second on the room page, observers' claim that Rust was tuned and others not, Elixir rewritten by others and then beating Rust on some tasks, the 'benchmarking how much attention each port got' and 'extra points for Elixir' quotes: confirmed. source-4 (Register, Oct 15): Browne, twice as fast as TS 7 and 12.5x TS 6, has not inspected the code, $400,000+ with GPT-5.6 Sol and GPT-6 Astra at ~84 percent, ~$24,000 with Claude Opus 5.5, nearly 100 percent conformance: confirmed, all Browne's own claims. source-1 (New Stack, Oct 20): up to 8x faster and 70% less memory than TS 7.0, Schwenk 'I wrote zero code', Opus 5.5, 19 GiB and 80 seconds (single-threaded check) then roughly seven seconds in CI, Rosenwasser HN quote, Miri/conformance/error-injection and the 'broken compiler with good marketing' quote, Chen quote, circular types, WebAssembly and garbage-collector point: confirmed. The 'dropped to roughly seven seconds' sentence compresses two statements but is faithful. source-5 (InfoQ, Curtis, Oct 27): one developer 5.3x faster and 72% less memory with exit code 2 vs 1; ImportStatic 2.7x against 7.1 beta ('the 8x headline is measured against the older baseline'); memory-model scepticism; experimental, no watch mode or language server; merging with tsc-rs: confirmed (InfoQ's 'quietly break a CI script' is InfoQ's own wording, the article's 'in one tester's account' is a negligible attribution stretch). INDEPENDENT-SOURCE COUNT: 3 outlets (The New Stack, The Register, InfoQ), all independent of the subjects; 6 sources total. The Register counts once as an outlet even with three articles, but the three articles are by one reporter and cover distinct stories; the 'at least 2 independent' test is met by outlets alone (3). No unverified outlet (the writer's Google News note) is relied on: every claim traces to the six cited URLs. Hosts thenewstack.io, theregister.com, infoq.com are on config/source_allowlist.txt (exact hostnames, lines 119, 118, 174); no allowlist edits.
Factual Accuracy: Title clauses are each supported and labelled. 'Large Gains': 15.9x (Copilot benchmark), 8x (tsrs), 36,260 vs 241 rps (Campfire), 95% of requests and 6x CPU (Habitat). 'Most figures are self-reported': the Herald's own tally in Implications (Copilot by Microsoft's Toub, Habitat by OpenAI, tsrs by Cloudflare/Schwenk, Browne's by Browne, Campfire by Hansson); The New Stack states it directly for Bun and GitHub; outside measurement exists only for tsrs (InfoQ's two testers) and Campfire (critics), which the article states. Defensible and explicit. 'Baselines are questioned': named parties are InfoQ (tsrs 8x vs 5.3x vs 2.7x baselines), ImportStatic, Rosenwasser, and the Campfire critics reported by The Register (one observer on X, Jaana Dogan unnamed in the article). 'Verification is a stated concern': Toub, Crossman, Schwenk, Chen, all named. Findings (two minor, below): (1) the intro says the Herald covered 'some of these projects individually, including ... the TypeScript 7.1 beta'; the linked TypeScript 7.1 beta article does not mention tsrs, Rust or any rewrite, so it is not coverage of one of the six projects. The Copilot and Bun links are accurate. (2) Disclosure gap, below. Internal links /article/2026-09/21-..., /article/2026-05/18-..., /article/2026-10/21-... all resolve to existing files. No orphan URLs (body URLs = sources array; chief:review body-sources check passed).
Overall Assessment: Meets the Analysis requirements (6 sources, 3 independent outlets, a topic examined across six cases, facts separated from labelled analysis) and the numbers and quotes check out against all six snapshots and a live re-fetch. Approved with two clarifications (link description; Anthropic/Bun and newsroom-model disclosure). Featured candidacy: 2 - cross-industry analysis of how AI-agent rewrites are measured and verified, relevant to a broad developer audience