Provenance Record

Well-structured News piece (1,012 words, within the 400-1200 range): Overview, What Was Launched, The Benchmark, Why TrueFoundry Is Giving It Away, and a 'What We Don't Know' closer that explicitly frames the benchmark as vendor-self-reported and not independently reproduced. All numeric specifics (14 tasks, $2.90 vs $11.80, $8.50 vs $11.80 vs $21, 3.8M/10M/16.5M tokens, 40/63/64 minutes, $5/$25 and $0.73/$2.28 per-1M pricing, 50,000-token compaction threshold) trace to a cited source.

All 4 sources read from local gzip snapshots in sources/2026-08/truefoundry-open-sources-trueforge-an-agent-harness-benchmarked-up-to-75-cheaper-than-claude-managed-agents/ (no re-WebFetch needed; sha256 of each decompressed file matches manifest.json). source-0.html.gz (VentureBeat, Carl Franzen, Aug 19 2026): confirms headline 30%-75% figure, $2.90 vs $11.80 (75% cheaper, different-model comparison) and $8.50 vs $11.80 (~30% cheaper, same-model Opus 4.8 comparison), 'San Francisco B2B machine learning startup co-founded in 2021 by former Meta engineers,' MCP-tool-schema-delay/subagents/file-offload/compaction description, 50,000-token default compaction threshold, sandbox-as-tool description, SQLite-local vs Docker Compose/Helm+Postgres/Redis production deployment, NetApp beta user and Automattic early user attributions, and all three Gutgutia quotes verified verbatim word-for-word against the article's quoted text. source-1.html.gz (TrueFoundry launch blog, Shubham Agarwal, 18 Aug 2026): confirms 'the harness we run in production,' MIT license, the exact `npx @truefoundry/trueforge` command, and the sandbox-as-tool 'only spins one up when the agent needs to run code' quote verbatim. source-2.html.gz (TrueFoundry benchmark-methodology blog): confirms the DevRev Enterprise-Bench 14-cross-system-task methodology verbatim ('14 cross-system tasks at varied difficulty that read like real B2B ops work'), the blind-LLM-judge / no-partial-credit grading description, the 'two obvious alternatives ... one blind judge: the closed Claude Managed Agents and the open-source deepagents' framing, the GLM-5.2-vs-Opus-4.8 table (~11/14 solved both sides, $2.9 vs $11.8, -75%), the same-model three-harness table (TrueForge $8.5/3.8M tokens/40min, Claude Managed Agents $11.8/10M/63min, deepagents $21/16.5M/64min, ~10/14 solved), the 'Claude Managed Agents can't run this setup at all: it's Anthropic-only, so the cheaper-model lever was never on the table' quote, and the exact pricing quote '$5 / $25 per 1M in/out for Opus 4.8; $0.73 / $2.28 for GLM-5.2' -- all verbatim matches. source-3.html.gz (GitHub truefoundry/trueforge): confirms MIT License and repository identity. No hallucinated quotes or fabricated specifics found anywhere in the article.

All specific numbers, quotes, and technical claims trace to the cited sources as documented above. No fabricated or unsourced specifics found. Every body URL is present in article.sources (no orphan-source provenance gap); one sources[] entry (github.com/truefoundry/trueforge) is not hyperlinked in the body but is used to independently confirm the MIT license and repo identity, which is acceptable background sourcing.

High-quality, thoroughly sourced submission. The one automated error-level finding was a false-positive prompt-injection match on ordinary technical prose (verified in full context and documented above), and the one warning was an expected allowlist gap for a vendor's own blog used transparently. All quotes are verbatim, all specifics trace to sources, the vendor-benchmark framing is honest and appropriately caveated, and the story is a genuinely new topic with a valid cross-reference to prior coverage. Approved without corrections.