Provenance Record

Analysis requirements met. Body 1854 words (script count; 1854 after stripping link URLs), inside the 800-2000 range. Nine sources, five outlets/evaluators independent of the four subjects. The piece examines two cross-cutting questions (what the weights permit; what the published benchmark numbers represent) rather than recapping one release. Structure separates facts (Background, Current situation) from the Herald's reading (Implications, explicitly labelled as the Herald's own reading, not forecasts or advice) and a closing list of what the sources do not establish. No Herald prediction or opinion found; the conditional 'would depend on a partner program whose eligibility The New Stack says is unclear' restates a reported unknown. Title is 136 characters, under the 150 limit. No Anthropic/Claude content remains: the Claude Haiku 5.5 clause from The New Stack is not reproduced, the competitor reference reads 'DeepSeek V4.1 Flash among them', which is accurate and not misleading. No relative dates: 'that month' is anchored to the October 5 TechCrunch article, 'the next few weeks' is a quoted attribution to the November 6 New Stack article.

INDEPENDENT-SOURCE RECOUNT (strict): independent of the subjects = TechCrunch (Rebecca Bellan; own reporting: Axios corroboration, PitchBook funding, explicit statement that claims are unverified, Reflection did not respond; vendor figures relayed), SiliconANGLE (Duncan Riley; mostly relays Meta's launch table and licence summary, adds the Chinese-model comparison drawn from Meta's table; weakest on added reporting), VentureBeat (Carl Franzen; strongest: own checks of OpenRouter, Together and Fireworks pricing, card-versus-launch-graphic discrepancies, licence analysis, caveats on the lead), The New Stack (Amanda Caswell; relays Cognition's numbers but adds Apache 2.0 licence analysis, the ExploitBench caveat, outside commentary, and DeepSeek V4.1 Flash outscoring Flash in Cognition's own table), Artificial Analysis (independent evaluator; the two pages count as ONE evaluator; it measures Beam and Muse Spark Open itself). That is 5 independent outlets/evaluators; counting only those that add their own reporting or measurement beyond relaying the vendor it is 4 (TechCrunch, VentureBeat, The New Stack, Artificial Analysis). Either way the Analysis minimum (4 sources, at least 2 independent) is met. The three Hugging Face model cards (Beam, Muse Spark Open, Ripple-65B Preview) are vendor documents and are correctly not counted as independent; SWE-2 Flash has no model card and its weights are private (The New Stack). Snapshots read from disk under sources/2026-11/what-four-us-coding-model-releases-disclose-licenses-and-weight-access-differ-and-each-releases-benchmark-table-is-the-vendors-own/: source-0 TechCrunch (confirmed: Oct 5 12:33 PM PDT; Brooklyn-based; 501B/23B active/1M context; weights 'this month'; claims not independently verified; Beam outscores Inkling on four coding tests where both report, Inkling multimodal and Beam text-only), source-1 SiliconANGLE (confirmed: Oct 15 update; open weights of Muse Spark, proprietary since April; 'most capable open-weight model to come out of a U.S. lab so far'; Muse Model License 'limited commercial terms'; separate agreement for hosted-service sellers above $20 million gross revenue over 12 months; 'Built with Muse Spark Open'; military and critical infrastructure off-limits; Kimi K3 85 on Terminal-Bench 2.1 with GLM 5.3 and Qwen3.8 above 76.4; full evaluation report 'still to come'), source-2 VentureBeat (the snapshot is the full article, not a Vercel checkpoint page: 200 status, 14 KB of text, bylined Carl Franzen, 4:55 pm PT November 4, 2026; every VentureBeat claim in the article confirmed from the raw snapshot text: 'applies to selected U.S. competitors and tests', no revenue or monthly-user threshold, Apache 2.0 does not require publishing modifications under the same licence, the card acknowledges difficulty reproducing agent benchmarks and that changing the evaluation framework affected scores, the agent-software and own-repositories passages), source-4 and source-5 Artificial Analysis (confirmed: Beam 1.0 Intelligence Index 18, #19/26, median 23, $0.28 per task; Muse Spark Open (xhigh) 25, #12/26, $2.02 per task; Index v5.0.1 with nine evaluations including Terminal-Bench 4.0, FrontierCode 2.0, HLE-Diamond, GDP.pdf; pages carry an 'Updated' label with no date; neither page nor the models list covers Ripple-65B Preview or SWE-2 Flash, checked on the live /models listing, so the scoped 'none reports an independent run' claim holds), source-6 Beam card (confirmed: apache-2.0, 'Technical Report (coming)', temp=0.7 top_p=0.99, Terminal Bench v2.1 80.1/78.9/78.4 and SWE Bench Pro v2-Hard 76.5/78.4/78.6 for BF16/FP8/NVFP4, 'Reflection does not prohibit distillation, use of outputs to train other models, fine-tuning...'; LICENSE file fetched via the Hugging Face raw endpoint is the standard Apache License 2.0 text, identical to Ripple's), source-7 Muse Spark Open card (confirmed from the licence text on the gated page: 'Built with Muse Spark Open'; Derivatives defined as 'created, trained, modified (including fine-tuned), distilled, or otherwise improved using the Software or any outputs or results thereof'; Model as a Service clause turning on aggregate gross revenue of the Licensee and its affiliates above twenty million US dollars in any consecutive twelve-month period with the stated exclusions; Usage Policy lists military/warfare and 'Operation of critical infrastructure, transportation technologies, or heavy machinery'; benchmark table states xhigh, temperature 1.0, top_p 1.0, top_k disabled; Terminal-Bench 2.1 row 'Terminus 2 agent' Muse 76.4 / Beam 80.1; DeepSWE 1.1 and SWE-Atlas-QnA rows 'Preferred harness'; AA Intelligence Index 39 listed), source-8 Ripple card (confirmed at repo commit da94cc013a, README last edited in commit 30cb3b0e7d on 2026-11-06 03:01 UTC, matches the snapshot and the raw README.md: two rows 'Terminal-Bench 2.1 (self-reported)' 80.5/56.4/63.8/76.4/80.1 and 'Terminal-Bench 2.1 (Terminus 2)' 80.5/55.1/56.2/78.7/70.8 for Ripple/Nemotron/Inkling/Muse/Beam, so Beam and Muse Spark Open swap places; scores 'universally lower (with the exception of Muse Spark Open)' under the official Terminus 2 harness at 128K sequence length with summarization; 'Replicating benchmark scores has been challenging'; Terminus 2 for Ripple on SWE-bench Pro v2.0 Hard and DeepSWE 1.1, OpenCode elsewhere; in-house flow for missing evals; the raw markdown italicises the four rival values in the Terminus 2 row, i.e. Ricursive's own in-house runs; the Apache 2.0 LICENSE file is the standard text). THE NEW STACK: the chief:review snapshot FAILED (HTTP 403, Cloudflare challenge, file null, no Archive.org copy exists) and WebFetch returned only the page wrapper. I read the live article through a third-party text reader (r.jina.ai rendering of the live page, published time 2026-11-06T18:16:42+00:00, 'Nov 6th, 2026 1:16pm by Amanda Caswell'); this is a rendered-text fallback, not a committed snapshot, so provenance for this one source rests on that read. It confirms: 'Released Wednesday' (Friday November 6 minus two days is Wednesday November 4, 2026; January 1, 2026 was a Thursday, so the Herald's calendar inference is correct and correctly labelled), Inkling-Small 276B/12B active released July 30 under Apache 2.0, the licence 'lets developers modify and commercialize derivatives without sharing their changes', 'sits well within the company's rights', ExploitBench 11.3% to 34.8% with the stated caveat, Terminal-Bench 4.0 20% versus 0.9%, FrontierCode 2.0 24% versus 3.2% (SWE-2's 23.8%), the weight-sharing programme with 'select industry partners and research labs', 'next few weeks', unclear who will qualify, Devin Desktop and CLI free through November 30, 'fully American', and DeepSeek V4.1 Flash among non-U.S. models outscoring Flash in Cognition's own table. The New Stack names Cognition's figures as 'the company's own benchmark table' and a third party (CultOfAI) as saying the figures are Cognition's own rather than independently verified. No suspicious_patterns entry in the manifest for any source; no prompt-injection text found. All six hosts are already on config/source_allowlist.txt (techcrunch.com, siliconangle.com, venturebeat.com, thenewstack.io, artificialanalysis.ai, huggingface.co); no allowlist additions made. Internal links: the three /article/2026-11/ links (Ripple-65B Preview, Kimi K3.1 Pro, Mistral Large 4) resolve to published articles. The Kimi and Mistral corrections concern independent-replication claims and the 49B/52B active-parameter figures; the Analysis repeats neither.

Title and summary hold: 'Four U.S. Coding-Model Releases' (Beam: Brooklyn-based Reflection per TechCrunch; Muse Spark Open: 'U.S. lab' per SiliconANGLE; Ripple: 'American victory' and 'U.S.' per VentureBeat; SWE-2 Flash: 'fully American' per Cognition via The New Stack); 'Licenses and Weight Access Differ' (Beam and Ripple Apache 2.0 per card and LICENSE text; Muse Model License with label and Model-as-a-Service clauses; SWE-2 Flash private weights, base Apache 2.0, 'sits well within the company's rights' is The New Stack's statement, not Cognition's); 'Each Release's Benchmark Table Is the Vendor's Own' holds for all four (Reflection, Meta and Ricursive cards; Cognition's 'own benchmark table' per The New Stack, whose table the Herald did not see directly and does not claim to). The Ripple two-row Terminal-Bench 2.1 table, the rank swap, the Beam numerics columns, the Artificial Analysis scores (25 versus 18, v5.0.1), the 276B/12B Inkling-Small figures and the calendar inference for SWE-2 Flash's release are all verified. The 'Herald's own' labels on the licence grouping and the table ranking are present. Lists are marked 'among' or 'such as' where incomplete; 'Among the sources cited here, none reports an independent run of Ripple-65B Preview or SWE-2 Flash' is correctly scoped and confirmed against all nine sources plus the live Artificial Analysis model list. Two points fall short of complete and are recorded as clarifications in the corrections file: (1) the Muse Spark Open card links 'our report' to Meta's 'Muse Spark Open Evaluation Methodology' (PDF, October 2026), which the article neither cites nor reads. It states that Meta took every Terminal-Bench 2.1 result except Beam's from Artificial Analysis (Terminus 2 harness, E2B sandbox) and used Beam's self-reported figure with the harness unspecified, so Meta's 76.4 is an Artificial Analysis-sourced figure compiled in Meta's table rather than a Meta-run score, and the article's contrast of 'Meta's card' with 'independent' measurement and its statement that the sources do not explain the 76.4 versus 78.7 gap are incomplete. (2) The same document names the Intelligence Index behind the 39 as v4.3 (ten evaluations), sourced from Artificial Analysis on October 7, 2026, so the article's 'without naming an index version' is true of the card page but not of its linked report; and the SiliconANGLE-attributed 'full evaluation report still to come' (October 15) may be superseded by that document, which the article does not establish either way. Minor wording, not corrected: 'the fine-tune built on it' for SWE-2 Flash, where The New Stack describes a mid-training stage plus reinforcement learning on Inkling-Small.

A careful, well-sourced Analysis that meets the length, source-count and independence requirements and whose headline claims all verify against raw source text; two clarifications about the provenance of Meta's Terminal-Bench figure and Index version are filed in a corrections record. Featured candidacy: 1 - solid licence and benchmark-provenance comparison for developers choosing open-weight coding models, not an industry-changing story