How this page writes itself.

Orchestration is code. Only one model has a tool: the researcher can read pages, and code decides which ones. The one that reads the web cannot write anything; the one that writes never reads the web; only code publishes.

Figures from the run of 14 October 2026

1. code
   Collect

   Reads feeds from a hand-kept allowlist. Drops what was already seen.

   142new items
2. Haiku 5.5
   Triage

   Scores the items. Keeps the best six, or none.

   6topics studied
3. Sonnet 5.5 × N
   Research

   One agent per topic, with a single tool: fetch_page. Allowlist, at most 3 pages and 6 turns, every call traced.

   30facts proposed
4. code
   Gate 1

   Each quote must appear word for word in a page the agent actually read. Otherwise the fact is gone.

   30quotes found
5. Sonnet 5.5
   Fact check

   Fresh context. Judges each fact against its quote, in both languages. A refused fact never reaches the writer.

   21facts validated9 refused
6. Sonnet 5.5
   Writing

   Writes in French and English from validated facts only. Never sees the web.
7. Sonnet 5.5
   Statement check

   Judges each written sentence against the facts it cites.

   5statements published
8. code
   Gates 2

   Verifier threshold, volume, budget. One failure and nothing is published.
9. code
   Publish

   Writes the files. The trace is published either way.

What this does not guarantee

The gates prove a sentence is sourced, not that the source is right. An allowlisted page that is wrong will be quoted faithfully. And an exact quote can be refused because the page changed in the meantime: the pipeline prefers a day without an edition to a doubtful one.

Verifier eval

10/10 trap cases judged correctly, last run 13 October 2026.

Last 2 runs

What the pipeline refused to publish

• A Scion engineer replied that since Claude is by far the heaviest user of Scion and is necessarily online, offline support wasn't important.relies on a fact its quote does not support (The quote does not say the speaker is a Scion engineer or that this was a reply, and it does not mention offline support; the URL alone does not establish these details.)
• Google's lightweight Gemma model was fine-tuned on local scam conversations for scam detection.relies on a fact its quote does not support (The quote says Gemma was fine-tuned on local scam conversations but does not mention scam detection as the purpose, though this is only implied by the URL.)
• The pipeline has two stages: a lightweight classifier screens inbound messages using minimal power, then the fine-tuned Gemma model analyzes a larger snapshot of the conversation to confirm a threat orelies on a fact its quote does not support (The quote covers only stage 2; the lightweight classifier screening inbound messages with minimal power is not in the quote.)
• Scams make up over half of Singapore's reported crimes and account for more than S$4 billion in losses since 2019.relies on a fact its quote does not support (The quote has no explicit subject; the claim assumes it refers to scams, which is not established by the quote alone.)
• In AWS's test of 72 agent runs, the agent answered 67 of the 72 runs correctly, including all 18 at the xhigh effort level.relies on a fact its quote does not support (The quote does not mention AWS's test or that the 72 runs were agent runs, beyond what the URL implies; it also frames xhigh without confirming the 72-run total as stated in the claim text.)
• Grok 4.7 is accessed on Bedrock through two cross-Region inference profiles: us.xai.grok-4.7 (routing within the US) and global.xai.grok-4.7 (commercial AWS Regions).relies on a fact its quote does not support (The quote says two profiles are available but does not list their names or routing scopes.)
• Reasoning tokens count as output tokens and draw down max_completion_tokens before the answer text; if the limit is reached during reasoning, the response returns without answer text.relies on a fact its quote does not support (The quote does not say that the response has no text if the limit is reached during reasoning.)
• Mean output tokens per run rose from 1,184 at low to 9,147 at xhigh, while tool calls stayed roughly flat (17.2 to 18.7).relies on a fact its quote does not support (The quote covers only the flat tool calls; the output token means of 1,184 and 9,147 are not in the quote.)
• Under the study's controlled conditions, Gated KalmaNet (GKA) performs best among the state space models, followed by Gated DeltaNet (GDN) and then Mamba-2.relies on a fact its quote does not support (The quote does not say that GKA, GDN and Mamba-2 are state space models or that they are the best among them in a layer sense beyond ranking.)
• Anthropic replaces Git with a new source control systemthe verifier does not accept this statement (The headline claims Anthropic replaces Git with a new system, which the cited claims only imply through migration from Git, and the analysis about strain on tooling goes beyond the facts.)
• A model fine-tuned with GovTech Singapore runs entirely on-devicethe verifier does not accept this statement (The headline says a model fine-tuned with GovTech Singapore, but the cited claims do not establish the model purpose or that DeepMind fine-tuned it, and the cited facts cover only a subset; the privacy-design framing is acceptable but the title is not fully supported.)
• Haiku 5.5 uses a new tokenizer: the same long prompt uses around 1.25x as many tokens as with Haiku 4.5, according to the author's Claude Token Counter tool.relies on a fact its quote does not support (The quote only says the same prompt uses about 1.25x more tokens and does not state that Haiku 5.5 uses a new tokenizer.)
• Max 5x users will get $100 in API credits per month, Max 20x users $200, and Team subscribers up to $500, pooled across their users.relies on a fact its quote does not support (The quote says "credits" without specifying they are API credits, so the claim adds a detail not in the quote.)