Side N
The Boundary Is Not the Model
Which sources back which section of this article, and what its evidence base looks like.
Which sources back which section
| Section | Sources behind it | Count |
|---|---|---|
| What happened | Anthropic’s incident disclosure and its containment write-up; the Claude Code routines documentation; Meta’s launch announcement; TechCrunch and Al Jazeera for everything OpenAI said; the Grok Bot reporting; the benchmark write-ups. | 8 |
| Timeline | Anthropic’s two posts carry five of the twelve entries; Meta’s launch and the September additions carry two; Al Jazeera and TechCrunch carry the last two; the routines documentation carries its own launch. | 7 |
| The argument | The benchmark write-ups for every figure and for the standard error; Anthropic’s two posts for the three failures; Al Jazeera for the quotations from OpenAI’s head of safety systems. | 4 |
| What others add | Meta’s launch announcement and the routines documentation describe two of the four designs at source; the other two are carried by TechCrunch and by the Grok Bot reporting. CNN carries Amazon’s refusal. | 5 |
| Conclusion | All ten. Each card states a fit rather than a ranking, and every trade-off named in one is a fact established earlier on the page. | 10 |
What the evidence base looks like
- 10Sources used on the page
- 8Read in full at their own pages, rather than through a summary
- 2Of the four vendors whose own pages could not be read at all — one refused every request, one returned no text
- 4That are a vendor describing its own system, including two disclosing its own failures
| Function | How many | How they were read |
|---|---|---|
| A vendor on itself | 4 | All four in full at their own pages. Two of them are a company publishing its own security failures in detail — a credential exfiltration its red team achieved 24 times in 25, and three real organisations compromised during evaluations — which is evidence against the publisher’s own interest and is weighted accordingly. |
| Press | 4 | All four in full at their own pages, each with a named byline and a timestamp. Three of them are the only route to what OpenAI said, because OpenAI’s own pages refused every request; one of those carries direct quotations from its head of safety systems. |
| Benchmark analysis | 1 | Figures located across published benchmark write-ups and cross-checked against each other. Two of them disagree on one score, and the page prints both rather than choosing. The vendor’s own announcement could not be opened to check any of them against it. |
| Reporting standing in for documentation | 1 | Everything about one of the four products. Its documentation returns an identical page for every address, with the text assembled in the reader’s browser, so none of it could be read here. The write-ups quoting that documentation disagree with each other on the single most consequential point, and the page says so instead of picking a side quietly. |
Reliable
Reliable, but with an interestWorth believing and reliableWeak on both, and used with careHonest, but thin
Worth believing