# Evidence Queue | AI Wars of 2026

> Research collected for AI Wars of 2026 but not yet used, and the reason each item is still waiting.

Canonical URL: https://ai-wars.correax.com/evidence-queue

## Evidence Queue

> **This page shows what this book has found but has not yet used, and why.**

Most publications show you what they included. This one also shows what it
collected and set aside, because the reasons are often more informative than the
additions.

### How evidence reaches a chapter

Research arrives here through four stages, and most candidates stop before the
end.

1. **Collection.** Automated producers search company newsrooms, release notes,
   investor pages, regulators, standards bodies, court dockets and independent
   reporting, in English and in the original language of the source. Each files a
   dated report listing what it found.
2. **Decision.** Every finding in every report is given a terminal state:
   applied, rejected with a reason, deferred with a named trigger, or superseded.
   Nothing is left undecided, and nothing is dropped silently.
3. **Verification.** Before a source enters the registry, the page it cites is
   read in full and the claim is checked against it. This stage removes more
   than any other.
4. **Citation.** A verified source only becomes part of the book when a chapter
   cites it. A record in the registry that no chapter uses is not yet evidence
   for anything.

### Why verification removes so much

A collector's summary is a lead, not evidence. Of the first nine sources checked
against their original pages, four carried a specific the cited page did not
state: a penalty amount, an article number, a publication date. In every case the
source was real and the surrounding claim was reasonable. Only the detail was
unsupported.

That is the failure a reader cannot detect, because the citation resolves and
the page opens. It is the reason this book reads the source before recording it,
and the reason the queue below moves more slowly than the collection does.

### What the states mean

A candidate in this queue is not a claim this book doubts. In most cases it is a
claim nobody here has checked yet. The state says which.

### Reading the tables

Each row is one finding. It names what the finding is about, where it came from
and when, and the specific thing that has to happen before it can enter the book.
That last column is the useful one: it is the actual bar, item by item, rather
than a general statement of standards.

The tables are generated from the intake records, so their counts and contents
cannot drift apart, and the build fails if the page falls behind. The subject
descriptions are written here rather than taken from the collector, because a
collector's heading paraphrases the claim, and a claim nobody has checked should
not be asserted on this page. Naming the subject is not the same as asserting it
happened.

### Where the evidence stands

| Stage | Count |
| --- | --- |
| Findings collected and given a decision | 76 |
| Judged ready for editorial review | 37 |
| Verified and recorded in the source registry | 13 |
| Waiting for one of the stages below | 24 |
| Marked unconfirmed by the collector itself | 28 |
| Held on watch for future relevance | 11 |

### Waiting to be read against the original page (16)

Collected and decided, but no editor has yet read the page it cites and checked the claim against it. This stage removes more than any other.

| Finding | Source | What it needs before it can enter the book |
| --- | --- | --- |
| AWS separating what an agent remembers long-term from what it saw in a session | aws.amazon.com, 2026-09-08 | A read of the announcement to establish what controls the separation. Agent memory is a Part III question and the mechanism matters more than the feature. |
| AWS letting customers debug which documents an agent is permitted to read | aws.amazon.com, 2026-09-09 | A read of the announcement. Part III records that governed data access outside Microsoft's ecosystem is thinly evidenced, so this is wanted, which is a reason for more care rather than less. |
| AWS offering automated modernisation of .NET applications | aws.amazon.com, 2026-09-09 | A read of the announcement. Availability of a migration tool says nothing about whether migrations succeed, so this can only ever be recorded as an offer. |
| AWS generating tests as part of automated code modernisation | aws.amazon.com, 2026-09-10 | A read of the announcement. Generated tests raise an obvious question the announcement may not answer: whether anyone checks that the tests are correct. |
| Microsoft on governing what AI agents cost to run | azure.microsoft.com, 2026-09-10 | A read of the post to separate what Microsoft is offering from what it claims the offering achieves. A vendor's account of its own cost controls is not evidence that customers save money. |
| Google putting Gemini on Windows as an application with a system-wide shortcut | blog.google, 2026-09-10 | A read of the post. Placing an assistant on a rival's operating system is a distribution move, and the book distinguishes reach that is offered from reach that is used. |
| NVIDIA on a target for AI data-centre capacity in Australia | nvidianews.nvidia.com, 2026-09-10 | A read of the release to separate the capacity target from capacity that exists. A target for 2027 is an intention, and the book does not record intentions as infrastructure. |
| Anthropic's September threat report on misuse of its own models | anthropic.com, 2026-09-10, found by two collectors | An editor to read the report and check every count against it. Anthropic is describing misuse of its own product, so the figures are the company's own attributions and cannot support an industry-wide rate. |
| Reuters on a chip startup adopting NVIDIA's interconnect technology | reuters.com, 2026-09-10 | A statement from D-Matrix or NVIDIA. A competitor choosing a rival's interconnect bears on the infrastructure contest and should not rest on reporting alone. |
| Reuters on US Senate scrutiny of how OpenAI handled an incident | reuters.com, 2026-09-10 | The Senate letters themselves. This is reporting on another outlet's reporting, which is two steps from the document, and the document is public. |
| Reuters on the EU cybersecurity agency obtaining access to frontier models for testing | reuters.com, 2026-09-10 | The European Commission statement the report describes. A regulator gaining test access is a governance development, and the governing body has published its own account. |
| Visa, Mastercard and Ant International on a shared way to identify AI agents in payments | reuters.com, 2026-09-10 | A statement from at least one of the three companies. An agreement between rivals on identity is materially interesting, which is exactly why it should rest on a primary source rather than on reporting. |
| Reuters on DeepSeek releasing a new model | reuters.com, 2026-09-10 | The primary release material from DeepSeek. This is secondary reporting, and the book records a release from the record of the release wherever one exists. |
| AWS extending what an agent can search from transcribed text to video and images directly | aws.amazon.com, 2026-09-11 | A read of the announcement to establish which content types are covered. This bears on how much context an agent can reach, which is a live question in Part III. |
| AWS putting agent-led incident investigation into the channel where on-call engineers already work | aws.amazon.com, 2026-09-11 | A read of the announcement to establish what the agent may do without approval. The control boundary is the part that matters, and it is not visible from the summary. |
| AWS on reducing the delay before a cached model starts serving | aws.amazon.com, 2026-09-11 | Confirmation of the stated figures and the model sizes they apply to. Performance numbers published by the vendor running the test are a claim about the product, not an independent measurement. |

### Waiting for a reading in the original language (5)

The primary source is not in English. Recording it needs the original passage, an aligned translation, and a note on any term that does not translate cleanly.

| Finding | Source | What it needs before it can enter the book |
| --- | --- | --- |
| France's data-protection regulator updating a traceability tool for openly published models | cnil.fr, 2026-08-26 | A reading of the French original. Regulator guidance is written precisely and the precision is the point. |
| China's foreign ministry responding to US allegations about copying AI models | mfa.gov.cn, 2026-09-09 | A reading of the Chinese original. The book holds this rebuttal but has not been able to retrieve the US advisory it answers, so recording one side alone would misrepresent the exchange. |
| Chinese thresholds for moving personal data across borders | cac.gov.cn, 2026-09-11 | A reading of the Chinese original. Thresholds are numbers in a legal instrument and must be read from the instrument, not from a description of it. |
| A Chinese state plan directing AI models into named rural government services | cac.gov.cn, 2026-09-11 | A reading of the Chinese original with the relevant passage quoted and translated. A state directive naming specific workflows is precise, and precision is lost in summary. |
| A Chinese AI chip company listing on the Shanghai STAR Market | cnfin.com, 2026-09-11 | A reading of the Chinese original, and preferably the exchange filing behind it. Financial facts should come from the filing rather than from coverage of it. |

### Outside the window that was searched (1)

The source predates the interval its report covered and was found incidentally, so it is held rather than recorded against a window it does not belong to.

| Finding | Source | What it needs before it can enter the book |
| --- | --- | --- |
| An APEC forum statement on AI, published in July | cac.gov.cn, 2026-07-24 | Nothing beyond a later pass. It was found while establishing regional coverage back to January and belongs to that earlier period rather than to the days these reports cover. |

### Already recorded from another page (1)

The same development is in the book from a different source.

| Finding | Source | What it needs before it can enter the book |
| --- | --- | --- |
| The EU Cyber Resilience Act reporting duties taking effect | digital-strategy.ec.europa.eu, date not recorded | Nothing. The book records this from the Commission's reporting-obligations page. Both collectors found this event independently at two different addresses, which is the clearest corroboration the intake has produced. |

### Not a development (1)

The finding describes what the research process could reach, not something that happened in the industry.

| Finding | Source | What it needs before it can enter the book |
| --- | --- | --- |
| China's algorithm filing registry becoming readable to the research process | beian.cac.gov.cn, date not recorded | Nothing. This records what the research could reach, not something that happened in the industry. The surface status is kept in the internal records where it belongs. |

### Marked unconfirmed by the collector

Collectors label a finding when they cannot confirm it themselves. 28 findings carry that label and are not eligible for the registry on their own. They are kept because a later, independent source may confirm one of them, and because a record of what could not be confirmed is part of an honest account.