How the survey used HALO

The survey Always-On Personal Intelligence applied HALO to 40 works: 34 research systems and 6 products. Every work went through read mode. Run mode tested one research system, AiGet, and one product with released code, EAIA. The results are on two pages, because research systems and products are reviewed from different sources:

• Research systems: papers and their released code.
• Products: product documentation, and code where a product releases it.

The tables on these pages repeat the survey’s published table.

Read mode

1. 1Sourcespapers, product documentation, and code, at pinned versions
2. 2Three oversight agentsGPT-6-Astra, Claude Opus 5.5, and Claude Fable 5.1, each reading alone
3. 3A label per cellevery channel and evidence step, with a quote and a locator
4. 4Majority votetwo of three agents decide each cell
5. 5Tablechannels and evidence for 40 works

Each oversight agent reads the sources listed in the work’s case file on its own and returns one judgment per cell: a label, a verbatim quote, and a locator such as a section, a file path, or a URL. Every review must validate against halo/read/schema.json, and the majority of the three agents decides each cell. The protocol is halo/read/protocol/PROTOCOL.md, and Read mode shows how to run it.

A product’s sources are its official pages, fetched once and pinned. The case file records each page’s address, the date it was fetched, and a SHA-256 hash of the text the agents read:

cases/dot/case.yaml

id: dot
kind: product
sources:
  - url: https://learn.chatgpt.com/docs/dots
    kind: product_documentation
    version: "Markdown documentation page fetched 2026-10-05"
    sha256: 2dd16a1beab72dac1e70eff29e79fee52d27831ef58adc56228e8bc93ecac6eb
modes: [read]

Run mode

1. 1Released codeonly the model clients are replaced, and an adapter declares the cut sites
2. 2Original runfixed inputs and one fixed backbone
3. 3Cut runsone or more channels or evidence steps removed, everything else kept
4. 4Blind judgecompares outputs without knowing which run they came from
5. 5Reportwhich cuts changed the output

Run mode needs released code. The adapter of a case runs the upstream code with only its model clients replaced, and declares where each channel or evidence step passes through that code so it can be cut at the receiving end. Run mode shows how to run a case.

Where the plugin fits

Read mode and run mode review a work once, from its sources or its code. The plugin follows an agent while it runs, which is the only way to watch a product whose code is closed. The products page lists what each product in the survey allows.