Products
Products are reviewed from their official documentation, fetched and pinned as described in How the survey used HALO, and from their code when they release it. Six products appear in the survey’s channel and evidence table.
| System | S→M | M→A | A→S | M→S | A→M | Rec. | Out. | Attr. | Rev. | Use |
|---|---|---|---|---|---|---|---|---|---|---|
| Wearable | ||||||||||
| Omicode | ● | ● | ● | ● | ● | ● | · | · | · | · |
| Bee | ● | ● | ◐ | ◐ | ● | ● | ● | · | · | · |
| Physical | ||||||||||
| ElliQ | ● | ● | ● | ◐ | ◐ | · | · | · | · | · |
| Digital | ||||||||||
| Dot | ● | ● | ● | ● | ● | ● | ● | ● | · | · |
| Muse | ● | ● | ● | ◐ | ● | ● | · | · | · | · |
| EAIAcode | ● | ● | ● | ◐ | ● | ● | ● | ● | ● | · |
| Reported or stated (of 6) | 6 | 6 | 5 | 2 | 5 | 5 | 3 | 2 | 1 | 0 |
Channels: ● reported, ◐ insufficient to decide, ○ explicitly absent. Evidence steps: ● stated in the source, · not stated.
What the table shows
Section titled “What the table shows”- Every product documents memory. All six report S→M and M→A.
- Documentation rarely says how memory shapes what a product notices. M→S is insufficient for four of the six.
- Few products document their own feedback loop. Five keep records of their actions, three keep outcomes, and two link outcomes to decisions. EAIA is the only one that states a revision: it rewrites its prompts from the user’s edits. No product’s sources report a run under a revised mechanism.
Run mode: EAIA
Section titled “Run mode: EAIA”EAIA is an L4 email assistant whose code is released, so run mode can test its evidence chain. It revises its prompts from edits to its drafts through a step called reflection. In the survey’s test, a simulated user with 5 unstated preferences edited drafts for 5 emails, and 4 new emails then checked whether the preferences carried forward. EAIA ran on GPT-6-Astra, GPT-6.1-Sol, and Claude Opus 5.5, with 2 runs per condition, and a test counted as changing the output when the unmodified run was preferred in at least 5 of 8 comparisons for at least one model.
| Cut during learning | Result |
|---|---|
| record: reflection loses the assistant’s questions and drafts | effects less consistent across models |
| outcome: the user’s feedback is withheld from reflection | all 3 models lost the required closing sign-off, and partnership replies no longer mentioned the designated colleague |
| attribution: feedback is paired with the previous email’s record | effects less consistent across models |
| revision and later use: revised prompts are generated but not stored | the same losses as withholding feedback |
Read mode left later use unmarked, because EAIA’s sources report no run that demonstrates it. Run mode supplied that evidence: replies in the reproduced configuration depend on stored revisions.
The plugin on a live product
Section titled “The plugin on a live product”Most products release no code, so run mode cannot cut inside them. The plugin follows a running product instead. It has been connected to three of the survey’s products so far: a ChatGPT dot, Grok Bot, and Claude Cowork. Products in the survey shows how HALO can attach to each of the others.