cd /news/ai-agents/four-reviewers-told-me-the-one-thing… · home topics ai-agents article
[ARTICLE · art-115481] src=dev.to ↗ pub= topic=ai-agents verified=true sentiment=· neutral

Four reviewers told me the one thing I couldn't fix by myself

A developer's open-source protocol for AI-device interaction was rejected by IEEE WF-IoT 2026 despite positive reviews, highlighting the challenge of self-evaluation. The reviewers praised the architecture but noted the lack of independent annotators, a gap the developer cannot close alone. The developer has since expanded the evaluation corpus and fixed bugs found through self-testing.

read4 min views1 publishedAug 30, 2026

What happens when you submit a solo open-source project to academic peer review: the reviews were good, the paper was rejected, and the gap they found was the one I had no way to close on my own.

I've been building DoSync, an open-source protocol that sits between AI agents and physical devices — I

wrote about

[why it exists here](https://dev.to/giulianiregspec/why-your-smart-home-ai-cant-open-the-door-in-an-emergency-and-how-to-fix-it-5d5i)

if you want the background.

I tested it the only way I could: on my own hardware, against my own

expectations. Then I sent it to IEEE WF-IoT 2026, because I wanted it examined by

people who weren't me.

Five reviews came back. Four were substantive. It was rejected.

And here's the part that took me longest to understand: the reviews were good. Two scored relevance 5 out of 5. One recommended minor revisions. The

Rejected anyway. Not because the idea was wrong. Because of this, from the most

detailed review:

The evaluation is small and self-judged: 15 scenarios with ground truth

defined by the author alone.

That sentence is correct, and no amount of rewriting section 4 was going to

change it.

It took me a few days to notice that the rejection was the thing I'd submitted

for.

I didn't send the paper hoping for a stamp. I sent it because after months of

being the only person who had ever evaluated this, I had no way to tell the

difference between "this works" and "this works the way I expect it to." Four

strangers read it carefully and told me. That is not a consolation prize — it is

the entire reason to submit.

What they said, in summary: the architecture holds up, and the evidence doesn't

reach as far as the claims. One reviewer went further and put the knife somewhere

specific. My central claim is that the project removes the manual burden of

writing device-by-device automation rules. Their objection:

if resolution quality depends on well-curated tags, the manual configuration

burden has not been eliminated, it has moved from rules to manifest curation

I've had months to find a counter-argument. I don't have one yet.

The request that came up in almost every review was for independent annotators.

Not more scenarios — different people. Someone other than me who had read a

scenario, looked at the devices available, and formed their own opinion about

which ones should respond.

I couldn't produce one. Every judgement about what the resolver should pick —

in all fifteen scenarios, in the paper and in the benchmark files — was mine.

When I disagreed with the resolver, I was the one who had decided what the right

answer was in the first place.

Everything else on the reviewers' list, I could work on alone, and did. The

evaluation used to run on a single home registry, which is a poor way to support

a claim of being domain-agnostic; there's now a

multi-domain corpus covering industrial and clinical registries, published so the numbers are

reproducible by someone who isn't me. The precision/recall conventions that made

two rows of a results table mathematically impossible are defined and the bug is

traced. There's a sensitivity analysis on the scoring weights.

None of that required anyone else. The annotators do. That asymmetry turned out

to be the most useful thing in the whole review packet.

Since the reviews, I've been doing the closest thing to third-party testing that

one person can do: installing my own project on machines that had never run it,

following my own published instructions, on hardware that owed me nothing.

It found more than a dozen defects in a week, on a codebase with about a thousand

passing tests. My favourite: device discovery had been silently dead in

production for days, because a dependency I'd added shipped an event loop that

declares a method and raises when you call it. My check was hasattr

. It

returned True

. Every test passed and nothing was ever discovered.

Not one of those was findable from the inside. My test suite ran on one operating

system, one event loop, one machine that already had everything configured — it

was measuring a world only I lived in. Which is, word for word, what the reviewer

said about my evaluation.

That's the part I'd pass on to anyone building something alone: the hardest work isn't the next feature. It's getting the thing into the hands of someone who doesn't already know how it's supposed to behave — including your own hands, on

The protocol is open, the corpus is public, and the one thing it still needs is

an opinion that isn't mine.

**GitHub:** [github.com/giulianireg-spec/dosync-protocol](https://github.com/giulianireg-spec/dosync-protocol)

**Web:** [https://dosync.dev/](https://dosync.dev/)

License: Apache 2.0

DoSync — the semantic layer between AI agents and physical systems.

(The reviews are summarised faithfully and the reviewers stay anonymous, as they should.)

── more in #ai-agents 4 stories · sorted by recency
── more on @dosync 3 stories trending now
sponsored brought to you by zahid.host 4,200+ EU-deployed projects
reading about agents? ship yours in a single git push.

Run your AI side-project on zahid.host

EU-based hosting, git-push deploys, automatic HTTPS, no cold starts. Free tier with a custom domain — perfect for shipping the agent you just read about.

$git push zahid main
Live at https://your-agent.zahid.host
Get free account → Pricing
from €0/mo · no card required
LIVE [news/four-reviewers-told-…] indexed:0 read:4min 2026-08-30 ·