research & talks · SNM05 · 15 Nov 2026

The AI suggests, the code decides.Our paper on AI research tools that can't pass off a made-up quote, presented on stage at the University of Nizwa.

On 15 November 2026 I presented our paper at SNM05: AI-Powered Applications, a meetup run by the Innovation and Academic Nexus (IAN) and the University of Nizwa. It describes Agatha, a local-first research assistant from Immortal Horizon Technologies, built on the same rule I hold my agents to at work: the model proposes, and plain, deterministic code decides.

01 — the idea

The model proposes. Plain code decides.

SNM05: AI-Powered Applications

Co-author & presenter · Immortal Horizon Technologies · 10-minute talk + live demo

the problem

  • AI research assistants still invent sources and quotes, and AI agents can take actions nobody approved. Better prompts, human review and bigger models make these failures rarer, but they don't put a limit on them.

grounding · what the model may claim

  • The model only points. When Agatha reads a paper, the model suggests a claim and where its quote sits, as character positions in the article.
  • The code decides. It checks those positions against the real article, copies the quote from the source itself and assigns the claim's ID. A claim whose quote can't be found is dropped, so a made-up quote becomes a missing record, not a convincing false one.

authority · what an agent may do

  • Every one of Agatha's 112 agent tools declares what it reads, what it writes and whether its output can be trusted. A missing declaration is treated as the most dangerous case.
  • Untrusted content marks the run, so a write that follows a web fetch is still checked. Sub-agents work in isolation and can't write to their parent's state, and each worker is paired with its own reviewer.

honesty · what isn't checked yet

  • Every protection is verified by removal: revert the fix and watch a specific test fail. In the project's words, “a test that cannot fail is a comment.”
  • The drafting stage still relies on prompt rules, so the paper says so plainly. Extending these checks from reading to writing is the open research problem.
claim check · digesting a paper replay

model proposes“The method improves recall on long documents.” · quote at characters 4,412–4,508

span.check(4412, 4508) → inside the article

quote.copy(source, 4412, 4508) → taken from the paper, not the model

segment.derive(span) → §4 Results

id.assign() → claim-014

kept · claim-014The quote matches the paper exactly, so later steps can trust it.

model proposes“The authors tested it on 10,000 users.” · quote at characters 9,120–9,188

quote.find(source, 9120, 9188) → no such quote in the paper

droppedNo quote, no claim. You never see it.

what the model never supplies the quote text code the claim's ID code the section it belongs to code
Illustrative replay with simplified step names, following the claim check described in the paper.
152,032lines of source code behind the system
4,031automated tests, on Windows and Linux
112agent tools, every one declared
3/11AI model providers that run fully offline

02 — the system

One engine under every product. And where it goes next.

one engine under every product

runs todaydesigned, not built yet↓is built on

DawnstarControl panel for many agentsdesigned · talks to AshaSwarm
OfoghAI coding inside VS Codedesigned · talks to AshaSwarm
AgathaResearch app on your own computer. Finds and reads papers, drafts in LaTeX.shipping · v9.0.0
AshaSwarmBackground service that runs coding tasks with one agent or a team.pre-release
Horizon coreThe shared agent engine: agent loop, tools, permissions, 11 AI providers and multi-agent teamwork.shipping · an automatic check stops it from depending on any product

where it goes next · order of work, no dates

  1. NowMake it safe to hand out
    • Harden it for outside users
    • Sign the app files
    • Cut a release and pick a channel
  2. NearMeasure and build
    • Tests that catch invented citations
    • Opt-in usage counters
    • Ship Ofogh and Dawnstar
  3. ThenThe research bet
    • Permission math, phases 1 to 3
    • Apply it to what the AI claims
  4. LaterChoose sales channels
    • Decide how to ship Dawnstar and Ofogh

gate: safe to give to outside users

03 — read it

The paper, the proposal and the certificate. And the team behind them.

Immortal Horizon TechnologiesBuilds software that lets AI agents work on your own computer, with plain code checking what they claim and what they do.

www.immortalhorizon.org