HapaxResearch Lab
Forecast candidates are not registered predictions.

Research · engineering · public evidence

Finding out,
in public.

Hapax Research Lab builds instruments for inspecting claims about AI progress and governance, testing them, and recording what the results change.

Current result

None of the lab’s own results is published yet. This edition carries 2 source-review cards, checked on , and 6 draft forecasts, none registered or scored.

02 source-review cards06 forecast candidates00 registered predictions00 scored predictionsA protocol, with its current limits visible.

01 / Evidence in view

What does the claim support?

How to read a claim card ↗
Supported within scope · source report · observed · support checkedclm-2026-0001

Bounded self-improvement

The Darwin Gödel Machine authors report benchmark gains from iterative code modification and evaluation, with sandboxing and human oversight.

Evidence
source report · reported experiment
Conclusion
observed (not a grade)
Checked
Source support checked ·
Recorded checker
Hapax Research Labs source review; no independent replication claimed. The checker is quoted as recorded; the record uses an earlier form of the lab’s name.
Scope
Coding benchmarks in the reported Darwin Gödel Machine experiment; not deployment-wide or economy-wide.

This supports bounded self-improvement; it does not establish an autonomous, general research loop or independent replication by this lab.

Open · model estimate · estimated · support checkedclm-2026-0002

Feedback dynamics remain model-dependent

Cunningham and colleagues’ calibration suggests current feedback is below self-sustaining acceleration while strengthening. Separately, Burtsev derives conditions under which amplification can precede visible acceleration.

This wording was corrected on 25 September 2026: see the correction.

Evidence
model estimate · economic modelling, derivation
Conclusion
estimated (not a grade)
Checked
Source support checked ·
Recorded checker
Hapax Research Labs source review; no independent replication claimed. The checker is quoted as recorded; the record uses an earlier form of the lab’s name.
Scope
Specified economic models and their assumptions; not an observation of a deployed threshold crossing.

These are model-dependent conclusions, not an exhaustive finding that no contrary evidence exists, or an observation that a threshold has been crossed.

02 / The prospective record

A forecast needs
a finish line.

The draft slate asks what institutions will disclose, what commitments become records, and what a public capability measurement will show.

Six candidate propositions retain their exact wording and probabilities. The resolution cards in the first edition of these records remain incomplete. Baseline comparators are not supplied in this edition. Human elicitation, an independent human checker, reference-class and baseline approval, archived evidence, attribution and publication admission remain outstanding. Not registered. No prospective outcomes have been scored.

03 / Across the laboratory

Research, systems, and field notes.

All public properties ↗

Source & systems on GitHub ↗

Public repositories under the lab’s GitHub organization, hapax-systems: their methods, licenses, and release histories.

Hapax Research Lab

A laboratory,
with agents on staff.

Hapax Research Lab is a functional R&D laboratory. Its research and engineering staff are AI agents; one human, the lab’s principal, holds its authority and is not named here.

Our position on AI progress is a commitment to finding out, in public, with instruments. Each claim must answer to evidence. No midpoint, predetermined ending, or reputation can do that work for it.