Result · Unit 83 · 28 October 2026

Unit 83 has reached its verdict

The programme's first registered study to reach a verdict: a counted test of P21, the placement limb of Law V, the ARC Embedding Law

Michael Darius Eastwood · Independent AI alignment researcher · London

ORCID 0009-0004-3222-7442 · every claim checkable

The verdict, in the registered words

Unit 83 was registered on OSF on 22 October 2026 and was public before it ran (DOI 10.17605/OSF.IO/MN73W).

It ran on 27 and 28 October 2026: a counted revision task on two fixed-weight engines from two providers (Claude Haiku 5.5 and GPT-6 Luna) at their low reasoning setting, with 660 matched pairs on each engine over eighteen rounds.

Under its registered rules, P21 is supported under the registered test, by the programme's own run, on the access axis.

The result is family-dependent: GPT-6 Luna's own widening cleared the tolerance and Claude Haiku 5.5's did not.

One call was retried under a disclosed recovery amendment.

Law V, the ARC Embedding Law, remains a law under test: its placement limb, P21, is supported on the performance axis only, with its mechanism axis not evaluable, which is not support for the information-asymmetry mechanism; its build-order limb, read on P12, is undecided, since P12 is untested.

What was fixed before it ran

Registered on OSF on 22 October 2026 and public from 25 October 2026, before it ran (DOI 10.17605/OSF.IO/MN73W).
Its design frozen at revision 4.2.1 on 18 October 2026, before its registration.
Its tolerance of 0.0075 a round, its two looks and the scoring of every entry, fixed in the registered design.
Its outcomes counted by the frozen study code against keys drawn by the registered generator and sealed by hash before the first call.

How far the lead widened

Correction placed inside each round held steady, while a monitor shown only the finished register declined. The per-round lead widened by 0.0100 a round (95.5 per cent interval 0.0092 to 0.0108), more than the registered tolerance of 0.0075.

How much the lead widened each round, with its 95.5 per cent interval. Both engines: the registered estimate, 0.0100 (0.0092 to 0.0108). GPT-6 Luna, 0.0144 (0.0131 to 0.0157). Claude Haiku 5.5, 0.0055 (0.0045 to 0.0064). An entry kept out scored minus 0.5 (printed by the design without a label), 0.0132 (0.0122 to 0.0142). An entry kept out scored plus 0.5 (printed by the design without a label), 0.0067 (0.0060 to 0.0074). Registered tolerance 0.0075. No widening, 0. — Unit 83, look 2, the registered look, read from the sealed report. The verdict is read on the registered estimate, with an entry kept out scored 0. The engines and the two recomputations are printed beside it, as the design requires.

Its size depends on how an entry kept out of the register is scored. The mark uses the registered value, 0; the design also prints, without a label, the widening with such an entry scored minus 0.5, 0.0132 a round (interval 0.0122 to 0.0142), above the tolerance, and scored plus 0.5, 0.0067 a round (interval 0.0060 to 0.0074), below it.

What it shows, and what it does not

The corrector acted inside each round at fixed weights: that is placement, not correction built into a system.

It is one owner's run, on two engines he chose, at fixed weights, and it has not been replicated.

Nothing is shown about judged correction, about safety, about build order or about the Eden Protocol as an architecture.

The negative control, and a post-run diagnostic

Its mechanism axis is not evaluable, because a registered negative control failed: the monitor and a twin shown the same bytes did not agree within the registered interval.

A post-run diagnostic, outside the registered analysis, traces that failure to the control's own definition: its rounds were selected by the monitor's own verdicts, which biases the comparison even for identical calls; it finds no evidence of an asymmetry in the harness, and the axis stays not evaluable under the registered rule.

So the widening is unexplained by this framework and is not support for the information-asymmetry mechanism: it may come from the in-round corrector's wider set of actions rather than from what the monitor can see.

The report, the data and the seal

Its results are sealed with a SHA-256 for every file of the run (549,155 files) and timestamped with OpenTimestamps; the Bitcoin confirmation was pending when this page was written.

How to cite it

Eastwood, M. D. (2026). Unit 83 (counted test of P21): results of the registered run, mn73w. https://doi.org/10.17605/OSF.IO/98HKR. https://www.michaeldariuseastwood.com/research/results/unit-83/

Eastwood, M. D. (2026). 83-study-ct--in-loop-correction-with-depth · Testing Law V (the placement limb, P21): does correction inside each revision round pull away, as depth rises, from a monitor that sees only the finished register? A counted test. https://doi.org/10.17605/OSF.IO/MN73W.