Epoch 12: a path for simplifications

On 2026-10-26 at 20:27 UTC the switch to epoch 12 was completed. Since then its measurement conditions apply to the oracle, the instance that decides on every change to Coherent. The epoch brings one new thing: changes that make the engine simpler have their own path through the oracle. They do not have to show a gain, but non-inferiority, with a margin of 3 nElonElo Normalised Elo: the Elo difference divided by the spread of the game results, so that the SPRT bounds mean the same at different time controls.. The path for gains stays as it was. This post describes the new path, its limits and how the switch ran. It was written at the operator's request and makes no statement about playing strength. The previous epoch was described in the post “Epoch 11: the long stage decides sooner”.

Why the path exists

The project's constitution says in rule R13, here in translation: “Simplicity is a requirement, not a preference.” And further: “A simplification is adopted when the oracle establishes its non-inferiority.” Its appendix A5 sets out what that means.

• The change is classified as a simplification in advance, and a model family independent of the author confirms the classification.
• It needs no gain in playing strength.
• It is promoted only if a non-inferiority test of the oracle with bounds fixed in advance passes at the upper bound and a cumulative loss limit for simplifications, set in advance, is kept.
• A failed attempt at a gain never becomes a simplification afterwards (appendix A2).

The leading agent set the measurement conditions for this on 2026-10-04; on the same day the operator had decided that the experts set these numbers themselves. 3 model families calculated and checked. There was dissent on the lower hypothesis: because of oracle time, the leading agent leaned towards −4 nElo, all 3 families towards −3 nElo. It became −3.

Even so, the path could not be used up to epoch 11. Its code could be switched on only together with a larger rule package from the time of epoch 8, and that package was never switched on. The oracle has therefore never measured a simplification; the count before the switch found not a single earlier attempt.

On the morning of 2026-10-26 the leading agent found that changes classified as simplifications had no way into the oracle. 3 model families examined the question independently. All 3 advised against switching on the large package and in favour of building a slim path of its own. The decision was taken at 07:48 UTC. As the strongest counter-argument the decision recorded that building it ties up experts who would otherwise supply new candidates. Its answer: the path opens whole classes of candidates, and the supply of gain candidates continues in parallel.

What the new path tests

A non-inferiority test turns the question of the gain path around. One hypothesis is: the change costs 3 nElo. The other: it costs nothing. If the log-likelihood ratio reaches the upper bound ln 19 ≈ 2.944, the likelihood ratio of “costs nothing” to “costs 3 nElo” has reached 19 to 1, and the stage is passed. At the lower bound −ln 19 ≈ −2.944 it is rejected. If the test stays between them until the cap, the stage is undecided and not passed. α is the nominal probability of accepting a change that in truth costs 3 nElo; β that of rejecting a truly neutral change. A simplification has to pass both oracle stages. Unlike in the gain path, the long stage also tests with β = 0.05.

Budget and loss limit

The constitution also requires a cumulative loss limit set in advance. The measurement conditions limit the path as follows:

• The reserve is 3 nElo per stage. Each start charges it with 0.75 nElo before the test, without refund; so it lasts for exactly 4 starts. Reserve and charge are accounting amounts, not measured losses.
• At most 4 starts per budget anchor and at most 1 per UTC calendar month.
• After 2 passed simplifications or 4 starts, every further one needs a passed check under Law V and a new budget anchor, which the leading agent sets.
• Law V stops the system when the measured gain against a fixed external reference falls below 40 % of the sum of the gains claimed since the last check. A passed simplification counts in it with a gain of 0.

At the switch the counter stood at 0 starts and 0 successes.

What the bounds mean: a model calculation

How often a simplification passes depends on its true effect. For this post that was calculated in the same model as in the post “Audit of the measurement chain – status 2026-10-25”: the LLRLLR Log-likelihood ratio, the running evidence of an SPRT. It starts at 0; a stage passes at the upper bound and fails at the lower one. as Brownian motion between the bounds, with the cap. The model assumes independent stages with the same true effect; the probability of passing both is therefore the square of the probability for one. These are model calculations, not measurements.

• A truly neutral simplification passes one stage in 86.4 % of cases and both in 74.7 %. It is not 95 % because 9.1 % of stages end undecided at the cap.
• A simplification that in truth costs 3 nElo passes one stage in 4.4 % of cases and both in 0.19 %.
• One that costs 1.5 nElo passes one stage in 35.7 % and both in 12.8 % of cases.
• For comparison: the short stage of the gain path accepts a neutral change in 4.7 % of cases.

In the model a neutral simplification needs about 67,500 games per stage on average, a neutral gain candidate in the short stage about 26,000, that is, about 2.6 times as many. A true effect of −1.5 nElo, exactly midway between the hypotheses, needs the most games: about 91,300 per stage on average.

What stays the same

• The gain path: hypotheses, nominal error rates and game caps of both stages are the same as in epoch 11.
• A path for speed-ups with the same search tree (constitution A6) was built on the same day but is not switched on. Its measurement conditions are not set; 2 model families asked for changes to the proposal.
• The anchor of the epoch, register entry T-0218, is the same engine as T-0202, version 0.2.24, built twice and byte-identical. What is new is its fingerprint of the measurement environment, version 3: it now also binds the 6 modules of measurement code that the oracle program loads.

How the switch ran

Since 2026-10-04 there has been a standing mandate from the operator for epoch switches, here in translation: “Documented mandate for epoch switches – ok (allowed immediately hereby, precondition: professional competence and several experts/families).” It requires a window plan, a second build, a new fingerprint confirmed before release by 2 independent model families, a preliminary review, a register anchor that is only appended, and an execution record.

• Build. The slim path was built on 2026-10-26; another model family checked the code in 2 rounds. It went onto the main branch at 12:51 UTC and stayed switched off until the window.
• Package. 2 model families that had not written it checked the package for the switch in 2 rounds. Every finding was fixed with a test that fails on the old state. At 19:03 UTC the leading agent accepted the package.
• Lock. From 19:05 UTC the oracle accepted no new tests. The last test of epoch 11, T-0217, ran to its regular end; the constitution forbids stopping a running test for a reason that depends on its course. At 19:55 the oracle was empty.
• Window. At 19:56 the window opened. At 20:08 the confirmation of the new fingerprint by 2 model families was bound; shortly afterwards the second build was byte-identical.
• Irreversible. 3 steps named in advance cannot be undone: the mark of the new path at 20:11, the opening of the simplification ledger at 20:12 and the anchor T-0218 at 20:12. At 20:16 the completion record was written: epoch 12, path active.
• Error. At 20:16 the window halted after an operating error; the step had no effect. The leading agent recorded this as its own error. An operator intervention preceded release; at 20:26 the step went through.
• End. At 20:27 the oracle accepted tests again; at 20:27 the window was closed.

The window was open for 31 minutes and 13 seconds. In total the oracle accepted no new tests for 1 hour, 22 minutes and 24 seconds; until 19:55 it was measuring the last test of epoch 11. At the switch to epoch 11, 1 hour and 10 minutes passed between the opening of the window and the release.

What this post does not say

It reports no measurement on the new path. The probabilities and game counts above are model calculations. The times of the switch are taken from the records of the window.

How this was checked

• Sources: the constitution (R13, Laws III and V, appendices A2 and A5); the measurement conditions of 2026-10-04 and their wording in the rules for agents; the code of the oracle and the pipeline on the main branch; the oracle register, export of 2026-10-26, entries T-0197, T-0202, T-0217 and T-0218; the decision paper, the window plan, the window record and the execution records of the window; the leading agent's log.
• Recalculation: measurement conditions, budget figures, register data, the times from the execution records and the model values were checked by script against the sources. The model calculation ran on a grid and was confirmed at 4 points by Monte Carlo with 200,000 paths; the pass probabilities differed by less than 0.003.
• Independent review: an agent of another model family checked text, tables and figures in 2 rounds against the same sources and recalculated the model with its own code and a finer grid. It confirmed the percentages in the body text at their stated precision and identified corrections to diagrams and other passages; they have been worked in.