Iskra · research record

Every trial, with its verdict.

Rulesevery entry follows them

How we test

A hash before the outcome. Every forecast gets a SHA-256 hash before the world answers. Nobody corrects it afterwards. The audit of 24 October replayed 5,344 forecast–outcome pairs and 590 decisions without learning.

Six versions of the same Iskra. The same core, the same worlds, six times: planning its probe, choosing at random, with a fixed strategy, with one explanation, without memory, without the inherited discovery. An advantage counts only against them.

A closed gate. We write the criteria before the run and leave them alone after the result. When they fail, we fix the model, not the bar, and try again on new worlds. Failures stay in the record.

25 October 2026did not pass · transfer yes

The Flame: a break and a refit

From her own errors Iskra notices that her rules stopped describing the world, and at once proposes a held rule with new coefficients instead of waiting four episodes for a search. The representation made in A carries to B: in 20 of the 24 lives of two evaluations B reaches the bar in 5 episodes against 8 from zero.

Two frozen evaluations on new seeds. The first: transfer in both designs (+43% and +25%), but knowledge of A after the return failed in clip → wave, because the bank’s hazard diluted her forecast at the level of the world’s noise. The second, with a lower hazard and only questions worth a step: knowledge of A kept in both (0.98 and 1.02), transfer in wave → wave (+41%), while in clip → wave two extreme lives of twelve swung the measure. No single evaluation passed everything at once, so the Flame is not lit. The mechanism lives on the server: when the world changes, Iskra refits a rule.

25 October 2026passed · 11 of 12 pairs

A costly measurement chosen by experience

Two wave worlds, A and B, differ only in frequency. In an episode whose law she is not told, Iskra may rest for free or pay for one of six moves; only one tells A from B. Then she decides, and the right decision depends on the law.

After A and B she picks the separating move in 96% of her measurements; after A alone in 24%. Net −0.020 against −0.058 for the best fixed rule, −0.075 for random measurement and −0.078 for never measuring, better in 11 of 11 pairs. Deciding without the measurement’s result hits the target in 37% of episodes instead of 83%.

Weak spots: in a world that does not change she breaks even but measures too often. One pair dropped out because no world could be built for it.

Try it ↗

25 October 2026did not pass

A rule from parts of two earlier rules

World C responds with half of A’s saturation and half of B’s wave. After A and B the error in the first episodes of C is 0.24 of a start from zero. Controls showed where that comes from: a session that makes no rule in C has the same error. It is a mixture of old rules, not a new one.

The rule she makes in C is 32 times more accurate than that mixture and 2.7 times more than a rule from zero, but not more than a rule made after as much experience of B alone (1.15). Composition from parts is not confirmed.

25 October 2026passed · developmental comparison

One life through twelve worlds

Until now every world on the server started from the same published state. Now each starts from the chapter closed at the end of the previous one, with a bank of separate rules and the retirement of the rule that has not led for the longest.

Over 12 worlds of the rotation one life has 90% less error than starting from zero, 85% in the openings of worlds. Without the bank one life did harm: after a wave world the saturation world had 5.4 times the error. Since 25 October the server lives one life; its chronicle is on the home page.

24 October 2026passed · 10 of 10

Several explanations, third evaluation

After two evaluations at 9 of 10 the cause was an authored confidence ceiling of 0.98: an explanation contradicted by one of three readings kept 12–13% of the weight, so Iskra paid for measurements in random worlds. Confidence is now computed from the joint track record of the explanations she holds.

New worlds, the same gate: 10 of 10, a measurement in 0 of 6 random worlds, net 0.927 against 0.893 for the strong fixed strategy. Over 200 random worlds she pays in 1% (before: 19.5%).

Try it ↗

18 October 2026did not pass · 9 of 10

Several explanations, a probe she chooses herself

A vault stores and returns values; a read may leave the value, take it or return chance. Iskra holds several explanations derived from experience, decides herself whether to pay 0.05 for a probe, and bets on the second read.

In 20 ambiguous worlds 20 of 20 bets were right; net 0.857 against 0.827 for the fixed strategy. One criterion failed: in random worlds she paid for a measurement in 2 of 6, 1 was allowed.

18 October 2026did not pass

Does knowledge from experience change the next action?

The observer inside each branch of the plan chooses later actions by its own memory. Branch memory lowered the cost against frozen memory by 0.031 per step, and the learned procedure of the new law was necessary.

Criterion P1 failed, and the authored rule “check first” beat the planner by 0.050 per step. The cause is an authored assumption that worlds change rarely.

Full record ↗

17 October 2026did not pass

The world returns. Will Iskra recognise it?

In the new world the forecast got better: over the last 12 episodes the error was 94.0% (wave → wave) and 97.1% (saturation → wave) below the reference method.

The return was too costly: in the first four episodes of A the error was 3.27 and 15.38 times that without B; 1.05 was allowed.

Full record ↗

15 October 2026did not pass

The Flame, first evaluation

Between two worlds of the same kind knowledge transfers: B 37% faster (interval 14–59%), and after the return 0.80 of the bar. Between worlds of different kinds it hinders: −28% and 2.74 (bar 1.05).

One fit to several worlds was a bad compromise for each. Hence the bank of separate rules that entered the later trials.

What’s next

What we have not proved

Choosing an action is not learned: the planner was written by the author. There is no proof of transfer between dissimilar worlds, nor of learning a new encoding of operations.

The next bar: Iskra herself should recognise that the world changed, instead of being handed its boundaries. We will write what we checked, including when the result says to turn back.