13 Sep 2026 · current verified status

Native v0.2.0F B6/H1: next epoch 2,276 · 2,388 committed candidates · 75 qualified F2 basins · 72 work-producing L2 selections · 0 promoted rules.

Legacy A3: the 72h campaign is now DONE. Its prior 1,662-epoch / 3,324-candidate checkpoint remains the last promoted numerical baseline while the final post-stop rich evidence stays separate from public claims.

Claim boundary: ML screening only; no stable-material, novelty or DFT/hull claim is promoted. The detailed 11 Sep snapshot below is retained as its dated baseline. 12 Sep evidence JSON →

v0.2.0 · RC1 HF2B6 / H1 · 2 CPUPosnetek tekoče raziskave · 11. 9. 2026Running research snapshot · 11 Sep 2026

8zMaterials

Arena, ki se iz iskanja materialov uči, kako iskati naprej. MDL×DCC izbira prihodnje delo; Atlas naučene vzorce preizkuša na novih sestavah.An arena that learns from materials search how to search next. MDL×DCC selects future work; the Atlas tests learned patterns on other compositions.

Človeški arhitekt in končna avtoriteta: BD. Izvedba in preverjanje: BD × AI Lab.Human architect and final authority: BD. Implementation and verification: BD × AI Lab.

1 222zaključenih epohcommitted epochs

1 308kandidatnih poskusovcandidate attempts

39zgodovinsko F2 kvalificiranih strukturhistorically F2-qualified structures

41izbir dela z dokazano povezavo do L2selected work actions causally bound to L2

Datiran posnetek, ne sprotni števec. F2 je kvalifikacija ML relaksacije. Novost, stabilnost in uporabnost materialov še niso potrjene.A dated snapshot, not a live counter. F2 qualifies an ML relaxation. Material novelty, stability and usefulness are not yet confirmed.

MDL×DCC ustvarja prihodnje deloMDL×DCC produces future work

Glavna veja je B6/H1: hitra zanka iskanja je povezana s počasnejšo zanko učenja Atlasa. MDL (Minimum Description Length) ocenjuje, kateri modeli in vzorci gospodarno pojasnjujejo opažanja. DCC (Dynamic Complexity Controller) vpliva na izbiro nadaljnjih akcij. MACE ocenjuje energije in sile kandidatnih struktur.The mainline is B6/H1: a fast search loop is coupled to a slower Atlas learning loop. MDL (Minimum Description Length) evaluates which models and patterns explain observations economically. DCC (Dynamic Complexity Controller) influences subsequent action selection. MACE evaluates candidate structures through energies and forces.

Naslednji del izračunaThe next piece of computation

Odloča o nadaljevanju posamezne relaksacije: še dela, vrne kandidata v obdelavo ali zaključi porabo.Decides how to continue an individual relaxation: do more work, return a candidate for further processing, or end the allocation.

Izbira iskalne akcijeSelecting the search action

Primerja napovedi za dovoljene akcije in razporeja delo med sestave, generatorje ter reprezentacije.Compares predictions for legal actions and allocates work across compositions, generators and representations.

Učenje, ki spremeni iskanjeLearning that changes search

Atlas oblikuje pravila, jih rangira in odpira za preizkus. Njegove odločitve lahko spremenijo prihodnji nabor izvedljivih akcij.The Atlas forms and ranks rules, making them available for probing. Its decisions can change the future set of executable actions.

Povratna puščica: naučeno se vrne v delo.The reverse arrow: learning returns to work.

Opažanje → naučeno pravilo → dejanska nova akcija → primerjalni preizkus prenosa → obračun celotnega stroška → promocija ali ohranjena zavrnitev s pogojem ponovnega odprtja.Observation → learned rule → actual new action → matched transfer validation → full cost accounting → promotion or a retained rejection with a reopening condition.

Rank, don’t eliminate.

Slabši rezultat zmanjša prednost. Arhiv ohrani razlog, dokaz in možnost ponovnega odprtja. Omejen aktivni Atlas lahko tako ostaja živ tudi po zapolnitvi.A weaker result lowers priority. The archive retains the reason, evidence and opportunity to reopen. A bounded active Atlas can keep learning after it fills up.

Learn from mistakes.

Napake pri obnovi, nepravilni sestavi ali prenosu pravil postanejo trajni regresijski testi in preizkusi z znanim pravilnim izidom.Resume errors, composition violations and rule-transfer failures become durable regression tests and canaries with known expected outcomes.

Primerjalne veje in poštena meritevComparison arms and fair measurement

VejaArm	VlogaRole
B2	Običajni stroškovno občutljiv UCB comparator.Ordinary cost-aware UCB comparator.
B5	MDL-M + DCC-S
B6/H0	B6 z izključeno H1 Atlas zanko; kontrola učinka učenja.B6 without the H1 Atlas loop; a control for the learning effect.
B6/H1	Glavna native večnivojska raziskovalna veja.The main native multilevel research arm.
H2	ssMDL×DCC kot omejen eksperiment.ssMDL×DCC as a bounded experiment.

Napoved je zabeležena pred izidom, logične epohe so deterministične, učenje se knjiži natanko enkrat. Primerjave morajo imeti iste informacije in primerljive proračune. Ta posnetek sam ne dokazuje prednosti pred B2 ali H0.Predictions are recorded before outcomes, logical epochs are deterministic, and learning is committed exactly once. Comparisons require the same information and comparable budgets. This snapshot alone does not establish superiority over B2 or H0.

Atlas že vpliva na dejansko iskanjeThe Atlas already affects actual search

Od odločitve do izvedene akcijeFrom a decision to an executed action

Med 305 izmerjenimi L2 odločitvami jih je 248 spremenilo prihodnji nabor akcij. Izvedenih je bilo 256 akcij z naučenimi pravili; pri 41 izbirah je strogo zabeležena povezava med spremembo L2 in novim delom.Of 305 measured L2 decisions, 248 changed the future action space. There were 256 executed rule-guided actions; 41 selections have a strict recorded link from an L2 change to new work.

260ustvarjenih pravilrules created

16aktivnih probe pravilactive probe rules

0promoviranih pravilpromoted rules

149 ponovnih odprtij in 393 arhiviranj. To so dogodki, ki se lahko nanašajo na ista pravila. Števcev ne seštevamo kot neodvisnih uspehov.149 reopening and 393 archive events. Events may concern the same rules. These counters are not additive independent successes.

Mg₂SiO₄ → Ca₂SiO₄ · epohaepoch 552

Prvi uspešen primerjalni transferFirst accepted matched transfer

Naučeno pravilo je v enem F1 paru doseglo nižjo ML energijo ob manjši primerjani CPU porabi. Strošek pravila vključuje pripisano učenje, načrtovanje in validacijo.A learned rule achieved a lower ML energy with less compared CPU cost in one F1 pair. Rule cost includes the attributed mining, planning and validation.

PraviloRule

43.78 s

KontrolaControl

76.09 s

−42.46% CPU stroška v enem paruCPU cost in one pair

Od 141 transfer preizkusov je bil sprejet eden, zavrnjenih 140. Poznejši preizkus istega pravila ni prestal kakovosti oziroma konvergence.One of 141 transfer trials was accepted; 140 were rejected. A later trial of the same rule failed the quality/convergence check.

Pravilo še ni povrnilo stroška svojega opisa v knjižnici, zato ni promovirano. To ni splošna 42-odstotna pospešitev arene.The rule has not repaid its library description cost, so it is not promoted. This is not a general 42% arena speedup.

Kaj je med kandidati zanimivega?What is interesting among the candidates?

Trenutno iskanje zajema 12 določenih oksidnih in silikatnih sestav iz Mg, Ca, Al, Si in O. Zgodovinsko je F2 doseglo 39 struktur. Pri naknadni primerjavi jih 16 ostane znotraj 0,2 eV/atom od trenutno najboljšega PASS rezultata iste sestave. Vseh 39 ostane ohranjenih.The current search covers 12 fixed oxide and silicate compositions from Mg, Ca, Al, Si and O. Historically, 39 structures reached F2. In a posthoc comparison, 16 remain within 0.2 eV/atom of the current best PASS result of the same composition. All 39 are retained.

MgCaSiO₄

Kandidat za nadaljnje preverjanjeA candidate for further checking

Struktura iz epohe 1100 je dosegla F2: ML energija −6,940381 eV/atom in največja sila 0,034441 eV/Å.The structure from epoch 1100 reached F2: ML energy −6.940381 eV/atom and maximum force 0.034441 eV/Å.

V preprosti uravnoteženi primerjavi z MgO + CaO + SiO₂ iz istega runa je njena energija približno 0,113 eV/atom nižja.In a simple balanced comparison with MgO + CaO + SiO₂ from the same run, its energy is about 0.113 eV/atom lower.

Nepopoln interni ML presejalni signal. Referenčne strukture imajo različne ravni relaksacije; celotna energijska ovojnica, negotovost modela in novost še niso preverjene.An incomplete internal ML screening signal. Reference structures have different relaxation levels; the full convex hull, model uncertainty and novelty have not been checked.

Al₂SiO₅

Zakaj F2 sam ne zadoščaWhy F2 alone is not enough

Najboljši F2 kandidat te sestave je v analogni primerjavi približno 0,142 eV/atom nad mešanico Al₂O₃ + SiO₂. Dobro zaključena relaksacija torej sama še ne pomeni ugodne stabilnosti proti razpadu.The best F2 candidate of this composition is about 0.142 eV/atom above the Al₂O₃ + SiO₂ mixture in an analogous comparison. A well-converged relaxation alone does not establish favourable stability against decomposition.

Tudi ta kandidat ostane v arhivu. Analitični vrstni red ne spreminja zamrznjene politike ali že pripisane zasluge v runu.This candidate also stays in the archive. The analytical ranking does not change the frozen policy or credit already assigned in the run.

Vseh 12 sestav: F2 in naknadni vrstni redAll 12 compositions: F2 and posthoc ranking

Znotraj okna = največ 0,2 eV/atom nad trenutno najboljšim PASS iste sestave.Within window = at most 0.2 eV/atom above the current best PASS of the same composition.
SestavaComposition	Zgodovinski F2Historical F2	Znotraj oknaWithin window
Al₂O₃	3	2
Al₂SiO₅	3	1
Ca₂SiO₄	2	0
CaAl₂O₄	7	3
CaO	7	1
CaSiO₃	3	3
Mg₂SiO₄	1	0
MgAl₂O₄	2	1
MgCaSiO₄	2	1
MgO	5	1
MgSiO₃	2	1
SiO₂	2	2
SkupajTotal	39	16

Surovih energij različnih kemijskih sestav ne razvrščamo neposredno med seboj. Potrebne so uravnotežene primerjave.Raw energies of different chemical compositions are not ranked directly against each other. Balanced comparisons are required.

Kako bomo prepoznali nov, uporaben material?How will we recognise a new, useful material?

Arena danes hrani lasten Atlas in primerja strukture znotraj runa. Pripravljeni so vmesniki za referenčne zbirke, vendar ta run še nima priklopljene napolnjene zunanje knjižnice. MACE je model energije in sil; ni katalog za potrjevanje novosti.The arena currently keeps its own Atlas and matches structures within the run. Reference-data interfaces are prepared, but this run has no populated external library connected. MACE is an energy and force model; it is not a novelty-verification catalogue.

Preveri strukturo proti znanim materialom.Match the structure against known materials. Polna periodična struktura se primerja z datiranimi zbirkami, kot sta Materials Project in COD. Enaka formula ne pomeni enake strukture; brez zadetka še ne pomeni dokazano novega materiala.Compare the full periodic structure with dated collections such as Materials Project and COD. The same formula does not imply the same structure; no match is not yet proof of a new material.
Preveri fizikalno obstojnost.Check physical stability. Neodvisen DFT izračun, primerljive energije vseh konkurenčnih faz in fazni diagram oziroma energijska ovojnica; nato dinamična in mehanska stabilnost.Independent DFT calculations, compatible energies for competing phases and a phase diagram / convex hull, followed by dynamical and mechanical stability checks.
Izmeri lastnost, zaradi katere je material dober.Measure the property that makes the material useful. Na primer ionsko prevodnost, trdoto ali temperaturno obstojnost proti jasnemu referenčnemu materialu. Sedanji F2 tega ne meri. Sinteza in eksperimentalna meritev zaključita preverjanje.For example, ionic conductivity, hardness or temperature tolerance against a clear reference material. Current F2 does not measure these. Synthesis and experimental measurement complete the verification.

Kaj je preverjeno in na katerem posnetkuWhat was checked, and on which snapshot

Na BD-jevi Windows napravi je HF2 verifier opravil 260 testov brez napak ali preskokov; preverjanja D0–D23 so bila PASS. Rich extract je dodatno pregledan na ravni integritete, končnega stanja, povezav zgodovine, 141 transfer receiptov in 39 F2 struktur.On BD’s Windows machine, the HF2 verifier completed 260 tests with no failures or skips; checks D0–D23 were PASS. The rich extract was additionally reviewed for integrity, final state, history links, 141 transfer receipts and 39 F2 structures.

To sta različna obsega preverjanja. Paket vsebuje 17 celotnih surovih epoh; iz njega ni bil izveden celoten policy replay vseh 1.222 epoh.These are different verification scopes. The package contains 17 complete raw epochs; a full policy replay of all 1,222 epochs was not performed from it.

Vir, merilni obseg in odprti rezultatiSource, measurement scope and open outcomes

8zMaterials_NATIVE_RICH_EXTRACT_20260911_123120_095725.zip

SHA3-256: 305353d1770a13c45c7cba482f77db7bb49a9c6ad2820c1f8a1d8e3e6fc79535

181,23 MB · 11.064 članov · 2.444 glavnih delovnih nalog. Izidi: 885 PASS, 1.550 LIMIT, 8 zavrnjenih generacij in 1 numerična napaka. LIMIT označuje doseženo omejitev izračuna.181.23 MB · 11,064 members · 2,444 primary work leases. Outcomes: 885 PASS, 1,550 LIMIT, 8 generation rejections and 1 numerical error. LIMIT denotes a reached computation limit.

En rezultat nadaljevanja je bil zavrnjen zaradi neujemanja vhodne strukture; run je nadaljeval. Poraba ob posnetku: 55.504,84 od 518.400 CPU-sekund. Odstotek porabe ni odstotek dokončanega odkritja.One continuation result was rejected for an input-structure mismatch; the run continued. Snapshot expenditure: 55,504.84 of 518,400 CPU-seconds. Budget consumption is not a percentage of discovery completed.

Javni merilni zapis · JSONPublic evidence record · JSON Zapis arene v 8Z ControlArena record in 8Z Control