SeismicShield-RL research infrastructure visual
Phase I infrastructure complete136 frozen ESM records475 planned atomic shardsFull-scale experiment deferred
Researcher hand-off guide

What SeismicShield-RL is, what it completed, and how to continue it

This page is the technical hand-off document for future researchers. It explains the project in plain language without hiding the important details: the physical problem, data, structural models, algorithms, budgets, information boundaries, reproducibility safeguards, stopping rationale, remaining work, and the exact scientific sequence required to resume the preregistered study.

Most important boundary: this release is a completed research-infrastructure contribution, not a completed algorithm-ranking experiment. No confirmatory structural-response outcome was used to claim that MAPPO, PPO, IPPO, NSGA-II, scalar GA, or random search is superior.

Contents

1. The project in one minute

SeismicShield-RL is a benchmark infrastructure for asking a difficult engineering/AI question: can reinforcement learning—especially multi-agent reinforcement learning—find better friction-damper retrofit layouts for buildings under earthquake and structural uncertainty than transparent heuristic and evolutionary optimizers?

The project does not optimize a single number. It studies a three-way trade-off among retrofit cost, maximum inter-story drift ratio (MIDR), and peak floor acceleration (PFA). The intended end product of Stage A is a frozen set of 768 selected designs across six stochastic methods, eight seeds, and sixteen structural states. Those selections are then supposed to face a completely untouched Tier-2 confirmatory campaign.

34
earthquake events
136
frozen ESM records
16
structural states
6 × 8 × 16
= 768 intended selected stochastic designs

2. The physical and optimization problem

A friction damper dissipates earthquake energy through controlled sliding. In this benchmark, the optimizer does not merely ask whether a building should have dampers. It decides, story by story, how many dampers to place and which slip-force level to use.

Per-story decision variables

  • Damper count: 0, 1, 2, 3, or 4.
  • Slip-force grid: 0, 50,000, 100,000, 200,000, 350,000 N.
  • If damper count is zero, slip force is physically irrelevant and is canonicalized so it does not create a fake distinct design.

Three objectives

  • Cost proxy — cheaper retrofit is better.
  • MIDR — lower maximum inter-story drift is better.
  • PFA — lower peak floor acceleration is better.

These objectives can conflict; a low-drift design may cost more or increase acceleration elsewhere. That is why Pareto analysis matters.

3. Frozen scientific design

Earthquake partitions

PartitionRecordsRoleAllowed before selection freeze?
Training52Learning / optimizationYes
Validation20Candidate and checkpoint selectionYes
Pilot16Software/runtime validation onlyNot for primary training/selection
Confirmatory48Final out-of-sample testNo

The repository does not redistribute restricted waveform bytes. Record identity is preserved through provenance and frozen processed-waveform SHA-256 values.

Structural uncertainty

Four building heights are represented: 3, 6, 10, and 20 stories. Each height has one nominal model and three frozen perturbations, producing 16 structural states. This prevents the benchmark from becoming a single-building optimization toy.

Algorithms and seeds

The stochastic method ladder is: random search, scalar genetic algorithm, NSGA-II, PPO, IPPO, and MAPPO. Eight frozen seeds are used:

1103, 2207, 3313, 4421, 5521, 6637, 7753, 8861

Fair budget principle

The project freezes computational budgets and selection rules before confirmatory evaluation. Non-policy optimizers receive 3,200 Tier-1 training calls per structural state. Learned policies share the frozen 51,200-call training budget across all 16 states with a balanced deterministic context schedule. Validation behavior is also prespecified rather than chosen after seeing results.

4. What the pipeline actually does

1 · AuthenticateVerify immutable source, manifests, contracts, seeds, and hashes.
2 · Prepare Stage AExpose training + validation only; keep confirmatory hydration/execution locked.
3 · Tier-1 learningEvaluate millions of structural-response calls while optimizers/policies search the design space.
4 · Validation selectionUse only the frozen validation partition to choose candidates/checkpoints.
5 · Selection freezeFreeze 768 method/seed/state designs and learned checkpoints at an immutable reference.
6 · Tier-2 confirmatoryOnly after the freeze, hydrate confirmatory data and evaluate selected designs with OpenSeesPy.
7 · Event-level inferenceAggregate by earthquake event and run the preregistered statistical analysis.
Completed through this release: infrastructure, authentication/preflight, deterministic planning, pilot runtime verification, and a merged selection-only workspace. The large Stage-A scientific run and everything after it remain deferred.

5. Reproducibility architecture

Immutable scientific source

Scientific code is frozen at tag confirmatory-v0.8.2-final, commit cecd3b6c27b5deb6cb6be7ddc478cfc407a45644.

Fail-closed gate

If source, manifests, numerical settings, algorithm bundle, seeds, or analysis contract cannot be authenticated, execution is supposed to stop rather than silently substitute a different state.

Information boundary

Confirmatory identity, waveforms, undamped responses, and outcomes are not available to the policy during training/selection.

Cryptographic provenance

All 136 processed-waveform hashes were independently reproduced during runtime preflight.

Atomic execution semantics

Infrastructure interruption means whole-shard restart. A solver/numerical failure returned by the simulator is a scientific outcome and consumes a call; it is not silently retried away.

Evidence discipline

Claims are linked to an evidence ledger so infrastructure findings are not confused with efficacy findings.

6. Exact compute plan and why the project stopped here

ComponentShardsCalls
Tier-1 feature precompute16832
Tier-1 non-policy train + validate3841,474,560
Tier-1 learned train + validate241,305,600
Stage A total4242,780,992
Tier-2 seeded confirmatory4836,864
Tier-2 support32,304
Grand total4752,820,160

Runtime preflight measured approximately 1.32904380175 s/call for Tier-1 and 1.98916977725 s/call for Tier-2 on the tested environment. Applying those measurements gives roughly 1,026.68 sequential Tier-1 hours for Stage A and 21.64 sequential Tier-2 hours.

Why stopping was scientifically preferable: the 24 learned Stage-A shards contain 54,400 calls each and are scientifically atomic under the frozen implementation. Breaking them merely to fit hosted-CI wall-clock limits would change the preregistered execution semantics. The project therefore stopped before the expensive run rather than reduce budgets, alter algorithms, or present a smaller convenience experiment as the registered study.

A persistent local or self-hosted compute target could run the study later. The stopping reason is therefore resource availability, not a discovered failure of the scientific protocol.

7. What this release actually achieved

  • A public preregistration and preserved protocol boundary.
  • An immutable scientific source and frozen execution/analysis contracts.
  • A 136-record ESM manifest with explicit training/validation/pilot/confirmatory roles.
  • Six stochastic benchmark methods under frozen budgets and seeds.
  • Sixteen structural states across four building heights and structural perturbations.
  • Tier-1 and OpenSeesPy Tier-2 runtime paths successfully exercised on pilot fixtures.
  • Independent reproduction of all 136 processed-waveform SHA-256 identities.
  • A deterministic 475-shard plan for the entire registered workload.
  • A selection-only workspace that keeps all 51 confirmatory Tier-2 shards locked.
  • Clear separation between infrastructure evidence and algorithm-efficacy evidence.

This means the release is not “nothing because the final ranking was not computed.” Its contribution is a ready, auditable experimental chassis for an expensive simulation-based AI benchmark.

8. How much does this move a future research project forward?

The correct way to measure the saving is by research phases removed, not by promising a universal number of days. A future team does not need to start from an empty repository and redesign the scientific controls.

Research taskStatus hereWhat a future researcher inherits
Research question and objective definitionsFrozenReady benchmark framing and normalization
Earthquake manifest and splitFrozen + hash-verifiedNo need to invent new train/validation/confirmatory partitions for the original study
Structural uncertainty ensembleDefined16 ready structural states
Baselines + RL/MARL method ladderDefinedComparable algorithms with equal-budget contracts
Seeds, checkpoint schedule, validation rulesFrozenRemoves post-hoc selection freedom
Tier-1 / Tier-2 runtime pathsPilot-validatedKnown operational numerical stack on tested fixtures
Execution decomposition475-shard planWorkload already partitioned and counted
Open-science provenanceImplementedDOIs, hashes, evidence ledger, immutable source
Full scientific computeNot runThis is the main remaining bottleneck

In practical terms, the project can remove entire setup, provenance, preregistration, benchmark-design, pilot-validation, and workload-planning phases. The actual calendar saving depends on the lab, hardware, and whether the researcher continues the original protocol or creates a derivative study.

9. Who can reuse this work?

Earthquake / structural-control researchers

Reuse the structural-world design, damper action space, response metrics, and Tier-2 OpenSeesPy confirmatory concept.

RL and multi-agent RL researchers

Test new policy architectures while inheriting a carefully separated training/validation/confirmatory benchmark design.

Multi-objective optimization researchers

Compare new evolutionary, Bayesian, surrogate-assisted, or Pareto optimizers against the frozen baseline ladder.

Reproducible-computing researchers

Study immutable source, fail-closed gates, evidence ledgers, cryptographic provenance, atomic shards, and pre-confirmatory selection freezes.

HPC / orchestration researchers

Use the deterministic workload plan as a case study for scientifically safe scheduling of expensive stochastic simulations.

Robustness / uncertainty researchers

Extend structural uncertainty, earthquake OOD evaluation, or transfer analysis while retaining a traceable benchmark core.

10. Exact scientifically valid continuation sequence

  1. Keep the immutable source tag and frozen contracts unchanged for the original registered study.
  2. Prepare the selection-only workspace.
  3. Hydrate training + validation only.
  4. Run all 424 Stage-A shards.
  5. Verify the exact Stage-A call count: 2,780,992.
  6. Freeze the 768 selected method/seed/state designs and learned checkpoints.
  7. Publish that selection freeze at an immutable Git reference.
  8. Only then authorize confirmatory data hydration.
  9. Run the 51 Tier-2 shards totaling 39,168 calls.
  10. Apply the frozen event-level analysis contract.
Important: if a future team changes algorithms, budgets, partitions, action grids, numerical settings, or selection rules, that work can still be valuable—but it should be labeled a new study/version, not presented as the untouched original preregistered confirmatory experiment.

11. Preregistered confirmatory analysis

The final comparison was designed around earthquake-event clusters rather than pretending each record/world is statistically independent. The frozen analysis defines 12 confirmatory event clusters, Pareto hypervolume analysis, cost slices, and event-level effects.

  • Primary H1 contrasts: MAPPO vs PPO, MAPPO vs NSGA-II, MAPPO vs IPPO.
  • Event-cluster bootstrap: 20,000 repetitions.
  • Exact two-sided sign-flip test: 4,096 configurations (= 2^12).
  • Multiple-comparison control: Holm, alpha 0.05.
  • H2 response metrics: MIDR and PFA across four prespecified cost ceilings.
  • Secondary robustness work includes unseen-event generalization and structural uncertainty.

None of these confirmatory inferential outputs have been generated for the final efficacy question in this release.

12. Repository map for a new researcher

Start here

README.md
docs/index.html
docs/researcher-guide.html

Final technical status

paper/TECHNICAL_REPORT_FINAL_INFRASTRUCTURE_RELEASE.md
paper/EVIDENCE_LEDGER.csv

Scientific execution contract

open_science/confirmatory_execution_v0.8.2.yaml
open_science/confirmatory_analysis_v0.8.2.yaml

Operational guidance

open_science/CONFIRMATORY_EXECUTION_OPERATIONS_v0.8.2.md
open_science/CONFIRMATORY_RUNTIME_RUNBOOK_v0.8.2.md

Preflight evidence

results/validation/confirmatory_runtime_preflight_v0.8.2.json

Persistent identifiers

Zenodo version DOI: 10.5281/zenodo.22067278
Concept DOI: 10.5281/zenodo.22067277
OSF: 10.17605/OSF.IO/64DTX

13. Scientific and engineering boundaries

Do not infer from this release: that any algorithm is superior; that selected policies generalize to unseen earthquakes; that a proposed damper layout is safe for a real building; or that the software provides code compliance, life-safety certification, retrofit approval, or earthquake prediction.

The platform is research software. Real-building use would require domain-specific validation and a licensed structural-engineering process. The most scientifically valuable fact about the stopping point is that the confirmatory outcome boundary remains uninspected, preserving a clean future continuation.

Araştırmacı devir teslim rehberi

SeismicShield-RL nedir, neyi tamamladı ve nasıl devam ettirilebilir?

Bu sayfa gelecekte projeyi devralacak araştırmacılar için teknik bir rehberdir. Fiziksel problemi, verileri, yapısal modelleri, algoritmaları, hesap bütçelerini, bilgi sınırlarını, yeniden üretilebilirlik korumalarını, durdurma gerekçesini, kalan işleri ve önceden kaydedilmiş çalışmanın bilimsel olarak doğru biçimde nasıl sürdürüleceğini sade ama teknik bir dille açıklar.

En önemli sınır: bu sürüm tamamlanmış bir araştırma altyapısı katkısıdır; tamamlanmış bir algoritma sıralama deneyi değildir. MAPPO, PPO, IPPO, NSGA-II, scalar GA veya random search'ün üstün olduğuna dair doğrulayıcı sonuç üretilmemiştir.

1. Bir dakikada proje

SeismicShield-RL şu soruyu ölçülebilir ve yeniden üretilebilir hale getirir: pekiştirmeli öğrenme ve özellikle çok-etmenli pekiştirmeli öğrenme, deprem ve yapısal belirsizlik altında sürtünmeli sönümleyici yerleşiminde klasik sezgisel ve evrimsel optimizatörlerden daha iyi maliyet–ötelenme–ivme dengeleri bulabilir mi?

Stage A'nın amaçlanan çıktısı, altı stokastik yöntem × sekiz seed × on altı yapısal durumdan oluşan 768 seçilmiş tasarımın doğrulayıcı veriler görülmeden dondurulmasıdır.

136
dondurulmuş ESM kaydı
16
yapısal durum
475
planlı atomik shard
2.820.160
toplam planlı yapısal yanıt çağrısı

2. Fiziksel ve optimizasyon problemi

Sürtünmeli sönümleyici, kontrollü kayma yoluyla deprem enerjisini dağıtır. Benchmark her kat için iki karar verir: kaç sönümleyici ve hangi slip-force seviyesi.

Kat başına karar uzayı

  • Sönümleyici sayısı: 0–4.
  • Slip-force grid'i: 0, 50.000, 100.000, 200.000, 350.000 N.
  • Sayı sıfırsa slip-force fiziksel olarak etkisizdir ve tasarım kimliği buna göre canonicalize edilir.

Üç amaç

  • Retrofit maliyet göstergesi.
  • Maksimum katlar arası ötelenme oranı (MIDR).
  • Tepe kat ivmesi (PFA).

Bu amaçlar birbirleriyle çatışabildiği için Pareto analizi gerekir.

3. Dondurulmuş bilimsel tasarım

BölümKayıtRol
Eğitim52Öğrenme / optimizasyon
Doğrulama20Aday ve checkpoint seçimi
Pilot16Yalnızca yazılım/runtime doğrulaması
Confirmatory48Nihai out-of-sample test

Dört bina yüksekliği vardır: 3, 6, 10 ve 20 kat. Her yükseklikte bir nominal ve üç dondurulmuş pertürbasyon vardır; toplam 16 yapısal durum. Stokastik yöntemler random search, scalar GA, NSGA-II, PPO, IPPO ve MAPPO'dur. Dondurulmuş seed'ler:

1103, 2207, 3313, 4421, 5521, 6637, 7753, 8861

4. Pipeline gerçekte ne yapar?

1 · Kimlik doğrulamaKaynak, manifest, kontrat, seed ve hash'leri kontrol eder.
2 · Stage AYalnızca eğitim + doğrulamayı açar; confirmatory bölümü kilitli kalır.
3 · Tier-1 aramaOptimizatörler/policy'ler milyonlarca yapısal yanıt çağrısı üzerinde tasarım arar.
4 · Doğrulama seçimiAday ve checkpoint seçimi yalnızca validation ile yapılır.
5 · Selection freeze768 tasarım ve learned checkpoint'ler değişmez bir referansta dondurulur.
6 · Tier-2 confirmatoryAncak bundan sonra confirmatory veri açılır ve OpenSeesPy değerlendirmesi yapılır.
7 · İstatistikSonuçlar deprem olayı seviyesinde önceden kaydedilmiş analizle karşılaştırılır.

5. Yeniden üretilebilirlik mimarisi

Değişmez bilimsel kaynak

confirmatory-v0.8.2-final etiketi ve cecd3b6c27b5deb6cb6be7ddc478cfc407a45644 commit'i.

Fail-closed gate

Beklenen bilimsel durum doğrulanamazsa çalışma sessizce başka duruma geçmez; durur.

Bilgi sınırı

Confirmatory kayıt kimliği, dalga örnekleri ve sonuçlar eğitim/seçim sırasında policy'ye verilmez.

Kriptografik provenance

136 işlenmiş dalga kaydının SHA-256 değerleri preflight sırasında bağımsız olarak yeniden üretildi.

Atomik shard semantiği

Altyapı kesintisinde shard baştan başlar; sayısal/solver failure ise bilimsel bir sonuç olarak çağrıyı tüketir.

Evidence ledger

Altyapı bulguları ile etkinlik iddiaları birbirine karıştırılmaz.

6. Hesap planı ve neden burada durdu?

BileşenShardÇağrı
Tier-1 feature precompute16832
Tier-1 non-policy train + validate3841.474.560
Tier-1 learned train + validate241.305.600
Stage A4242.780.992
Tier-2 toplam5139.168
Genel toplam4752.820.160

Preflight ölçümü Tier-1 için yaklaşık 1,329 s/çağrı, Tier-2 için 1,989 s/çağrı verdi. Bu, Stage A için yaklaşık 1.026,68 sıralı saat, Tier-2 için yaklaşık 21,64 sıralı saat demektir.

Durdurma gerekçesi: 24 learned shard'ın her biri 54.400 çağrı içerir ve dondurulmuş implementasyonda bilimsel olarak atomiktir. Yalnızca hosted-CI süre sınırına sığdırmak için bunları bölmek preregistered yürütme semantiğini değiştirirdi. Bu nedenle bütçeleri küçültmek veya deneyin daha küçük bir versiyonunu “asıl çalışma” gibi sunmak yerine büyük hesaplama öncesinde duruldu.

7. Bu sürümün somut başarısı

  • OSF preregistration ve korunmuş confirmatory sınırı.
  • Dondurulmuş bilimsel kaynak, execution contract ve analysis contract.
  • 136 ESM kaydı ve açık partition rolleri.
  • 16 yapısal durum ve altı stokastik yöntem için adil bütçe mantığı.
  • Tier-1 ve OpenSeesPy Tier-2 yolunun pilot fixture'larda başarılı çalışması.
  • 136 waveform hash'inin bağımsız doğrulanması.
  • Tüm deneyin deterministik 475-shard planı.
  • 51 confirmatory Tier-2 shard'ını kilitli tutan selection-only workspace.

8. Gelecekteki araştırmacıya ne kadar iş kazandırır?

Kazancı “her laboratuvar için aynı sayıda gün” diye vermek doğru olmaz. Daha doğru ölçü, baştan yapılması gerekmeyen araştırma aşamalarıdır. Yeni ekip; veri split'ini, seed'leri, bütçeleri, checkpoint seçimini, confirmatory bilgi sınırını, provenance sistemini, pilot runtime kontrolünü ve 475-shard iş planını yeniden icat etmek zorunda değildir.

Başka bir deyişle proje; benchmark tasarımı + open-science kontrolü + veri kimliği + pilot doğrulama + yürütme planlaması aşamalarını hazır bir başlangıç noktasına dönüştürür. Asıl büyük eksik kalan iş, tam ölçekli compute ve sonrasındaki selection freeze/confirmatory analizdir.

9. Kimler kullanabilir?

Deprem / yapısal kontrol

Sönümleyici action space, yapısal dünya tasarımı ve Tier-2 confirmatory yaklaşımını kullanabilir.

RL / MARL

Yeni policy mimarilerini veri sızıntısına dirençli bir benchmark üzerinde sınayabilir.

Çok amaçlı optimizasyon

Yeni evrimsel, Bayesian veya surrogate-assisted yöntemleri mevcut baseline ladder ile karşılaştırabilir.

Reproducible science

Immutable source, fail-closed gate, evidence ledger, hash provenance ve selection freeze desenlerini yeniden kullanabilir.

HPC / orchestration

Bilimsel semantiği bozmadan pahalı stokastik işlerin planlanması için 475-shard yapısını inceleyebilir.

Robustness / uncertainty

OOD deprem, yapısal belirsizlik ve bina yüksekliği transfer çalışmalarını genişletebilir.

10. Orijinal preregistered çalışmaya nasıl devam edilir?

  1. Immutable source ve frozen contract'ları koru.
  2. Selection-only workspace'i hazırla.
  3. Yalnızca 52 training + 20 validation kaydını hydrate et.
  4. 424 Stage-A shard'ın tamamını çalıştır.
  5. Toplam 2.780.992 Stage-A çağrısını doğrula.
  6. 768 seçilmiş tasarımı ve learned checkpoint'leri dondur.
  7. Selection freeze'i immutable Git ref olarak yayımla.
  8. Ancak bundan sonra confirmatory hydration'a izin ver.
  9. 51 Tier-2 shard / 39.168 çağrıyı çalıştır.
  10. Dondurulmuş event-level analizi uygula.
Algoritma, budget, partition, action grid veya selection rule değişirse bu çalışma yine değerli olabilir; ancak artık orijinal preregistered confirmatory çalışmanın aynısı değildir ve yeni sürüm/çalışma olarak etiketlenmelidir.

11. Önceden kaydedilmiş istatistik

  • 12 confirmatory deprem-event cluster.
  • Ana karşılaştırmalar: MAPPO–PPO, MAPPO–NSGA-II, MAPPO–IPPO.
  • 20.000 event-cluster bootstrap tekrarı.
  • 4.096 exact sign-flip konfigürasyonu.
  • Holm çoklu karşılaştırma düzeltmesi, alpha 0,05.
  • MIDR ve PFA için dört önceden belirlenmiş maliyet tavanında ikincil analizler.

12. Yeni araştırmacı için dosya haritası

Başlangıç

README.md
docs/index.html
docs/researcher-guide.html

Nihai teknik durum

paper/TECHNICAL_REPORT_FINAL_INFRASTRUCTURE_RELEASE.md
paper/EVIDENCE_LEDGER.csv

Bilimsel kontratlar

open_science/confirmatory_execution_v0.8.2.yaml
open_science/confirmatory_analysis_v0.8.2.yaml

Operasyon

open_science/CONFIRMATORY_EXECUTION_OPERATIONS_v0.8.2.md
open_science/CONFIRMATORY_RUNTIME_RUNBOOK_v0.8.2.md

13. Ne iddia edilmemeli?

Bu sürüm herhangi bir algoritmanın üstünlüğünü, gerçek bir binanın güvenliğini, yönetmelik uygunluğunu, retrofit onayını veya deprem tahmin yeteneğini kanıtlamaz.

En değerli korunmuş özellik, confirmatory sonuç sınırının görülmemiş olmasıdır. Uygun compute gelecekte bulunursa çalışma temiz biçimde sürdürülebilir.

Guía de transferencia para investigadores

Qué es SeismicShield-RL, qué completó y cómo puede continuarse

Esta página es el documento técnico de transferencia para investigadores futuros. Explica, con lenguaje accesible pero preciso, el problema físico, los datos, los modelos estructurales, los algoritmos, los presupuestos computacionales, las fronteras de información, las protecciones de reproducibilidad, la razón de la detención, el trabajo pendiente y la secuencia científicamente válida para reanudar el estudio prerregistrado.

Límite esencial: esta versión es una contribución de infraestructura de investigación completada, no un experimento terminado de ranking de algoritmos. No se ha producido evidencia confirmatoria que establezca superioridad de MAPPO, PPO, IPPO, NSGA-II, scalar GA o random search.

1. El proyecto en un minuto

SeismicShield-RL convierte en un benchmark reproducible la pregunta: ¿puede el aprendizaje por refuerzo—especialmente el multiagente—encontrar mejores diseños de amortiguadores de fricción bajo incertidumbre sísmica y estructural que optimizadores heurísticos y evolutivos transparentes?

El objetivo de Stage A era congelar 768 diseños seleccionados: seis métodos estocásticos × ocho semillas × dieciséis estados estructurales, antes de exponer la evaluación confirmatoria.

136
registros ESM congelados
16
estados estructurales
475
shards atómicos planificados
2.820.160
llamadas planificadas

2. Problema físico y de optimización

Un amortiguador de fricción disipa energía sísmica mediante deslizamiento controlado. El benchmark decide por planta cuántos amortiguadores colocar y qué nivel de fuerza de deslizamiento usar.

Variables por planta

  • Número de amortiguadores: 0–4.
  • Grid de fuerza: 0, 50.000, 100.000, 200.000, 350.000 N.
  • Si el número es cero, la fuerza se considera físicamente irrelevante y la representación se canonicaliza.

Tres objetivos

  • Proxy de coste del retrofit.
  • Máxima deriva entre plantas (MIDR).
  • Aceleración máxima de planta (PFA).

Como los objetivos pueden entrar en conflicto, se necesita análisis de Pareto.

3. Diseño científico congelado

ParticiónRegistrosFunción
Entrenamiento52Aprendizaje / optimización
Validación20Selección de candidatos/checkpoints
Piloto16Validación de software/runtime
Confirmatoria48Prueba final fuera de muestra

Hay edificios de 3, 6, 10 y 20 plantas. Cada altura incluye un modelo nominal y tres perturbaciones congeladas: 16 estados estructurales. Los métodos estocásticos son random search, scalar GA, NSGA-II, PPO, IPPO y MAPPO. Semillas:

1103, 2207, 3313, 4421, 5521, 6637, 7753, 8861

4. Qué hace realmente el pipeline

1 · AutenticarVerifica fuente inmutable, manifests, contratos, semillas y hashes.
2 · Stage AAbre solo entrenamiento + validación; la parte confirmatoria permanece bloqueada.
3 · Búsqueda Tier-1Los optimizadores/policies exploran diseños mediante millones de evaluaciones estructurales.
4 · SelecciónLos candidatos/checkpoints se eligen únicamente con validación.
5 · FreezeSe congelan 768 diseños y checkpoints en una referencia inmutable.
6 · Tier-2Solo después se hidratan datos confirmatorios y se evalúa con OpenSeesPy.
7 · InferenciaLos resultados se agregan por evento sísmico y se aplica el análisis prerregistrado.

5. Arquitectura de reproducibilidad

Fuente científica inmutable

Tag confirmatory-v0.8.2-final, commit cecd3b6c27b5deb6cb6be7ddc478cfc407a45644.

Gate fail-closed

Si el estado esperado no puede autenticarse, la ejecución se detiene en vez de sustituir silenciosamente otro estado.

Frontera de información

La identidad, señales y resultados confirmatorios no se entregan a la policy durante entrenamiento/selección.

Procedencia criptográfica

Los 136 hashes SHA-256 procesados fueron reproducidos independientemente en el preflight.

Semántica atómica

Una interrupción de infraestructura implica reinicio completo del shard; un fallo numérico del simulador se conserva como resultado científico.

Evidence ledger

Se separan explícitamente hallazgos de infraestructura y afirmaciones de eficacia.

6. Plan computacional y razón de la detención

ComponenteShardsLlamadas
Tier-1 feature precompute16832
Tier-1 non-policy train + validate3841.474.560
Tier-1 learned train + validate241.305.600
Stage A4242.780.992
Tier-2 total5139.168
Total general4752.820.160

El preflight midió aproximadamente 1,329 s/llamada en Tier-1 y 1,989 s/llamada en Tier-2: unas 1.026,68 horas secuenciales para Stage A y 21,64 horas para Tier-2.

Por qué se detuvo: cada uno de los 24 shards learned contiene 54.400 llamadas y es científicamente atómico bajo la implementación congelada. Dividirlo solo para ajustarlo a límites de CI cambiaría la semántica prerregistrada. Se prefirió detener el proyecto antes del cálculo masivo a reducir presupuestos o presentar un experimento reducido como si fuera el estudio registrado.

7. Logros reales de esta versión

  • Prerregistro OSF y frontera confirmatoria preservada.
  • Fuente científica y contratos de ejecución/análisis congelados.
  • Manifest de 136 registros ESM con roles explícitos.
  • 16 estados estructurales y seis métodos estocásticos con presupuestos fijados.
  • Rutas Tier-1 y OpenSeesPy Tier-2 probadas con fixtures piloto.
  • Reproducción independiente de los 136 hashes de waveform.
  • Plan determinista completo de 475 shards.
  • Workspace de selección que mantiene bloqueados los 51 shards Tier-2 confirmatorios.

8. ¿Cuánto trabajo ahorra al investigador siguiente?

No es responsable prometer el mismo número de días para todos los laboratorios. La ventaja se mide mejor por fases de investigación que ya no hay que diseñar desde cero. Un equipo futuro hereda las particiones de datos, semillas, presupuestos, reglas de selección, frontera confirmatoria, provenance, validación piloto y descomposición de la carga.

En la práctica, elimina gran parte de las fases de diseño del benchmark, control open-science, identidad de datos, validación de runtime y planificación de ejecución. El gran cuello de botella que permanece es el compute a escala completa, seguido del freeze de selección y la evaluación confirmatoria.

9. ¿A quién le sirve?

Ingeniería sísmica / control estructural

Puede reutilizar el espacio de diseño de amortiguadores, estados estructurales y concepto confirmatorio Tier-2.

RL / MARL

Puede probar nuevas policies sobre una separación entrenamiento-validación-confirmatoria cuidadosamente definida.

Optimización multiobjetivo

Puede comparar nuevos métodos evolutivos, bayesianos o con surrogate contra los baselines.

Ciencia reproducible

Puede reutilizar patrones de source freeze, gates fail-closed, evidence ledger, hashes y selection freeze.

HPC / orquestación

Puede estudiar planificación segura de simulaciones estocásticas costosas mediante el ledger de 475 shards.

Robustez / incertidumbre

Puede ampliar OOD sísmico, incertidumbre estructural y transferencia entre alturas.

10. Secuencia exacta para reanudar el estudio original

  1. Mantener la fuente inmutable y contratos congelados.
  2. Preparar el workspace de selección.
  3. Hidratar solo 52 training + 20 validation.
  4. Ejecutar los 424 shards de Stage A.
  5. Verificar exactamente 2.780.992 llamadas.
  6. Congelar los 768 diseños seleccionados y checkpoints learned.
  7. Publicar el freeze en una referencia Git inmutable.
  8. Solo entonces autorizar datos confirmatorios.
  9. Ejecutar 51 shards Tier-2 / 39.168 llamadas.
  10. Aplicar el análisis por evento congelado.
Si se modifican algoritmos, presupuestos, particiones, action grid o reglas de selección, puede ser un excelente estudio nuevo; pero debe presentarse como nueva versión/estudio, no como la ejecución intacta del experimento prerregistrado original.

11. Análisis estadístico prerregistrado

  • 12 clusters de eventos sísmicos confirmatorios.
  • Contrastes principales: MAPPO vs PPO, MAPPO vs NSGA-II, MAPPO vs IPPO.
  • 20.000 repeticiones de bootstrap por cluster de evento.
  • 4.096 configuraciones exactas de sign-flip.
  • Corrección de Holm, alpha 0,05.
  • Análisis de MIDR y PFA bajo cuatro techos de coste prespecificados.

12. Mapa del repositorio

Inicio

README.md
docs/index.html
docs/researcher-guide.html

Estado técnico final

paper/TECHNICAL_REPORT_FINAL_INFRASTRUCTURE_RELEASE.md
paper/EVIDENCE_LEDGER.csv

Contratos científicos

open_science/confirmatory_execution_v0.8.2.yaml
open_science/confirmatory_analysis_v0.8.2.yaml

Operaciones

open_science/CONFIRMATORY_EXECUTION_OPERATIONS_v0.8.2.md
open_science/CONFIRMATORY_RUNTIME_RUNBOOK_v0.8.2.md

13. Límites científicos y de ingeniería

Esta versión no demuestra superioridad algorítmica, seguridad de un edificio real, cumplimiento normativo, aprobación de retrofit ni capacidad de predecir terremotos.

El software es una plataforma de investigación. Su mayor ventaja científica al detenerse aquí es que la frontera confirmatoria sigue sin inspeccionarse, por lo que una continuación futura puede mantenerse limpia.