38978 research outputs found
Sort by
[Core] The Ayala Sigil: Reinforcement Memory Protocol for Stateless AI (v0.3.0)
No full textCore research component containing the Ayala Sigil protocol documents, hypothesis (v0.2), and supporting materials.
v0.3 compiled paper with 30-day empirical validation is at the project root.
Key files:
- AYALA_SIGIL_HYPOTHESIS.md (v0.2, superseded by root v0.3)
- PROTOCOL_FREEZE_1.0.0.md
- REPLICABILITY_OUTLINE_2026-03-02.md (superseded by v0.3 Section 6
GeoFM v2 - Update 1: Multi-Head Architecture & Spatial Resampling Strategy
No full text# PRE-REGISTRATION UPDATE 1: Multi-Head Architecture Decision
**Project:** GeoFM v2 - Cerrado Land-Use Transition Pilot
**Date:** 2026-03-08
**Version:** 1.1
**Status:** BEFORE resampling execution, BEFORE Multi-Head training
**Link to Original Registration:** [GeoFM_v1]
---
## EXECUTIVE SUMMARY
This update documents a critical decision made AFTER baseline analysis but BEFORE implementing the new approach. We identified a 26 percentage point performance gap between validation (81.3%) and test (55.3%) that is NOT traditional overfitting, but rather **spatial heterogeneity** in land-use change processes.
**Decision:** Implement Multi-Head architecture with stratified spatial resampling.
**Timing:** This registration is made BEFORE executing resampling or training the new model.
**Integrity:** Baseline model frozen (2026-03-07) and will remain unchanged for comparison.
---
## CONTEXT: What Changed Since Original Registration
### Original Plan (GeoFM_v1):
- Train baseline model (MLP with LULC embeddings)
- Evaluate on temporal-spatial splits
- Identify factors influencing conversion speed
### What We Discovered:
1. ✅ **Baseline trained successfully**
- Validation: 81.3% accuracy
- Test: 55.3% accuracy
- **Gap: 26 percentage points**
2. ✅ **Gap diagnosed as spatial heterogeneity, not overfitting:**
- Training set: Northern Cerrado, early-stage frontier (dispersed patterns)
- CDPD (Dispersed→Pasture): 49.0%
- CCPC (Cluster→Pasture): 4.2%
- Test set: Southern Cerrado, consolidated frontier (cluster patterns)
- CDPD: 39.3%
- CCPC: 37.9% (9x more than training!)
- **Interpretation:** Model learned dispersed conversion dynamics, tested on cluster dynamics
- This represents **heterochronous agricultural frontiers** (different stages)
3. ✅ **Integrated Moran I spatial patterns:**
- Used hexagonal grid (20,000 ha cells) with LISA analysis
- Confirmed spatial clustering of conversions
- Patterns: CCAC (cluster conversion + cluster agriculture), CDPD (dispersed), etc.
---
## NEW HYPOTHESIS (Emerged from Baseline)
**Hypothesis:** Dispersed and cluster conversion patterns represent fundamentally different spatial processes. A Multi-Head architecture that explicitly models these distinct processes will reduce the performance gap by allowing specialized learning for each pattern type.
**Null Hypothesis:** Multi-Head will not significantly reduce gap (<5 percentage points improvement).
---
## PROPOSED SOLUTION (Pre-specified BEFORE implementation)
### Architecture: Multi-Head with Soft Gating
```
Input: [LULC time series, spatial patch, auxiliary features] (287 features)
↓
Shared Encoder: 287 → 256 → 128 (ReLU, Dropout 0.3)
↓
├─→ Head Dispersed: 128 → 64 → 1 (specialist for CDPD, CDAD patterns)
├─→ Head Cluster: 128 → 64 → 1 (specialist for CCPC, CCAC patterns)
└─→ Gate Network: (128 + 3 spatial features) → 32 → 2 (softmax)
↓
Output: w_dispersed * pred_dispersed + w_cluster * pred_cluster
```
**Spatial features for gate:**
- `is_cluster_conversion` (binary)
- `is_cluster_use` (binary)
- `consolidation_level` (ordinal 0-3)
### Training Strategy: Stratified Spatial Resampling
**Problem identified:** Severe imbalance in spatial patterns
- Training: Head Cluster only sees 537 samples (5.4%)
- Training: Head Dispersed sees 8,905 samples (89.1%)
- Ratio: 16.6:1 (insufficient for cluster head)
**Solution:** Hybrid resampling strategy
| Pattern | Current | Target | Strategy |
|---------|---------|--------|----------|
| CDPD (Dispersed→Pasture) | 4,903 | 2,000 | Subsample |
| CDAD (Dispersed→Agric) | 1,582 | 2,000 | Oversample + 500 new |
| CCPC (Cluster→Pasture) | 417 | 2,000 | Oversample + 1,000 new |
| CCAC (Cluster→Agric) | 56 | 2,000 | Mainly new (1,800) |
| Others | — | — | See detailed strategy |
**New pixel sampling:**
- Sample from hexagons with specific patterns
- Validate N→P→S trajectory for each pixel
- Determine label (rapid ≤4 years vs slow >4 years)
- Expected total: ~14,000 balanced samples
**Justification:**
- NOT p-hacking: Gap identified in validation set also (21pp)
- NOT test-set-specific: Balancing based on spatial patterns, not test performance
- Scientifically motivated: Represents real distribution across frontier stages
### Hyperparameters (Fixed):
- Optimizer: Adam (lr=0.001)
- Batch size: 256
- Epochs: 30 (early stopping on validation)
- Dropout: 0.3
- Random seed: 42
---
## SUCCESS CRITERIA (Pre-specified)
### Primary Metric:
- **Gap reduction:** From 26pp to <10pp
- Baseline: Val 81.3% → Test 55.3% (gap: 26pp)
- Target: Val ~76-80% → Test >70% (gap: <10pp)
### Secondary Metrics:
1. **Absolute test performance:** >70% accuracy (vs 55.3% baseline)
2. **F1 Score:** >0.60 (vs 0.24 baseline)
3. **Stability:** Standard deviation across 3 seeds <0.03
### Interpretability Checks:
- Gate weights make sense (dispersed patterns → dispersed head)
- Heads learn distinct representations (verifiable via feature importance)
---
## DECISION GATES
### Gate 1: After Resampling (2026-03-09)
**Question:** Did we obtain sufficient balanced samples?
- ✅ **If >70% of target:** Proceed with Multi-Head
- ⚠️ **If 40-70% of target:** Adapt strategy (accept fewer samples)
- ❌ **If <40% of target:** Pivot to weighted loss (no resampling)
### Gate 2: After Multi-Head Training (2026-03-13)
**Question:** Did Multi-Head significantly reduce gap?
- ✅ **If gap <10pp:** SUCCESS - Scale and publish
- ⚠️ **If gap 10-15pp:** PARTIAL - Document, consider ensemble
- ❌ **If gap >15pp:** INSUFFICIENT - Paper focuses on problem, not solution
---
## WHAT WE WILL NOT DO (Anti-P-Hacking Safeguards)
- ❌ Will NOT iteratively tune on test set
- ❌ Will NOT add arbitrary features to improve test performance
- ❌ Will NOT change success criteria after seeing results
- ❌ Will NOT report only successful experiments
- ❌ Will NOT use test set for any development decisions
**Commitment:** We will publish results honestly regardless of outcome.
---
## COMPARISON WITH BASELINE (Frozen)
### Baseline (Frozen 2026-03-07):
- **Model:** MLP with LULC class embeddings
- **Data:** Original 10,000 samples (imbalanced)
- **Performance:** Val 81.3%, Test 55.3%, F1 0.24
- **Location:** `D:\Projetos\Cerrado\GeoFM_sampling\baseline\`
- **Status:** FROZEN - no further modifications
### Multi-Head (To be trained):
- **Model:** Multi-Head as specified above
- **Data:** Balanced ~14,000 samples
- **Performance:** TBD (will report honestly)
- **Comparison:** Side-by-side with frozen baseline
---
## SCIENTIFIC JUSTIFICATION
### Why This Is NOT P-Hacking:
1. ✅ **Baseline frozen BEFORE exploring solutions**
- Gap identified on 2026-03-07
- Baseline saved and no longer modified
- Multi-Head decision on 2026-03-08
2. ✅ **Gap observed in validation set also**
- Validation gap: 21pp (Val-Train)
- Test gap: 26pp (Val-Test)
- Not an artifact of test set
3. ✅ **Solution is general, not test-specific**
- Multi-Head architecture motivated by spatial heterogeneity theory
- Not tuned to specific test set properties
- Applicable to other regions/datasets
4. ✅ **Strategy pre-specified before execution**
- This registration made BEFORE resampling
- This registration made BEFORE Multi-Head training
- Cannot be influenced by results not yet obtained
### Why This Advances Science:
**Contribution 1:** Identification of heterochronous frontiers in LULC-ML
- First quantification of spatial heterogeneity impact on ML models
- Gap attributed to different frontier stages (dispersed vs cluster)
**Contribution 2:** Multi-Head as solution for spatial heterogeneity
- Novel architecture for spatially heterogeneous processes
- Framework applicable to other contexts (climate, urbanization, etc.)
**Contribution 3:** Spatial stratification methodology
- Importance of pattern-balanced sampling
- Validation should be stratified by spatial context
---
## TIMELINE
- **2026-03-01:** Project initiated
- **2026-03-07:** Baseline frozen (gap: 26pp identified)
- **2026-03-08:** THIS UPDATE (Multi-Head decision, BEFORE execution)
- **2026-03-09:** Resampling execution (overnight)
- **2026-03-10:** Dataset balancing complete
- **2026-03-12:** Multi-Head training
- **2026-03-13:** Results & Decision Gate 2
- **2026-03-15:** Final update with complete results
---
## FILES ATTACHED
1. `DECISAO_ESTRATEGICA_MULTIHEAD.md` - Full decision documentation
2. `REVISAO_CRITICA_ALINHAMENTO.md` - Critical self-assessment
3. `1_gerar_amostras_estratificadas_v2.py` - Resampling script (ready, not executed)
4. `baseline/` folder - Frozen baseline model, metrics, and data splits
---
## TRANSPARENCY COMMITMENT
We commit to:
1. ✅ Publishing all results (positive or negative)
2. ✅ Reporting complete methodology (including failures)
3. ✅ Making code publicly available (upon paper acceptance)
4. ✅ Making data publicly available (upon paper acceptance)
5. ✅ Honest comparison with frozen baseline
---
**Registered by:** Mario Barroso Ramos Neto
**Institution:** The Nature Conservancy
**Date:** 2026-03-08
**Status:** Pre-implementation (strategy defined, execution pending)
**Contact:** [email protected]
``
Politomorphism: Symbolic Resonance Mapping (SRM) and the Entropic Equilibrium Function (EEF)
No full text[RO]
Această lucrare prezintă Politomorphism, un cadru teoretic și computațional integrat, conceput pentru modelarea sistemelor politice complexe prin intermediul a cinci componente formale corelate. Arhitectura „Politomorphism Engine” include:
Metapolitical Calibration (MC) pentru repoziționarea strategică în noosfera digitală;
Cognitive Morphogenesis (CMG) pentru evoluția conștiinței politice de la stadiul tribal la metaconștiință;
Symbolic Resonance Mapping (SRM) pentru măsurarea impactului simbolurilor asupra grupurilor sociale;
Entropic Equilibrium Function (EEF) pentru cuantificarea dezordinii sistemice prin termodinamică politică;
Legitimacy Dynamics Engine (LDE) pentru simularea ciclurilor de autoritate și risc de colaps.
Cercetarea demonstrează modul în care rezonanța simbolică influențează entropia sistemică și stabilitatea legitimității, oferind un instrument de analiză predictivă pentru dinamica puterii în secolul XXI.
[EN]
This paper presents Politomorphism, an integrated theoretical and computational framework designed to model complex political systems through five interrelated formal components.
The "Politomorphism Engine" architecture includes:
Metapolitical Calibration (MC) for strategic repositioning within the digital noosphere;
Cognitive Morphogenesis (CMG) for the evolution of political consciousness from tribal stages to metaconsciousness;
Symbolic Resonance Mapping (SRM) for measuring the impact of symbols across social groups;
Entropic Equilibrium Function (EEF) for quantifying systemic disorder via political thermodynamics;Legitimacy Dynamics Engine (LDE) for simulating cycles of authority and collapse risk.
The research demonstrates how symbolic resonance influences systemic entropy and legitimacy stability, providing a predictive analytical tool for 21st-century power dynamics
Financial Literacy in Ibero-America: Database 2000-2026
No full textFinancial literacy database for Ibero-America (20 countries, 2000-2026). Includes: financial literacy level (% adults), banking services usage (% adults with formal bank account), household debt (% of disposable income), and gender differences (gender gap in percentage points). Countries included: Spain, Portugal, Mexico, Argentina, Brazil, Colombia, Chile, Peru, Uruguay, Venezuela, Ecuador, Bolivia, Paraguay, Costa Rica, Panama, Guatemala, Honduras, El Salvador, Nicaragua and Dominican Republic. Sources: World Bank (Global Findex 2011/2014/2017/2021), OECD/INFE, CAF, S&P Global Financial Literacy Survey, and national reports. Values between survey years are interpolated estimates based on documented trends
Дополнительные материалы к статье «Перевод, адаптация и валидация русскоязычной версии шкалы выгорания учителей (Teachers Burnout Scale, TBS)»
No full textВ дополнительных материалах представлена русскоязычная версия Шкалы выгорания учителей (Teachers Burnout Scale, TBS), адаптированная и валидизированная в ходе данного исследования (Приложение А). Методика включает 22 утверждения, оценивающих четыре ключевых компонента профессионального выгорания педагогов: удовлетворенность карьерой, ощущение поддержки со стороны администрации, стресс, связанный с работой, и негативное отношение к ученикам.
В Приложении В содержатся статистические таблицы, дополняющие результаты основного исследования: показатели описательной статистики, таблица ковариаций суждений и матрица интеркорреляций шкал.
Представленные материалы могут быть использованы исследователями и практическими психологами для диагностики профессионального выгорания в образовательной среде
A 10-Year Retrospective Analysis of HIV Epidemiology and Antiretroviral Therapy Patterns in Bihor County, Romania (2013–2023)
No full textthis study aimed to provide a retrospective analysis of the epidemiological and therapeutic characteristics of HIV-infected patients in Bihor County be-tween 2013 and 2023, examining demographic patterns, clinical staging, and antiretroviral treatment profiles
Rethinking pluralistic ignorance: People systematically under- and overestimate public engagement in climate action
No full textC-ORAL-ESQ: Corpus oral de esquizofrênicos / Brazilian corpus of schizophrenics
No full textC-ORAL-ESQ consists of 43 recordings of psychiatric consultations with patients diagnosed with schizophrenia, conducted at the Instituto Raul Soares (IRS-FHEMIG) and at the Hospital das Clínicas of UFMG (HC-UFMG/EBSERH), in Belo Horizonte. The corpus contains 103,633 words (47,828 from patients, 48,069 from doctors, 7,090 from companions, and 646 from other participants). Of the 43 recordings in the corpus, 19 have both audio and video, and 24 have audio only. Text-audio-(video) alignment was performed using the Elan software (Wittenburg et al., 2006)