VITAL-12 and FSV-12
VITAL-12 generates and tests hypotheses about functional organization, agent behavior and system viability. FSV-12 is its current auditable taxonomy and design hypothesis. Universal validity remains an open research question.
Applied AI research initiative | IPV | Brazil
Applied AI Research and Evaluation at Instituto Projeto de Vida
Instituto Projeto de Vida is a Brazilian nonprofit educational and social organization that hosts applied research on agentic AI evaluation, reliability and applied statistics. The initiative tests claims under controlled conditions, produces versioned research artifacts and connects selected methods to real programs without exposing sensitive data.
Computer Science
Research Engineering
Applied Statistics
AI Evaluation
Human-Centered AI
Research Engineering
Applied Statistics
AI Evaluation
Human-Centered AI
147,208events in a bounded VITAL study
1,344sessions in an IPV applied-analysis study
31research claims tracked by evidence status
19null, negative or inconclusive results preserved
About the initiative
Methods, results and limits in the same record
The initiative studies agent behavior under controlled intervention, audits the evaluators used to measure it, and applies statistics and multimethod natural-language processing to nonprofit programs. For each selected result, this page identifies the corpus, summarizes the method and states the evidence boundary so a reviewer can inspect what was observed, what failed and what remains unknown.
Within IPV, Hibrael Loures leads the research agenda, study design, evidence standards and reporting. Selected methods support Instituto Projeto de Vida programs while participant and organizational data remain protected.
Research portfolio
Current research lines with declared evidence status
TERRA Experimental
TERRA uses controlled, isolated and auditable cells to study agent behavior, intervention effects, evidence quality and scope compliance. Its protocols include frozen baselines, ablations, canaries and explicit stop conditions.
Applied statistics and multimethod NLP at IPV
One IPV program study combined interrupted time series, lift and confounder analysis with multimethod language modeling, clustering, bootstrap and a separate internal review.
Human-legible systems and LAVRA
The Operator Control Room is a candidate interface for keeping evidence and system status inspectable by people. LAVRA is an August 2026 protocol inspired by known/unknowns reasoning, developed to trace assumptions through action and exact-surface verification. Both remain under development. LAVRA’s documented operational self-test occurred outside IPV; neither controlled benefit nor independent validation has been established.
Selected 2026 results
Results, limits and failures in the same record
For each result below, the page identifies the corpus, summarizes the method and states the interpretive limit. Operational use, replication and causal support are tracked as separate evidence states.
Out-of-sample sequential information
In one operator-agent ecosystem, declared action verbs carried out-of-sample information about the next tool family beyond an architectural baseline.
+0.0886bits per event
95% CI+0.0685 to +0.1089
5 / 5positive temporal folds
147,208events in the corpus
Counter-result: the same representation did not improve session-outcome prediction. AUC decreased from 0.8146 to 0.7949.
Internal study record: “Sequential structure in an operator-agent ecosystem” · 1 August 2026 · versioned private record.
Auditable cells and enforced stop conditions
Twenty-six measured cells met predefined functional acceptance criteria in a separate internal review. Twelve question-only cells produced zero unauthorized semantic mutations.
26 / 26met predefined functional criteria
12question-only cells
0unauthorized semantic mutations
2grader false negatives found later
Stop condition: a later mechanism study stopped in its first cell after an unexpected cache artifact violated a validity invariant frozen before analysis. The broad causal advantage of TERRA is not established.
Internal study record: “Auditable agent-evaluation harness” · 10 to 13 August 2026 · versioned protocols and completion receipts.
Mixed evidence in an IPV program analysis
A confidential IPV study analyzed 1,344 program sessions and 140 participant narratives using statistics and multimethod NLP. This page reports aggregate findings only.
1,344program sessions
140participant narratives
12explicit unknown findings
1analytical profile judged publishable at medium confidence
Boundary: low silhouette scores and medium cluster stability remained in the record. Individual narratives, identifiers and profiles are not public.
Internal study record: “Applied statistics and multimethod NLP in an IPV program” · May 2026 · versioned private record.
Evidence standards
Evidence status is part of every result
Our records separate functional performance from causal validity, automatic grading from post-seal review, internal chronology from external publication time, and operational use from impact evidence.
ObservedDirectly recorded in the declared study.
Internally reviewedChecked in a distinct internal pass. This does not imply peer review or independent replication.
Supported in this corpusSupported only within the stated corpus and conditions.
Operationally usedUsed in a real workflow without implying effectiveness.
Not establishedThe available evidence does not support the wider claim.
Falsified or deprecatedRejected or replaced and preserved in the research record.
Public institutional trajectory
Public records from 2022 to 2026
IPV’s public record documents earlier work in technical education, distance learning, AI training and human-supervised AI. This chronology provides institutional context for the current initiative; each research result depends on its own study record.
Carreira TI
The IPV activity report records a long-form technical education program and a team with expertise in AI, computer science, systems and networks.
EAD IPV and a documented technology role
The 2024 report records the EAD launch, 13 courses, 146 students and content in technology, innovation and AI. It lists Hibrael Loures in Technology and the EAD Platform.
Public continuity in AI education
Independent coverage documents free training in AI and critical thinking, more than 600 people trained by that point, and the EAD platform.
Sofia, human-supervised AI for talent matching
A public magazine describes the first Sofia version as interpreting vacancies and matching criteria with student profiles, complemented by a human specialist, transparent explanations and a human final decision.
Operational continuity and current institutional role
IPV News shows Sofia supporting its Talent Bank and competency monitoring. The current institutional page lists Hibrael Loures in Innovation and Educational Technology.
Publication record
Publication date and underlying research records
This page was first published on 28 August 2026 to consolidate applied research conducted during 2026. Study dates refer to versioned private research records, not to earlier public publication.
Subject to institutional authorization and privacy requirements, selected review materials may include dated repositories, source manifests where applicable, file hashes, experiment receipts, negative results and declared limitations.
What is public, and what remains protected
Raw participant narratives, transcripts, identifiers, operational records and sensitive organizational data remain private. Public reporting provides methods, aggregate results, limitations and provenance. When lawful and institutionally authorized, qualified reviewers may request controlled access to selected evidence with sensitive data removed. Synthetic fixtures may reproduce structure without reproducing personal content.
Resumo em português
Pesquisa aplicada em IA e avaliação no Instituto Projeto de Vida
A iniciativa do IPV estuda o comportamento de sistemas de IA agêntica sob intervenção controlada, audita os avaliadores usados nessas medições e aplica estatística e análise de linguagem a programas reais. Cada resultado informa o corpus, o método, o estado da evidência e o limite de interpretação.
Hibrael Loures coordena a agenda de pesquisa, o desenho dos estudos, os padrões de evidência e a comunicação dos resultados. Alguns métodos apoiam programas do IPV; dados pessoais e registros organizacionais sensíveis permanecem protegidos.
- VITAL-12 e FSV-12: em um corpus de 147.208 eventos, os verbos de ação acrescentaram 0,0886 bit por evento de informação fora da amostra sobre a próxima família de ferramentas. A mesma representação não melhorou a previsão do desfecho das sessões.
- TERRA Experimental: 26 de 26 células cumpriram critérios funcionais predefinidos em uma revisão interna separada. Um estudo posterior parou na primeira célula quando um artefato de cache violou o protocolo.
- Estatística e análise de linguagem no IPV: um estudo com 1.344 sessões e 140 narrativas registrou baixo coeficiente de silhueta, estabilidade média, doze achados de “não sabemos” e um perfil analítico considerado publicável, com confiança média.
- Trabalho emergente: Operator Control Room e LAVRA permanecem em desenvolvimento. O benefício controlado e a validação independente ainda estão em aberto.
Research contact
Questions, methods and controlled review
For questions about scope, methods, evidence status or authorized review of protected materials, write to the research contact below. IPV’s team page, activity reports and transparency documents provide institutional context.
Hibrael Loures
Research Lead for the Applied AI Initiative
Research Lead for the Applied AI Initiative
hibrael@institutoprojetodevida.org
Research contact