01
Semantic Intake
Universal, domain, and enterprise semantics shape the Query Contract before the first reasoning pass.
BEYOND LLM + RAG + VECTOR SEARCH + AGENTS
AI that measures uncertainty,
investigates what matters next,
adapts to new evidence—
and knows when it has learned enough to decide.
POLOXI.ai is a mathematical and information-theoretic framework that transforms LLMs, RAG, vector search, knowledge graphs, enterprise data, APIs, multimodal evidence, rules, and agents into a recurrent evidence-adaptive reasoning system.
MODEL-AGNOSTIC•RETRIEVAL-AGNOSTIC•EVIDENCE-GROUNDED•AUDITABLE
POLOXI.ai REASONING STATE ITERATION 04
QUERY CONTRACTWhich conclusion is best supported—and what evidence could still change it?
REASON→RETRIEVE→MEASURE→INVESTIGATE
UNCERTAINTY Hₜ0.42
INFO GAIN IGₜ+0.18
01 Candidate A0.84
02 Candidate B0.71
03 Candidate C0.46
HIGHEST INFORMATION VALUEInvestigate the evidence gap most capable of changing the rankingIV = 0.31 →
RECOMPUTE AFTER NEW EVIDENCE
Repeat only while another investigation is justified
Repeat only while another investigation is justified
ONE ADAPTIVE REASONING CORE. MANY AI CAPABILITIES.
LLMsRAGVector SearchKnowledge GraphsEnterprise DataAgents
THE TECHNOLOGY
From generated answers to
evidence-adaptive investigation.
A conventional AI stack may still reach the same final answer. POLOXI.ai is designed to govern the harder part: whether the answer has been investigated enough to justify it. The framework measures uncertainty, identifies the next evidence gap with the highest decision value, allows new evidence to reshape the candidate space, recomputes the competitive state, and stops only when governed convergence criteria are satisfied.
THE POLOXI.ai ADAPTIVE LOOP
Et → Bt → Ct → Mt → St → Ht → IVt → Investigate → Et+1 → Recompute
Repeat while another investigation is justified. Evidence changes understanding. Understanding can change the candidate space. The candidate space changes what POLOXI.ai investigates next.
02
Dynamic Discovery
Candidate answers and hypotheses can emerge, merge, disappear, or re-enter as evidence changes.
03
Measured Uncertainty
Shannon entropy quantifies the current competitive uncertainty instead of hiding it behind prose.
04
Information Value
POLOXI.ai selects the unresolved branch predicted to provide the greatest decision value next.
05
Candidate Competition
Evidence support is evaluated across branches with deterministic weighting, coverage, constraints, and identity normalization.
06
Governed Convergence
The system can investigate again, answer, escalate, abstain, or authorize a downstream action based on governed criteria.
THE DIFFERENCE
The question is not merely
“Can AI produce an answer?”
Has AI investigated enough of the right evidence to justify that answer?
CONVENTIONAL AI STACK
Query → Retrieve → Generate / Rank → Answer
LLM + semantics + DB/API/RAG can be excellent for well-defined or sufficiently covered problems. But uncertainty, candidate discovery, evidence gaps, and stopping can remain implicit.
VS
POLOXI.ai
Discover → Compete → Measure → Investigate → Learn → Rediscover → Recompete
A recurrent stateful process where the answer space, evidence, and leading conclusion can change during investigation—and each change is attributable.
“The answer is not assumed. It emerges through evidence.”
THE ENTERPRISE AI STACK
Turn individual AI capabilities into
one governed investigation system.
POLOXI.ai operates across probabilistic and deterministic capabilities. LLMs interpret language. Vector and RAG systems retrieve. SQL and APIs provide authoritative facts. Knowledge graphs expose relationships. Rules enforce business constraints. Agents execute tools. POLOXI.ai governs how those capabilities participate in an evidence-changing reasoning loop.
LLMsSemantic interpretation
RAGDocument retrieval
Vector SearchSemantic similarity
Knowledge GraphsEntities & relationships
SQL / APIsAuthoritative facts
RulesConstraints & calculations
Multimodal AIImage / audio / video evidence
AgentsTools & execution
POLOXI.aiMeasure uncertainty → select high-value investigation → acquire evidence → recompute → govern convergence
REAL-WORLD MARKET IMPACT
Problems where the evidence itself
can change the answer.
These examples intentionally go beyond simple retrieval. A one-pass LLM or fixed RAG retrieval may produce a plausible first answer; POLOXI.ai is designed for the subsequent investigation when unresolved evidence can change the candidate set, ranking, hypothesis, or final decision.
01
Insurance
Complex problem
A $12M commercial-property renewal has three viable placement strategies. The incumbent is +18% with a restrictive water-damage exclusion. Carrier B is +11% but previously flagged roof condition. A layered placement is +14% and adds complexity. Loss history is acceptable, but inspection, CAT exposure, underwriting notes, remediation status, and coverage terms conflict. Which strategy should be recommended, and is there unresolved evidence capable of changing the winner?
Worked POLOXI.ai solution
POLOXI.ai normalizes the three strategies and evaluates coverage fit, appetite, losses, CAT exposure, property condition, exclusions, premium, and underwriting evidence. Carrier B leads initially, but roof-remediation uncertainty keeps Hₜ above the stopping threshold. Roof evidence has the highest IVₜ, so the latest inspection and underwriter note are retrieved. They confirm remediation and removal of the condition. Eₜ₊₁ is updated, IGₜ is measured, and all strategies are recompeted. The system converges on Carrier B, retains layered placement as fallback, and records the evidence that resolved the uncertainty.
02
Finance
Complex problem
Five U.S. companies appear attractive for a 10-year investment, but valuation, growth, profitability, balance-sheet strength, competitive moat, AI disruption, regulation, management execution, and geopolitical exposure conflict. Candidate A leads the first ranking. Is A actually the strongest evidence-supported choice, or is there unresolved evidence that should change the ranking?
Worked POLOXI.ai solution
POLOXI.ai normalizes the candidates and constructs Candidate × Branch competition. Candidate A initially leads. Regulatory exposure is weakly covered and has the highest IVₜ. POLOXI.ai investigates current filings and regulatory evidence and finds a material risk that weakens A. After Eₜ → Eₜ₊₁, the matrix is rebuilt and Candidate B becomes the leader. The result records that the winner changed because of newly grounded regulatory evidence—not model variation—and another loop occurs only if sufficient information value remains.
03
Research
Complex problem
Multiple high-quality studies reach conflicting conclusions about Treatment A. Which explanation best accounts for the disagreement, which conclusion currently has the strongest evidence, and what additional evidence would most reduce the remaining uncertainty?
Worked POLOXI.ai solution
POLOXI.ai forms competing hypotheses around population differences, methodology, dosage, study design, and outcome definitions. Hypothesis A initially has the strongest support, but entropy remains high because one subgroup is weakly represented. That subgroup becomes the highest-information research gap. Additional evidence is retrieved and grounded; Hypothesis B strengthens for the subgroup and changes the overall evidence distribution. POLOXI.ai reports the updated hypothesis state, information gained, unresolved uncertainty, and whether another research loop is justified.
04
Cybersecurity
Complex problem
What actually caused an enterprise security incident when identity logs, endpoint telemetry, network activity, cloud events, user behavior, and threat intelligence are incomplete or contradictory—and which evidence should investigators acquire next?
Worked POLOXI.ai solution
POLOXI.ai discovers competing hypotheses: credential compromise, malicious insider activity, compromised endpoint, and cloud-token abuse. Credential compromise initially leads, but endpoint process lineage has the highest candidate-discrimination value. New endpoint telemetry reveals abnormal token use from a compromised endpoint, weakening the initial explanation and strengthening an endpoint-compromise → token-theft chain. POLOXI.ai recompetes, tests remaining uncertainty, and either selects the next investigation or converges on the supported incident chain for analyst review.
SEMANTIC FABRIC
Meaning begins at intake.
Grounding continues through every loop.
POLOXI.ai combines LLM semantic interpretation with governed semantic metadata. The LLM proposes meaning; Universal, Domain, and Enterprise semantic layers resolve canonical concepts, relationships, rules, metrics, permissions, and sources of truth.
UNIVERSAL COREPerson · Organization · Document · Event · Risk · Decision
DOMAIN PACKVocabulary · Ontology · Entities · Relationships · Rule templates
ENTERPRISE LAYERLocal aliases · Data mappings · Rules · Metrics · Ownership · Permissions
GROUNDING REGISTRYSQL · APIs · Documents · RAG · Search · External sources
NSF SBIR PHASE I — RESEARCH DIRECTION
Test whether the loop
actually improves the outcome.
POLOXI.ai does not assume that conventional AI cannot reach the same answer. Phase I validation is designed to test whether recurrent evidence governance reaches evidence-supported answers more reliably, reproducibly, explainably, and information-efficiently—especially when interpretation, evidence, or candidate space changes during investigation.
Research & Collaboration →01Candidate Identity StabilityMeasure normalization, duplicate fragmentation, erroneous merges, and ranking effects.
02Dynamic Candidate DiscoveryMeasure whether evidence appropriately expands or contracts the answer space, including late-winner discovery.
03Repeated-Run StabilityCompare winners, Top-K overlap, entropy variance, evidence coverage, and stopping consistency.
04Model DependenceBenchmark lower-cost and premium models under identical POLOXI.ai governance.
05Information EfficiencyMeasure uncertainty reduction, information gain, retrievals, tokens, latency, and cost.
LLM + Semantics + DB/API/RAGVSLLM + Semantics + DB/API/RAG + POLOXI.ai
AGENTIC AI DECISION GOVERNANCE
Agents can execute.
POLOXI.ai governs what deserves execution.
POLOXI.ai can use information value to select the next tool or specialized agent, incorporate the returned evidence, recompute the decision state, and determine whether to investigate again, escalate, answer, abstain, or permit an authorized action.
Goal→POLOXI.ai→Agent / Tool→Evidence→Recompute→Governed Action
COMPANY & TEAM
Built from enterprise
software experience.
POLOXI.ai was conceived and developed by Nelson Castillon, B.S. Computer Science, with more than 26 years of professional software and technology experience spanning Insurance, Healthcare, E-Commerce, Marketing Campaigns, Digital Advertising, and Custom Software development.
The architecture grew from a practical research question: can probabilistic AI remain flexible while an independent mathematical layer governs semantics, evidence, uncertainty, investigation, competition, iteration, and convergence?
26+Years software
experience
experience
6+Core industry
domains
domains
1Adaptive reasoning
framework
framework
POLOXI.ai
Investigate enough of the
right evidence.
Mathematical reasoning infrastructure for enterprise, research, and agentic AI.
research@poloxi.ai →Replace this placeholder address with your production contact before launch.