Eitan Lavi · Quorumz Research
Transcript · Synthetic English narration · Based on the Quorumz whitepaper and September 2026 research manuscripts
PRODUCT OPPORTUNITY
Personal AI can learn from experience and feedback that never reach a central model. HiNet proposes connecting these human-bound specialists through cooperation, while owners retain their private records and authority. The opportunity depends on turning that local experience into useful, permitted service.
Source: Abstract
ARCHITECTURE CONTRIBUTION
Our proposed expert is a complete personal agent. A2RL develops its owner's preferred behavior, writing, task completion and skills. A separate cooperation learner develops that iCore's choices about help. Ordinary mixtures already learn experts and gates; our distinction is independently governed personal learning and recursive task cooperation. The framework uses established methods and states the limits of its own claims.
Source: 2 The personal core
PROPOSED OWNER LEARNING
The owner teaches the personal core through permitted corrections and outcomes. Completed collaborations teach a different model which peers to call, when to branch and when to stop. Proposed nightly cycles train versioned candidates and check the combined pair before release. Even individually improved components can fail together. Learning remains subject to independent owner permissions.
Source: 8 Nightly consolidation
LOCAL RECURSIVE ROUTING
Each iCore selects from its own permitted relationships and referrals. A colleague can suggest another willing specialist. The request moves through successive local decisions, with no required central expertise catalog. Natural quorums offer a starting point; evidence of competence determines whether those connections are useful.
Source: 4.1 Forward split join and return; 4.5 When natural quorums reduce search cost
RECURSIVE COOPERATION
A question can travel along a sequence of referrals, or split into parallel work by a temporary quorum. Each invited iCore can make the same decision again with its own learned policy. Replies return through the recorded calls for synthesis. No single router chooses every step. Budgets, deadlines and clear completion rules keep a collaboration bounded.
Source: 3 Connectivity
BOUNDED CALLS AND RETURNED ANSWERS
A question can branch into parallel contributions, then bring replies back for review. Each step has a local permission check and a share of the original budget. The execution follows causal events across a mesh of peers, not a permanent central hub. Enforcing budgets requires protected local accounting.
Source: 4.2 Conserving a query budget through local grants; 4.1 Forward split join and return
ARCHIVED SCORES / ORACLE CEILING
In our earlier specialist tests, the best single model scored about fifty-six percent and the best merge about sixty-eight. Perfect routing approached ninety-three percent. That is an ideal selection ceiling, not deployed performance. The gap motivates learning better cooperation while preserving specialists that interfere when merged.
E4 / SYNTHETIC EVIDENCE
The E4 study found a small-quorum gain in learned private knowledge. Three iCores answered about forty-five percent of joint questions, compared with about nineteen percent for the selected standalone. Adding irrelevant members reduced accuracy to about twenty-one percent at twenty members on those same questions. That failed the registered scaling threshold. It motivates selective cooperation that preserves existing competence, not a claim that size alone improves intelligence.
Source: Holding the original cases fixed
CONDITIONAL OWNER-FEEDBACK ADVANTAGE
Our finite construction proves an expected-reward advantage when owners provide more informative feedback than the central learner receives. Learned local routes make their expertise callable, with communication costs counted separately. It establishes one defensible advantage regime. It does not establish that our network beats frontier systems with equal information and resources.
Q1 / CONDITIONAL QUALITY BOUND
The new quality bound separates personal-core error from cooperation error. Missing the right expertise, misjudging a contribution, or poor selection can consume the specialists' advantage. The mathematical condition is explicit: available competence must exceed those losses relative to the chosen comparator. This is a conditional result about declared tasks, not a frontier-model benchmark.
Source: 5 A2RL quality
E1 TO E3 / CONDITIONAL EFFICIENCY
The efficiency case has two parts. Restrained recursive calling can keep expected active work bounded even as the available population grows, provided discovery and maintenance are also counted. Parallel required contributions can return sooner than the same work done serially, when capacity and links permit. Neither property alone proves lower total energy. The full cost must fit below the comparator at the required answer quality.
Source: 7 Energy
PLANNED ECONOMY
The planned marketplace rewards useful contributions rather than requiring owners to hand over their original private records. Experience can become a service, subject to permission and privacy checks. Payments and settlement still need implementation, and earnings depend on demand and demonstrated usefulness.
Source: 15. Economics & Membranes
CONDITIONAL ENFORCEMENT
Protected local controllers enforce permission independently of the AI. Owners can revoke future access, while remote permissions expire under stated conditions. Distribution alone guarantees neither privacy nor moral alignment, and previously disclosed information cannot be recalled. These boundaries are part of the design, not promises to solve through scale.
NEXT IMPLEMENTATION MILESTONE
The next milestone is a bounded, owner-authorized quorum: useful personal learning, reliable referrals, returned evidence, and every resource cost visible. A large network becomes valuable through callable expertise and trusted cooperation. Population size alone is not a capability metric.