collective-intelligence

Ailin¹ Collective Intelligence

⭐ Star the repo and back a new, more collective and collaborative era of AI

TL;DR: Ailin¹ makes 103,224 AI models collaborate inside one collective model, coordinated through 32 strategies instead of routed to a single one. Structured diversity, independent reasoning, and a full decision audit trail on every request: more reliable, resilient, and auditable than any single-model integration, and proven against the frontier in the open.

→ Quickstart · See the evidence · Full docs

Thousands of AI models coordinate inside one collective model.

Structured diversity, independent reasoning, and full decision provenance on every request, designed to make outputs more reliable, more resilient, and more auditable than a single-model integration. Every day a new model launches claiming to be the best. This is the layer where they work together. Full documentation: ailin.guide. Quickstart · The next frontier · Why a collective · The evidence · Always at the frontier · How it works · Contributing · Docs

Collective intelligence: the next frontier of AI

The AI industry has been focused on building bigger individual models. Ailin¹ takes a complementary approach: a collective of 103,224 AI models (live production count, 2026-08) that can collaborate, debate, critique, and synthesize together, applying structured diversity to problems where a single model is a single point of training, architecture, bias, and failure.

This is not multi-model routing. This is not an API gateway. This is Collective Intelligence: a system where models from every major architecture (frontier APIs, open-weight challengers, and our own model family) coordinate through dozens of strategies, with the goal of higher reliability, broader evaluation coverage, and fuller auditability than any single-model integration provides.

The principle is grounded in research on collective intelligence and cognitive diversity: Hong & Page's "diversity trumps ability" result and Woolley et al.'s work on collective performance (see the public Bibliography). Ailin¹ applies that principle as an engineering platform: a discovery engine that indexes 103,224 models, dozens of coordination strategies, an audit substrate that records every coordination decision, and a closed-loop training pipeline. Some of these layers are production-grade today and others are still maturing; the docs carry status badges so you always know what is shipping versus what is on the roadmap.

Why a collective beats the biggest single model

Frontier models keep getting bigger, and the strongest single model at any moment is remarkable. But a single model is always a single point of training, architecture, failure, and bias. A well-coordinated collective addresses each of those structural limits in a way that scale alone cannot.

The effect compounds. These are not six independent features; they are six facets of a single structural choice: coordinate many models well, and the result is more reliable, more governable, more durable, and, on the expanding set of tasks where correctness can be objectively verified, measurably more accurate than every frontier flagship we tested (97% vs 68–82%, receipts below).

Proven against the frontier, in the open

We test the thesis against ourselves, publicly, with objective grading: pinned judges, machine-checkable answers wherever a task allows one, and the raw per-execution data committed to this repository (full report · raw CSVs + scripts · regenerate every table yourself).

✅ Validated: the collective beats every frontier flagship on verifiable tasks.

• 97% objective accuracy (37/38) vs. 68–82% pooled for GPT-5.5-pro, Claude Opus 4.8, Gemini 3.1 Pro, and Grok 4.3
• Across every run, the verifier never selected an objectively wrong answer
• A pool of sub-frontier open-weight models, coordinated well, out-answered every flagship on the same tasks (leaderboard with every n and caveat, §3)

The current frontier of the thesis (measured honestly, driving the roadmap):

Every number above is backed by the raw per-execution data and reproducible scripts committed in this repository: run the harness yourself, on your own workload, and hold us to it.

Tens of thousands of models, always at the frontier

The Ailin¹ collective does not depend on hardcoded model lists or manual provider integrations. A continuous discovery engine scans the global AI ecosystem and automatically absorbs new models as they are released.

The result: a live collective of 103,224 models across ~90 provider integrations that stays current with the ecosystem. When a new model is published by a discovered source, the discovery engine absorbs it without code changes, configuration, or downtime.

Semantic discovery, zero hardcoded models

The discovery engine scans dozens of sources in parallel:

• Native provider APIs
• Cloud hubs
• Model aggregators
• Open-model repositories
• Private inference endpoints

But the sources aren't the point, how models are selected is.

Every discovered model is analyzed and indexed automatically (no manual mapping) across: capabilities, performance profile, pricing, context window, modalities, architecture. Routes are health-gated, a model is advertised only after being proven live.

Model selection is fully semantic. When a request arrives, the collective does not pick from a static list. It assembles the ideal team of models based on the task's requirements, the chosen strategy, and the desired outcome profile (maximum quality, best cost-benefit, lowest cost, fastest response). The right models are elected in real time, for every single request. When tomorrow's "best model ever" launches, the collective absorbs it; it doesn't compete with it.

Own models in the same arena

The ailin model family and its training flywheel are part of the design: coordinator checkpoints trained on the engine's own coordination traffic, competing in the same pool as every third-party model, no routing privilege. The audit substrate ships today; production coordinator weights are still in development (honest status, always current).

Collective strategies as falsifiable hypotheses

32 registered strategies (consensus with convergence floors, blind debate, expert panels, devil's-advocate consensus, cost-cascade, best-of-N with objective verification), each labeled with honest reachability (auto-selectable / explicit-only / roadmap), each falsifiable by the experiment harness in this repo. Strategies earn their place with evidence, or lose it.

Multimodal + deterministic file generation

Multimodal generation (images, audio, video) routed by capability, plus deterministic file rendering (DOCX, XLSX, PDF, PPTX, ZIP, code) from any structured-output chat model, proven in production.

Governance that enterprises actually need

The same audit trail that proves our benchmark claims governs your production traffic: governance as first-class principle, not overhead.

Architecture at a glance

The system, end to end. Discovery feeds team assembly, every execution path converges on the same provenance-generating arbitration step:

…

In text: a request enters through the OpenAI-compatible API from any OpenAI SDK or curl client (only the base_url changes). Strategy resolution applies the ailin-auto conservative cascade and hands off to team assembly, which does semantic selection over the live model catalog fed continuously by the discovery engine (health-gated, zero hardcoded models). The assembled team runs in execution, which manages fallback chains and a budget governor, talking bidirectionally to ~90 provider integrations. Execution's output goes to arbitration, which applies quality gates and the deterministic verifier, producing the final response with full provenance (ailin_metadata).

How a request flows

Zoomed in on one request, which of the three paths above it takes, and why:

flowchart LR
    A[OpenAI-compatible request] --> B{Strategy resolution
ailin-auto cascade}
    B -->|simpl