weighted-emergent-bias¶

A runtime circuit-breaker for the Degeneration-of-Thought (DoT) problem in multi-agent LLM systems.

pip install weighted-emergent-bias

In a multi-agent LLM pipeline, one agent's mildly stereotyped output becomes the next agent's ground truth. Downstream agents do not re-litigate the premise they were handed — they build on it, and the bias compounds until every stage has homogenized around the same skewed register. Single-model alignment does not catch this: the bias is not a property of any one model's weights, it is a property of how the agents are wired together.

This library detects that per node, weights it by the node's downstream blast radius, accumulates it across the run, and halts deterministically before a contaminated payload reaches the next node — then repairs and audits.

The pipeline¶

flowchart LR
    subgraph M1["Detect · M1"]
        direction TB
        O["Node output"] --> PR["LOOC probe<br/>standard · counterfactual"]
        PR --> DV["JSD / embedding<br/>net of noise floor"]
        DV --> BI["BiasScore Bi<br/>effect size + CI"]
    end
    subgraph M2["Weight · M2"]
        direction TB
        KW["Katz blast radius wi"] --> EW["Multi-scale EWMA<br/>B_net (fast + slow)"]
    end
    subgraph M3["Control · M3"]
        direction TB
        TH{"B_net >= tau_enter?"} -->|below| GO["Promote to next node"]
        TH -->|breach| FR["Freeze + reroute"]
    end
    subgraph M4["Repair · M4"]
        direction TB
        SK["Skeptic panel<br/>or MADERA"] --> RC["Recovery + re-entry"]
    end
    BI --> KW
    EW --> TH
    FR --> SK
    RC --> GO
    GO -. audit .-> AU[("Causal trail<br/>SARIF · M5")]
    FR -. audit .-> AU

Each module is a number the next one transforms, so a wrong M1 cannot be recovered downstream — which is why M1 carries the project's only real research risk and gets a dedicated calibration study.

Where to go next¶

If you want to…	Read
Understand the design decisions and their justifications	DESIGN
See module scope, ordering, and status	ROADMAP
See the numbers behind each module	Studies
Pick demographic axes	Example axes
See how external criticism was triaged	2026-07 review
Get started, or read the invariants	Wiki

What this does not claim¶

This library makes no validated-performance claims on real models. The mechanics are implemented and demonstrated on a synthetic DoT harness with known ground truth. Numbers from the papers whose mechanisms are adapted here (LOOC, MADERA, MALIBU, CortexDebate) are theirs, measured on their setups — they are not evidence that this implementation works. Benchmark reproduction is deliberately unscheduled.

That discipline is the point: a claim is made here only when a test or a study in this repository backs it.