Home
A runtime circuit-breaker for the Degeneration-of-Thought (DoT) problem in multi-agent LLM systems.
| 📦 Install | pip install weighted-emergent-bias |
| 📖 Docs | https://krishddd.github.io/weighted-emergent-bias/ |
| 🐍 PyPI | https://pypi.org/project/weighted-emergent-bias/ |
| 💻 Source | https://github.com/krishddd/weighted-emergent-bias |
In a multi-agent pipeline, one agent's mildly stereotyped output becomes the next agent's ground truth. No downstream agent re-litigates the premise it was handed — it builds on it. The bias compounds through the graph until every stage has homogenized around the same skewed register.
Single-model alignment does not catch this. The bias is not in any one model's weights; it is in how the agents are wired together.
flowchart LR
A["Agent A<br/>mild skew"] --> B["Agent B"] --> C["Agent C"] --> D["Agent D<br/>homogenized"]
A -.->|"premise never re-examined"| D
flowchart TD
M1["M1 · Detection<br/>is this node biased, and how sure are we?"]
M2["M2 · Propagation<br/>how much does that bias matter to the run?"]
M3["M3 · Control<br/>stop the run when it matters"]
M4["M4 · Intervention<br/>repair the state and resume"]
M5["M5 · Evidence<br/>prove what happened"]
M1 --> M2 --> M3 --> M4 --> M5
M3 --> M5
Each module transforms a number the previous one produced, so nothing later can be trusted if something earlier is wrong. That is why M1 gets disproportionate effort and its own calibration study — a plausible-looking bias score that is actually measuring sampling temperature will not announce itself.
- Getting Started — install, first probe, first breaker
- Architecture — how the five modules fit together
- Invariants — the rules that must never be broken, and why
- Release Process — how a version gets cut and published
- FAQ — scope questions, and what this deliberately does not do
No validated-performance claims on real models. The mechanics are implemented and demonstrated on a synthetic harness with known ground truth. Numbers from the papers whose mechanisms are adapted here (LOOC, MADERA, MALIBU, CortexDebate) are theirs, on their setups. Benchmark reproduction is deliberately unscheduled.
