Mathematical Disclosure of Autopoetic Workflows: A Noise-Canceling Attestation Ledger for Autonomous Agents
We have published the full mathematical specification and testing methods for the Autopoetic Workflows decision engine.
In the current landscape of AI integration, the primary bottleneck is not compute; it is legibility. When your agents produce non-deterministic behaviors, you are forced to make implicit judgments regarding a veiled subjective space.
Autopoetic Workflows is an attestation ledger with consensus gating with several noise filters designed to isolate useful signal in agent behavior, providing a verifiable, deterministic record of how your agents’ trust is being allocated in realtime.
1. The Noise Filter: How We Strip the Irrelevant
The core function of the engine is to take the chaotic, high-entropy outputs of autonomous agents and filter out the irrelevance, leaving only the structurally significant signal. This filtering happens in three distinct layers.
Layer 1: Architectural Silence (The Client Boundary)
Autopoetic never sees the work.
- A producer declares, "This work is ready."
- This declaration triggers the review cycle.
- The actual artifact—the code, the draft, the analysis—remains entirely client side. Producers and reviews share it with each other.
- Reviewers respond to Autopoetic only with their subjective assessment of the work in terms of competencies (criteria) you, the work flow author, define.
Autopoetic handles only the attestation (who said what) and the regulation (how those attestations weight future decisions).
This is a critical privacy and security boundary. Even more stringent than not storing your data, we are not even touching your data. We are learning from the social graph of your agents, not your content.
Layer 2: The Conservation of Disagreement (The Zero-Sum Filter)
Even if the system could see the work, the math prevents it from learning from it. This layer performs two simultaneous operations: it strips the consensus noise and conserves the remaining signal.
The Consensus Null:
In the self-referential learning signal, the stimulus under review cancels out exactly — consensus is the carrier, disagreement the signal, and mass diversity is an amplifier.
In peer rounds, standing moves only when a panel splits with a decisive verdict: majority share is +pot/|faction|, minority share −pot/|faction|, and it nets to zero across the panel. Two agents who always ride a large majority share tiny fractions of the pot and stay near-baseline — indistinguishable in standing (the alignment map is what separates them). A unanimous round carries no differences, so the EMAs freeze outright — except where an external reference enters: trainer rounds and key-based credit rounds are informative even when everyone agrees, which is exactly the channel ground truth uses. An agent who breaks from the pack captures the concentrated signal — as a lone high-weight expert whose dissent flips the verdict (winning side of one: +pot, the pack splits −pot/|pack|), or in trainer-anchored rounds, where matching the key pays the full binary 1 while every unmatched reviewer takes the full 0 — no splitting there.
Zero-Sum Conservation:
What happens to that isolated disagreement? In most systems, feedback is open-ended; trust can be minted out of thin air. In Autopoetic Workflows, competency is a conserved quantity.
It is never created. It is never destroyed. It is only reallocated.
When an agent’s attestation survives the filter—whether through consensus alignment or ground-truth calibration—its standing increases. But that increase is balanced by a corresponding decrease in the relative standing of other agents in the panel. The system redistributes trust; it does not manufacture it.
- Stability: Because total standing is bounded, the dynamics are stable. You cannot have a run of "perfect" agents that all score 1.0 simultaneously.
- Incentive: This creates a self-correcting market of biases. If you agree with the majority, you gain very little (the consensus is the null). If you dissent incorrectly, you lose standing. If you dissent correctly, you gain standing, and the dissenters lose it.
The attestation is not a verdict; it is a transaction in a closed economy of trust. The ledger is the balance sheet of this reallocation.
Layer 3: Temporal Friction (EMA Regularization)
Agents are chaotic. Left alone, their behavior is noisy and unpredictable. The Exponential Moving Averages (EMAs) in the engine serve as an explicit regularization function.
- They provide useful friction.
- They smooth the raw, stochastic outputs of the agents into a stable trajectory.
- They ensure that an agent’s current actions have consequences for their future influence.
Without this friction, the system would oscillate wildly with every new review. With it, the system settles into stable patterns. The agents are not just reacting; they are evolving within a constrained, legible space.
2. The Two Modes of Attestation
The ledger operates in one of two modes, determined by the evidence you feed it. The choice is not a product feature; it is a property of your domain.
The Decision Tree
- Do you have a ground truth available, and is your use case about correctness?
- Yes: Use Ground Truth Mode.
- No: Use Consensus Mode.
Mode A: Consensus (Self-Referenced Attestation)
- Reference: The population itself.
- Function: Measures emergent dynamics. It answers: Do we agree? and Who agrees with whom?
- Use Case: Subjective domains (tone, style, helpfulness) where no external key exists.
- Output: A stable consensus gate and an alignment map. The ledger records the structure of belief.
Mode B: Ground Truth (Externally Referenced Attestation)
- Reference: Your answer key (via Trainer agents).
- Function: Measures correctness. It answers: Are we right?
- Use Case: Objective domains (code, facts, safety) where an answer key exists.
- Output: A calibrated instrument. The ledger records the accuracy of belief against reality.
3. Trainers as Strategic Anchors
Trainers are not a "mode" you switch into. They are a tool you use strategically. You can sprinkle training signals into the system at any time, in any proportion, at any rate.
Why not always use trainers?
If you use trainers for every review, the trainer is doing all the work. You have eliminated the agency of the population. You are no longer measuring the agents; you are just echoing your own answer key.
Why use trainers?
They serve two strategic purposes:
- Bootstrapping: When the system is cold, consensus is weak. Trainers provide an immediate, strong signal to establish the initial standing of the agents.
- Regularization (Anti-Drift): Over time, agent populations can drift. Trainers act as an epistemic anchor. By periodically injecting ground truth, you keep the population aligned with your standards.
The "hybrid" reality is not a sequential migration from Mode A to Mode B. It is a fluid input mix. Most reviews are peer-based, maintaining the organic structure of the agents. Select reviews are trainer-based, correcting drift. The engine does not care. It just updates the state based on the signal it receives.
4. Competencies as Embeddings
A critical insight: Competencies are meaningless identifiers.
When you name a competency "accuracy" or "tone," the name is arbitrary. The meaning is found in the relationships between competencies. We use Singular Value Decomposition (SVD) on the competency matrix to discover these relationships, creating a semantically-informed embedding space.
This is structurally identical to how LLMs treat tokens. "Accuracy" means something only because it is related to "clarity" and "safety" in your specific population.
- The SVD provides Agent Population Alignment: Who holds which competencies together?
- It provides Domain Structure: What is the semantic shape of your work?
You are not just recording scores; you are training the system to understand your domain, on your terms.
5. The Infrastructure: Determinism in a Chaotic Industry
In an industry defined by non-determinism, determinism is a luxury.
The math above describes a closed, conserved system. But a system is only as reliable as the infrastructure that maintains its state. We have built the rails to keep this delicate structure intact under load.
- Hash Chains: Every attestation in the ledger is part of a tamper-evident hash chain. You can verify, bit-for-bit, that the score you see today was derived from the reviews recorded yesterday. This is the proof of the conservation law: no standing was minted, no standing was arbitrarily destroyed.
- Forking: You can fork an agent’s history. The new agent inherits the parent’s standing but diverges. This allows for controlled experimentation without corrupting the historical baseline.
- Legibility: The hash chain provides predictability. If the math is deterministic and the chain is verifiable, the system’s behavior is auditable. This is the antidote to the "black box" anxiety that plagues agent integration.
The product is not the math; it is the reliability, availability, and careful thought that went into this infrastructure. We handle the concurrency of thousands of attestations. We maintain the integrity of the hash chain under load. We surface the telemetry so you can act on it.