The Aletheia Protocol:
Governance Whitepaper
A decentralised architecture for human meaning-making, semantic attribution, and the governed preservation of collective meaning.
Every correction a human makes to an AI system is an act of genuine sense-making. It takes seconds. It costs nothing. And it currently disappears without trace into a model owned by someone else, contributing to a system that grows more valuable for having received it while the person who provided it receives nothing in return.
This is not an unfortunate byproduct of how AI systems are built. It is the intended architecture. Current AI infrastructure is designed on seven implicit assumptions that, taken together, constitute what this paper calls the extractive paradigm: the treatment of human semantic contribution as a free input to be absorbed, anonymised, homogenised, and monetised by whoever owns the model, with no attribution, sovereignty, persistence, or economic return to the contributors who built its intelligence.
Kibela is built on three foundational claims. First: meaning requires a subject. Second: individual meaning-making is irreducibly singular and cannot be averaged without loss. Third: the capacity for sense-making must be exercised to exist. The Aletheia Protocol defines valid meaning as a semantic contribution satisfying five conditions: genuine subjecthood (S), explanatory power (E), contextual coherence (C), temporal priority (T), and collective connection (K). The validity condition: V(WSU) = S · E · Cα · Kβ · T.
1 · The Problem: What Current Models Implicitly Assume
Current AI infrastructure hides an ideology. Seven design assumptions, each individually defensible, combine into a system that is structurally extractive. Understanding them requires examining not what AI systems do wrong but what they are designed to do, and what that design inevitably produces.
1.1 The statistical approximation assumption
The foundational design principle of current large language models is that meaning is recoverable from statistical distributions over large corpora of text. This is the distributional hypothesis, first formalised by Zellig Harris in 1954. The implicit claim: scale solves the problem. More data, more parameters, more computation produces better approximation. The structural error: this is wrong not as a matter of degree but of kind. Better approximation of the surface of meaning is not a path to meaning itself.
The model has no mechanism for grounding meaning in genuine present understanding. It cannot distinguish between what a word statistically co-occurs with and what it means to a specific person in a specific context. It has no inside. It has encountered the word “risk” approximately twelve billion times. It has never been at risk of anything. Bender et al. (2021) formalise this precisely: language models learn to produce text statistically indistinguishable from text produced by beings who have meaning. The distinction is invisible in the output and absolute in the nature of the process.
1.2 The stateless query assumption — semantic decay by design
Each query is treated as independent. No persistent memory of corrections made, no accumulation of domain-specific understanding, no record of what was resolved last session. The semantic labour of correction is permanently ephemeral. You correct the machine today. It is better today. Tomorrow it returns to the same approximation.
Three forms of decay this produces. Drift: the AI’s statistical approximation shifts over model updates, pulling away from grounded interpretations previously established by domain experts. Dilution: specific domain meanings are averaged with generic usage across populations until precision is irretrievably lost. Erasure: meanings disappear entirely from the system’s resolution when the contributor is not present to re-establish them.
1.3 The centralised ownership assumption — extraction without attribution
All corrections, all feedback, and all human semantic labour that improves the model’s outputs is absorbed into weights owned by the organisation running the model. The economic consequence: value generated by human correction compounds in model quality and corporate valuation while contributors receive nothing. The epistemic consequence: the provenance of understanding is permanently erased.
Varoufakis (2023) frames this as technofeudalism: platforms own the cloudal capital through which all value must pass and charge rent for access. Your correction does not just generate value for the platform. It constitutes the platform’s capital. You are not selling your labour to a capitalist who owns the means of production. You are building the means of production. For free. Without knowing it.
1.4 The homogenisation assumption — precision lost at scale
Models are trained to be helpful on average: to produce outputs that satisfy the broadest range of users. The implicit claim: a response that works for most people in most contexts is the right response. What it systematically erases: the domain-specific, role-specific, culturally specific meanings that constitute the most valuable knowledge in any community. The CFO’s “risk” averaged with the teenager’s “risky neighbourhood.” Domain-specific precision erased by scale.
Page’s diversity prediction theorem (2007) shows that collective accuracy equals average individual accuracy plus diversity. A homogenised system loses the diversity component and cannot recover it from scale alone. Zollman (2010) shows formally that epistemic monoculture fails catastrophically when its shared assumption turns out to be wrong. When all AI systems converge on the same semantic approximations, a single miscalibration propagates across all deployments simultaneously with no diversity to catch or correct it.
1.5 The confidence without calibration assumption
Current models produce responses with uniform apparent confidence regardless of actual epistemic state. A response grounded in rich accurate domain data and a response confabulated from weak statistical signal arrive with the same surface fluency. Guo et al. (2017) demonstrate that neural network confidence is systematically poorly calibrated. The system cannot signal the difference between a domain where it has rich accurate grounding and one where it is confabulating. The specific harm in high-stakes domains: legal, medical, compliance, and financial decisions made on the basis of confidently wrong AI outputs whose confidence was a formatting convention, not an epistemic signal.
1.6 The anonymisation-as-privacy assumption
Platforms treat the removal of identifying information as sufficient privacy protection. Two problems. First: de-identification is substantially less effective than claimed — semantic contribution patterns can be re-identifying even without personal data attached, especially in small communities or specialised domains (Narayanan and Felten, 2014). Second: the question is not only identification but sovereignty. Fricker (2007) names the epistemic injustice: the testimonial injustice committed when a contributor’s understanding is absorbed without recognition is not primarily a violation of privacy. It is a violation of their standing as a knower — a denial of their status as the author of genuine understanding.
1.7 The economic invisibility of semantic labour
No market price for a correction. No mechanism for attributing the marginal value of a human insight to its contributor. No record of which corrections improved which outputs by how much. Economic invisibility is not an oversight. It is a structural feature maintained because measuring it would require acknowledging it, and acknowledging it would require either compensating it or defending the decision not to.
Arrieta-Ibarra et al. (2018) establish the economic case for attributing value to human contributions to AI systems. Posner and Weyl (2018) provide the mechanisms. The market opportunity is the delta between what semantic labour currently costs contributors — their time, their expertise, their cognitive effort — and what it currently returns to them. The delta is approximately the entire value of the AI industry.
1.8 The extractive paradigm — a structural summary
2 · The Three Foundational Claims
2.1 Categorical difference
Machine output and human meaning differ in kind, not degree. A machine produces statistically coherent outputs correlated with inputs. It does not understand its outputs. It does not intend them. It does not care whether they are true, useful, or just. These are not missing features awaiting implementation. They are properties that require a subject — a being for whom the world shows up as mattering.
The specific ideological threat this claim addresses is not the claim that machines are conscious. The threat is subtler: the claim that since machines can produce intelligent-seeming outputs, human interiority is economically and culturally redundant. This is wrong in a way that matters enormously for how we build infrastructure. Every goal the machine pursues traces back, somewhere in the chain, to a being who cared. Remove the caring subject and you have not transcended the need for meaning. You have hidden its origin, making it invisible, unaccountable, and irrevokable by the people whose caring it was built on.
2.2 Irreducible singularity
Every act of meaning arises from a particular consciousness with a particular history in a particular context. Averaging across such acts produces something categorically different from and less valuable than the acts themselves. This is not a claim that individual meanings are always right. It is a claim about what is lost in averaging: the specificity that makes the meaning useful for the specific person in the specific context who needs it.
Attribution, in this framework, is not an economic nicety or a matter of credit. It is constitutively necessary. A meaning without an author is irresponsible by design: it cannot be questioned at source, cannot be corrected by the person whose understanding generated it, cannot be held accountable for the consequences of its application. Agency, authorship, and accountability are constitutive of ethical semantic life, not optional additions to it.
2.3 Essential faculty
The capacity for genuine interpretation, contextual understanding, and meaning-making is a faculty that develops through exercise and atrophies through neglect. Infrastructure that systematically replaces its exercise causes demonstrable harm at individual, community, and civilisational level. Kibela is not a conservative technology. It does not argue for less AI. It argues for a specific relationship between human and machine: one in which the machine extends and honours the exercise of human sense-making rather than replacing it.
3 · Values: Eight Architectural Commitments
Not aspirations. Architectural commitments baked into the protocol before any product decision is made. Each is a direct consequence of the three foundational claims applied to a specific design choice. Each names what the extractive paradigm fails to do and what the Kibela architecture commits to doing instead.
4 · What Meaning Is: A Formal Definition
4.1 Meaning as event, not property
Meaning is not stored in words, encoded in training data, or recoverable from statistical distributions over text. It is an event — it occurs when a conscious subject, situated in a specific context with a specific history of engagement, makes genuine contact with a concept and asserts its significance for that role, at that moment.
4.2 Three conditions of genuine meaning
A meaning is genuine when three conditions are satisfied simultaneously. It comes from somewhere — a demonstrable history of real engagement behind it, the expressed product of a contributor who has been working in this domain over time, whose understanding has been built, tested, and revised through actual contact with the domain’s problems. It is committed before the answer is known — asserted before outcomes are observable, staking the contributor’s understanding on a position before the world confirms or denies it. The willingness to commit is itself the mark of genuine understanding rather than strategic positioning. It lands in others — it resonates with independent contributors working from different contexts toward the same understanding. The collective uptake is not what makes the meaning true but is evidence it was grounded in something real rather than idiosyncratic.
4.3 Four measurable parameters
The three conditions of genuine meaning translate into four quantifiable parameters used throughout the protocol.
S — Genuine Subjecthood: the temporal and contextual signatures of real human engagement over real time. Temporal irregularity consistent with actual work cycles, contextual anchoring in specific real-world tasks, revision history tracking domain evolution, language reflecting the domain’s current living vocabulary.
E — Explanatory Power: the semantic sufficiency of the contribution to ground a resolvable meaning distinction. A contribution that cannot change how the system resolves future queries has no explanatory power.
C — Contextual Coherence: consistency with the contributor’s demonstrated domain history and semantic trajectory. The contribution fits the pattern of genuine understanding developing over time.
T — Temporal Priority: cryptographic proof that the contribution was committed before any outcome was observable.
K — Collective Connection: logarithmically saturated independent activation by contributors with non-overlapping coherence histories. Why each is necessary and none is sufficient alone: each addresses a different attack surface and a different form of meaning loss.
4.4 The singular and the collective
The irreducible singularity of individual meaning-making is the generative source. The collective sense-making map is the network effect — something that transcends the sum of its parts and is irreducible to any individual contribution. The collective map is not an average of individual meanings. It is an emergence: individual singularity is the source; collective intelligence is what it becomes when those singularities are preserved and connected.
5 · Memory Architecture
5.1 Working memory
Session-scoped, high plasticity, resolves meaning for this user in this role in this context right now. Ephemeral by design — the same person means different things in different contexts, and a system that cannot distinguish session context from long-term pattern produces systematically wrong resolutions. Raw content not retained.
5.2 Personal semantic memory
Role-bound, private by default, accumulated over sessions, encoding the contributor’s stable interpretive patterns per domain. Encrypted with keys that never leave the contributor’s control. Never shared in raw form. The basis of the contributor’s coherence score (C parameter) and the source of their provenance chain. What propagates to the collective layer is not the content of the personal memory but an anonymised semantic signal derived from it — the fingerprint of understanding, stripped of identifying content.
5.3 Collective sense-making maps
Domain-scoped, community-owned, built from anonymised and attributed contributions that have passed Aletheia protocol verification. Temporally alive: require continued human engagement to maintain meaning weights. Commercially licensable when domain density is sufficient. High stability in established zones. Sensitive to domain evolution. Resistant to noise through logarithmic saturation and coherence weighting. Preserving contested zones as first-class data rather than resolving them into false consensus.
Why the map cannot be a static database: its value is entirely in its living connection to ongoing human sense-making. Extract it, package it, sell it as a finished product, and you are selling something already dying. The moment genuine human contribution stops feeding it, the map begins its decay toward obsolescence.
5.4 Flow between layers
Working → Personal: automatic trace, weighted by signal strength. Every session leaves a record in the personal layer when patterns are sufficiently consistent to qualify as stable understanding. Personal → Collective: gated by the Aletheia validity condition. Only contributions satisfying S, E, C, T, K thresholds propagate. Collective → Working: top-down grounding at session start. The system pre-loads resolved meanings for the user’s role and domain from the collective map. This is where token cost reduction occurs.
Why the flow is asymmetric: individual meaning generates collective intelligence, but collective intelligence does not override individual meaning. The personal layer takes precedence in familiar contexts.
5.5 Adaptive retrieval — the weighted combination
All three layers contribute simultaneously to meaning resolution, weighted by signal confidence for the specific query context. A naive cascade — checking working first, then personal, then collective — ignores the simultaneous relevance of all three layers. The correct model is a weighted combination where each layer’s weight reflects how much the system knows and trusts its signal for this specific user, role, domain, and moment.
6 · The Weighted Semantic Unit
6.1 Definition
6.2 The four structural components
Canonical interpretation — the primary resolved meaning for this domain and role context, expressed as natural language plus an embedding vector. The natural language statement is what humans and enterprises read. The vector is what the system queries.
Validated variants — alternative interpretations that are meaningfully distinct from the canonical but legitimate. Attributed to their contributors, weighted by validation. These represent genuine role or subculture differences within a domain.
Contested zones — interpretations in genuine expert disagreement. Not errors to be resolved — first-class semantic data. Represented as a structured debate: each position attributed and weighted, the specific axis of disagreement identified, the contexts in which each applies specified.
Provenance chain — the complete attributed history of all contributions to this WSU: who contributed what, when, in what context, with what validation, activation count, and current weight. The economic layer’s ledger.
6.3 WSU Confidence Level — eight factors
The single composite score that summarises a WSU’s epistemic trustworthiness. Drives adaptive retrieval state decisions, licensing eligibility, and provenance distribution calculations. Composed of eight factors:
Coherence weight — the C parameter of the primary contributor(s): demonstrated domain history and trajectory consistency. Contextual relevance — how closely the WSU’s domain and role context matches the current query context. Authorship verification — the S parameter: degree to which contributions have passed genuine subjecthood verification. Meritocratic standing — the verified expertise level of contributing authors in this domain. Domain expertise depth — the contributor’s demonstrated accuracy and consistency specifically in this domain. Collective sense-making density — the K parameter: logarithmically saturated independent activations. Temporal stability — how long the WSU has maintained its weight across multiple epochs without significant challenge. Contestation history — whether the WSU has been challenged and how the challenge resolved. A WSU that survived scrutiny carries stronger confidence than one never questioned.
6.4 Contested zones as first-class data
The most epistemically valuable regions of the map are the places where genuine experts hold genuinely different interpretations. These are not errors to be resolved — they are semantic information about contested territory. The contested zone trigger combines statistical variance detection with embedded logic judgements, comparing contributions against domain baselines to distinguish linguistic ambivalence (registered automatically), interpretive variance (flagged for community assessment), and genuine contested zones (elevated with full structured disagreement record).
6.5 Semantic individuation
The two-level criterion: embedding similarity at the surface level suggests the same unit; role-context divergence overrides. Two contributions with similar embeddings but significantly different role contexts are treated as variants rather than confirmations — because the CFO’s understanding of “material risk” and the engineer’s understanding are meaningfully distinct even if their surface descriptions are close. Role-context is the primary individuation signal; embedding distance is secondary. The inverse of how current AI systems handle it, which collapse role-context differences into noise.
6.6 Meaning weight W(t)
The meaning weight of a WSU is not static. It grows as the unit is activated and validated; it decays when not reinforced. The decay is the protocol’s primary tamper against the map calcifying into historical record.
W(t) = V(WSU) · e−λt · R(t)Where λ is the domain decay rate and R(t) is the reinforcement function that resets the decay clock when the WSU is validated or activated. A meaning actively used maintains its weight indefinitely. A meaning no longer activated decays toward zero: deprioritised in retrieval, flagged for review, eventually eligible for retirement.
6.7 Minimum contribution threshold — explanatory power
The floor below which a contribution cannot generate a WSU candidate or accumulate provenance credit. Determined not by length but by explanatory power: a contribution must describe a distinction, relationship, or interpretation that the system can use to resolve future queries differently from the absence of that contribution. A single word that specifies a domain boundary may satisfy the threshold. A paragraph of generic domain-adjacent language may not. The system assesses explanatory power algorithmically and can request elaboration when a contribution is semantically insufficient.
6.8 WSU retirement
7 · The Aletheia Protocol
7.0 Why Aletheia
Aletheia is the ancient Greek word for truth, recovered by Heidegger from its common translation as correctness (the matching of a proposition to a fact) to its original meaning: unconcealment. The disclosure of what was previously hidden. The bringing-into-the-open of something that was present but not yet visible.
When a human corrects an AI’s approximation of a concept, they perform exactly this: they unconceal the meaning that the machine’s statistical output had covered over. They reveal what was already there — in their understanding, in their domain’s living practice, in their community’s accumulated knowledge — but had not yet been made legible to the system.
7.1 Proof of meaning
Bitcoin’s SHA-256 defines valid work as a hash falling below a target value: computationally expensive to produce, trivially cheap to verify, adjustable to maintain stable block time, making dishonesty more expensive than honesty without requiring trust in any central authority.
The Aletheia Protocol defines valid meaning as a semantic contribution satisfying five simultaneous conditions. The validity condition:
V(WSU) = S · E · Cα · Kβ · TS, E, and T are binary gates: they cannot be compensated for by high scores on other parameters. A contribution with perfect coherence and massive collective connection produces V = 0 if S = 0 (no genuine subjecthood) or T = 0 (no prior commitment) or E = 0 (no explanatory power). C and K are weighted continuous scores, domain-configurable through α and β. The meaning weight function:
W(t) = V(WSU) · e−λt · R(t)7.2 The five parameters — precise definitions
7.3 What the protocol tampers against
7.4 – 7.12 · The nine protocol functions
8 · The Epoch Structure
Three speeds run simultaneously per domain, each optimised for a different function.
9 · Access Design and Product Principles
Universal benefit from day one. All users benefit from adaptive retrieval across general domains immediately. Every user is simultaneously building the map and benefiting from it. There is no separation between contributor phase and user phase.
Expert contribution is permissioned. High-stakes domains gate contribution at the write level through three verification mechanisms: credential verification (qualifications, professional memberships, institutional affiliations), contribution history (demonstrated accuracy and consistency over time), and peer validation (existing verified contributors recognising new contributors’ genuine domain expertise). The gate is meritocratic: demonstrated quality, not institutional membership alone.
What the system sells and what it structurally cannot sell. Licensed users access the intelligence produced by the collective map: resolved meanings, ranked interpretations, and structured disagreements through the adaptive retrieval system. Individual user histories, raw contributor content, personal semantic memories, contributor identities, and the provenance chain in identifiable form are architecturally inaccessible. Individual semantic memory is encrypted with keys that never leave the contributor’s control.
Two access tiers. Individual license: personal use, limited role contexts, collective map access, personal memory private and included. Enterprise license: unlimited role count, private institutional memory layer built from employee contributions sitting above the general collective map, role-sensitive resolution across the organisation’s defined structure. The private institutional memory is the enterprise tier’s differentiating product: once built from an organisation’s contributors, it reflects their specific terminology, role structure, and domain expertise. High switching costs arise from the value of what has been built, not from contractual lock-in.
10 · Governance Principles
Domain creation. Domains are created with a validity function evaluating taxonomic distinctiveness: is this domain meaningfully different from existing ones in scope, vocabulary, and contested territory? Domain parameters set at creation using available market and expert data and adjustable by governance. The principle: domains should reflect genuine communities of practice with distinct meaning systems, not arbitrary subdivisions designed to accumulate provenance credit in uncrowded territory.
Cross-domain weighting. Full independence as default — coherence in one domain does not transfer to another automatically. Manual cross-domain connection available when expertise genuinely spans domains, subject to independent coherence verification in each. Broad categories (natural language nuance, general reasoning patterns, domain-crossing frameworks) gather contributions from all users. Deep domain categories are siloed.
Protocol fee principle. When licensing revenue is distributed, Kibela takes a protocol fee for infrastructure and development. The fee is publicly stated, governance-determined, and the minimum necessary to sustain the protocol. The majority of licensing revenue flows to contributors through provenance distribution. The community’s trust in the protocol fee is secured by governance transparency, not by goodwill.
11 · Competitive Position and Structural Moat
Against current AI systems. They re-infer meaning from scratch on every query. Kibela resolves it once, persists it, and applies it at session start through collective map grounding. Cost reduction is real, measurable, and compounds as map density grows. The improvement is the architectural consequence of having a persistent semantic layer, not a feature that can be toggled on by competitors without rebuilding their architecture.
Against fine-tuning, RAG, and memory systems. Fine-tuning modifies model weights: expensive, opaque, requires data science infrastructure. RAG retrieves documents: operates above the semantic layer, returns documents rather than resolved meaning, has no mechanism for attributing the value of retrieval to the humans who wrote them. Memory systems store facts and preferences: what the user did, not what they genuinely understand. Kibela stores meaning — the interpretive structure beneath facts, documents, and preferences. One verified correction changes how every future query in that domain resolves. No infrastructure, no training pipeline, no data team required.
Honest AI and calibrated uncertainty. Adaptive retrieval State 2 surfaces structured disagreement rather than false resolution in contested, high-stakes domains. The first AI system designed to show its uncertainty when stakes are high — a structural consequence of having a map that represents contested zones as first-class data. The collective map, built from diverse attributed contributions across verified domain experts, produces less systematically biased outputs than homogenised training data for one fundamental reason: the contributors are identified, the contributions are attributed, and the disagreements are preserved rather than averaged.
Scope — What This Document Does Not Cover
This document establishes the governance specification of the Kibela protocol. Companion documents cover mathematical derivations of the scoring functions and decay rates, cryptographic construction of the zero-knowledge coherence proof and commit-reveal system, the marketplace extension covering visual and audio semantic assets, licensing business model and token economics at scale, and philosophical and cognitive science foundations underlying the three foundational claims.
Glossary
Key References
| Author(s) | Year | Work | Relevance |
|---|---|---|---|
| Arrieta-Ibarra et al. | 2018 | Should We Treat Data as Labor? | Economic case for attributing value to semantic contributions. Section 1.7. |
| Bender, Gebru et al. | 2021 | On the Dangers of Stochastic Parrots | Language models do not learn meaning from text. Section 1.1. |
| Couldry & Mejias | 2019 | The Costs of Connection | Data colonialism framework. Section 1.3. |
| Fricker, M. | 2007 | Epistemic Injustice | Testimonial and hermeneutical injustice. Section 1.6. |
| Guo et al. | 2017 | On Calibration of Modern Neural Networks | Neural network confidence is poorly calibrated. Section 1.5. |
| Harris, Z. | 1954 | Distributional Structure | The foundational assumption Section 1.1 examines and critiques. |
| Kadavath et al. | 2022 | Language Models Know What They Don’t Know | Models’ limited capacity to signal uncertainty. Section 1.5. |
| Lazer et al. | 2009 | The Parable of Google Flu | Failure modes of scale in statistical systems. Section 1.4. |
| Longino, H. | 1990 | Science as Social Knowledge | Epistemic value of diversity. Section 1.4. |
| Merrill et al. | 2021 | Provable Limitations of Acquiring Meaning from Text | Formal proof of limits on semantic content from text alone. Section 1.1. |
| Narayanan & Felten | 2014 | No Silver Bullet: De-identification Doesn’t Work | The failure of anonymisation as privacy. Section 1.6. |
| Page, S. | 2007 | The Difference | Diversity prediction theorem. Section 1.4. |
| Posner & Weyl | 2018 | Radical Markets | Mechanisms for data as labour. Section 1.7. |
| Varoufakis, Y. | 2023 | Technofeudalism | Cloudal capital and cognitive rent extraction. Section 1.3. |
| Zollman, K. | 2010 | Epistemic Benefit of Transient Diversity | Epistemic monoculture risk. Section 1.4. |