The N'Ko Compute Network

The N'Ko Compute Network

A Protocol for Human-Linguistic Computation on Bitcoin

Mohamed Diomande
March 2026

---

Abstract

Every existing compute network, from Bitcoin to Akash to Render, treats workers as interchangeable machines. The worker's identity, language, and culture are irrelevant to the protocol. This paper proposes a fundamentally different architecture: a compute network where the worker's linguistic and cultural competence IS the valuable computation, and the protocol pays for it in STX on Bitcoin's Layer 2.

The N'Ko Compute Network combines three production systems into a single protocol: (1) the EPOCH Protocol, eight Clarity smart contracts on Stacks that route compute tasks and manage treasury distribution with $21.64 proven on testnet; (2) an N'Ko ASR pipeline with a 3,024-entry codebook, FarmRadio Whisper fine-tuned for Manding languages, and Gemini Flash OCR running on Vast.ai GPU infrastructure; and (3) a distributed mesh coordination layer with real-time task routing across five machines, proven at scale with Evolution World invariants and NUMU bus messaging.

The thesis is simple. Over 40 million people speak Manding languages. Virtually zero AI training data exists in N'Ko, the script engineered by Solomana Kante in 1949 to give those languages a native written form. This protocol turns N'Ko speakers into compute nodes. Their linguistic competence, the ability to validate tonal accuracy, catch cultural misuse of proverbs, generate authentic text, is computation that no GPU can perform. The economic flywheel: learn N'Ko, become a worker, train the AI, AI improves, demand for workers grows, more people learn N'Ko. Cultural preservation becomes economic activity.

---

1. Problem Statement

1.1 The Data Desert

N'Ko (Unicode U+07C0-U+07FF) is one of the best-engineered scripts in existence. Every phoneme maps to exactly one grapheme. Tone is marked explicitly. There are zero spelling exceptions. From a computational linguistics perspective, it is structurally superior to Latin orthography for representing Manding languages.

Yet modern AI is effectively blind to it. Activation profiling of Qwen2-72B-Instruct across all 81 transformer layers reveals a 2.90x "translation tax" on N'Ko compared to English: L2 norm ratio of 2.90x, 30-60

Every Bambara ASR system in production, all eleven publicly available models, produces Latin-script output. A child in Kankan, Guinea who speaks Maninka and reads N'Ko cannot dictate a text message, search the web, or interact with any AI system in their own script. The entire ASR field has been writing Manding languages in a foreign orthography designed for French linguists.

The root cause is not technical difficulty. It is a data problem. There is almost no digitized N'Ko text for training. There are almost no N'Ko speech-to-text pairs. The data does not exist because the economic incentive to create it does not exist. No company will pay to annotate N'Ko data because no product generates revenue from N'Ko users because no AI works in N'Ko. The loop is closed.

1.2 The Compute Network Problem

Existing compute networks solve the wrong problem for this domain. Bitcoin's Proof-of-Work converts electricity into security. Ethereum's Proof-of-Stake converts capital lockup into security. Akash, Render, and io.net sell GPU cycles as a commodity. In all cases, the worker is a machine. The worker's identity is a public key. The worker's cultural knowledge, linguistic skill, and contextual judgment are irrelevant.

But the computation required to build N'Ko AI is not parallelizable matrix multiplication. It is: does this N'Ko sentence sound natural to a native speaker? Is this proverb used correctly in this context? Does this translation preserve the tonal meaning? These are judgment tasks that require fluency in both the language and the culture. A native N'Ko speaker is literally a more capable compute node than an H100 for this class of problem.

1.3 The Economic Disconnect

Amazon Mechanical Turk pays $0.01-$2.00 per microtask. Scale AI pays $15-50/hour for annotation work, primarily to workers in the US and Europe. Appen pays $5-15/hour, with African annotators often receiving $1-3/hour. These platforms extract value from workers in developing economies and route it to AI companies in wealthy ones.

More importantly, none of these platforms build lasting infrastructure for the worker's community. A Kenyan annotator labeling English images builds training data for models they will never use and languages they already speak. The work creates no cultural asset. When the contract ends, nothing remains.

---

2. Proposed Solution: Proof-of-Linguistic-Competence

The N'Ko Compute Network introduces a new consensus mechanism: Proof-of-Linguistic-Competence (PoLC). Workers prove their value not by burning electricity or locking capital, but by demonstrating the ability to perform linguistic computation that machines cannot.

2.1 The Flywheel

The protocol creates a self-reinforcing cycle:

1. Apps pay protocol fees. Any AI-powered application (translation, ASR, content moderation, education) that processes Manding languages pays a fee in STX to the protocol's treasury smart contract.

2. Treasury distributes to workers. The EPOCH treasury contract, already deployed on Stacks testnet, distributes STX to workers based on verified task completion.

3. Workers produce training data. Every completed task (translation pair, transcription, cultural review, original content) becomes training data that improves the underlying AI models.

4. Better AI creates more demand. As models improve, more applications can serve Manding-speaking users, generating more fees and more work.

5. More work attracts more workers. Economic opportunity drives N'Ko literacy. Learning to read and write N'Ko becomes a path to income.

6. Cultural preservation accelerates. The act of working on the protocol IS the act of preserving the language. Every task produces a digitized artifact in N'Ko.

This is not a speculative design. Each component exists in production today. The contribution of this paper is the protocol that connects them.

2.2 What Workers Do

Workers on the N'Ko Compute Network perform five categories of linguistic computation:

Translation Validation. Given a machine-generated N'Ko translation, workers rate accuracy, naturalness, and cultural appropriateness on a structured rubric. Two to three workers validate each translation. Consensus triggers payment.

Audio Transcription. Workers listen to Manding-language audio and produce N'Ko transcriptions. The pipeline uses FarmRadio Whisper for initial Latin-script transcription, then workers verify the N'Ko transliteration and correct errors that the 28-rule cross-script bridge cannot catch.

Training Data Creation. Workers generate original parallel text: a sentence in N'Ko with its French or English equivalent, tagged with domain, register, and cultural context. The five-world generation system (Everyday, Formal, Storytelling, Proverbs, Educational) guides workers toward high-value data categories.

Cultural Review. AI-generated N'Ko content is submitted to workers who evaluate whether proverbs are used correctly, whether register is appropriate for context, whether tonal marking is accurate, and whether the text would be understood by a native speaker in its intended setting.

Original Content. Workers write original text in N'Ko, news summaries, educational materials, conversational examples, covering domains where no training data exists. Content is tagged, reviewed by a second worker, and ingested into the training pipeline.

---

3. Technical Architecture

3.1 System Overview

The protocol consists of three layers: the chain layer (EPOCH smart contracts on Stacks), the coordination layer (mesh task routing with Supabase and NUMU bus), and the worker layer (mobile-first task delivery with SMS fallback).

                    APPLICATIONS
                         |
                    [STX Payment]
                         |
              +----------v-----------+
              |   EPOCH TREASURY     |    Stacks/Bitcoin L2
              |   (Clarity Contract) |    8 contracts, 38 tests
              +----------+-----------+
                         |
              +----------v-----------+
              |   CHAIN ORCHESTRATOR |    Routes tasks by type,
              |   + AGENT REGISTRY   |    matches to workers
              +----------+-----------+
                         |
              +----------v-----------+
              |   MESH COORDINATOR   |    Supabase + NUMU bus
              |   task_routing       |    Real-time distribution
              |   work_claims        |    Conflict resolution
              |   mesh_events        |    Event propagation
              +----------+-----------+
                         |
              +----------v-----------+
              |   WORKER ENDPOINTS   |    Mobile app, SMS, USSD
              |   Task queue         |    Offline-capable
              |   Result submission  |    Multi-format upload
              +----------+-----------+
                         |
              +----------v-----------+
              |   N'Ko ASR PIPELINE  |    FarmRadio Whisper
              |   3024-entry codebook|    Gemini Flash OCR
              |   Cross-script bridge|    5-world generation
              +----------------------+

3.2 The Chain Layer: EPOCH Protocol

The EPOCH Protocol consists of eight Clarity smart contracts deployed on Stacks, Bitcoin's Layer 2 with Proof-of-Transfer consensus. Clarity is a decidable, non-Turing-complete language where the full call graph is analyzable statically, meaning every possible state transition is knowable before deployment. The contracts have passed 38 unit tests and two red-team security audits.

Treasury. Holds protocol funds in STX. Accepts deposits from applications and distributes to workers based on verified task completion. The treasury contract implements a fee splitter: 85-90

Fee Collector. Accepts STX payments from applications requesting linguistic computation. Validates payment amount against current task pricing. Forwards funds to treasury with metadata linking payment to specific task types.

Agent Registry. Maintains an on-chain record of all registered workers. Each worker entry includes: wallet address, language proficiencies (self-declared, then verified), skill ratings (updated by consensus outcomes), availability status, and total tasks completed. The registry is the identity layer; it answers the question "who can do this task?"

Chain Orchestrator. The routing brain. When a task arrives (via fee collector payment), the chain orchestrator queries the agent registry for qualified workers, selects candidates based on skill rating and availability, and emits events that the off-chain coordination layer picks up. This is the key architectural innovation: the blockchain does not just record state, it actively routes computation.

Time Log. Records timestamped entries for all protocol activity: task submissions, worker assignments, completion events, payment distributions. Provides an immutable audit trail. Each entry is a Clarity tuple with epoch timestamp, actor principal, action type, and a 32-byte data hash.

Micro-DEX. A minimal on-chain exchange allowing workers to swap STX for other tokens or stablecoins. Workers in West Africa often need to convert to local currency equivalents. The micro-DEX integrates with DeFi Llama, Velar, and STXTools APIs for pricing.

Marketing Engine. Manages referral bonuses and growth incentives. When an existing worker refers a new worker who completes their first verified task, both receive a bonus from the treasury. The contract enforces caps to prevent gaming.

Compute Market. Originally designed for GPU task routing (pricing LLM inference at 100-500 micro-STX per 1K tokens), this contract is extended to support human computation tasks. Pricing for linguistic tasks is determined by: task complexity (translation < transcription < cultural review < original content), language pair, and current worker supply for that task type.

3.3 The Coordination Layer: Mesh Task Router

On-chain transactions on Stacks finalize in approximately 30 seconds (post-Nakamoto upgrade). Human task assignment requires lower latency. The coordination layer bridges this gap.

Four Supabase tables handle real-time coordination:

• task_routing: Maps incoming tasks to worker pools based on language, skill tier, and geographic availability. Updated in real-time as workers come online or go offline.
• work_claims: Implements optimistic concurrency control. When a worker claims a task, the claim is recorded with a timestamp. If the worker does not submit a result within the timeout window, the task is released back to the pool.
• mesh_events: Event bus for system-wide notifications. Worker completed a task, new task batch available, payment distributed, all propagated as events.
• compute_contributions: Tracks each worker's cumulative output, quality scores, and earnings. Feeds into the on-chain agent registry for periodic state synchronization.

The NUMU bus (WebSocket on port 7890) provides sub-second event propagation across the mesh. When a task is submitted, the sequence is:

1. Application calls fee collector contract (on-chain, ~30s finality)
2. Chain orchestrator emits a task-created event
3. Stacks event poller (off-chain, polling every 5s) detects the event
4. Event orchestrator writes to mesh_events table and broadcasts via NUMU
5. Worker endpoints receive the task notification in <1 second
6. Worker claims the task (off-chain, instant)
7. Worker completes the task and submits the result
8. Consensus validators (2-3 other workers) verify the result
9. Coordinator writes completion to chain (batched, every ~5 minutes)
10. Treasury releases payment to worker

3.4 The Worker Layer: Mobile-First Delivery

Over 80

Primary: Mobile Web App. Progressive Web App optimized for low-bandwidth connections. Tasks are pre-fetched when connectivity is available and cached locally. Workers complete tasks offline and results sync when connection returns. The interface renders N'Ko text natively using the Unicode block and system fonts available on modern Android and iOS devices.

Fallback: SMS/USSD. For workers without smartphones or reliable data, simple tasks (binary validation: "Is this translation correct? Reply 1 for yes, 2 for no") can be delivered via SMS. USSD menus provide task selection for feature phones. Results are routed back through the coordination layer.

Task Packaging. Each task sent to a worker includes: the source material (text, audio URL, or image), the expected output format, the payment amount in STX, the deadline, and the number of consensus validators required. Audio tasks include a compressed audio file (<500KB) playable on any device.

3.5 The Pipeline Layer: N'Ko ASR and Training

The N'Ko ASR pipeline is the primary consumer of worker output and the primary generator of new tasks. It runs on Vast.ai GPU instances (A100, $0.89/hr) and the Mac4/Mac5 compute cluster (M-series Apple Silicon, always-on).

ASR Architecture. A frozen Whisper large-v3 encoder feeds a character-level CTC decoder. A 4-state finite-state machine encoding N'Ko syllable phonotactics guarantees 100

Codebook. 3,024 entries mapping N'Ko characters, syllables, and common words to their phonemic representations. The codebook is the bridge between the acoustic model (which understands sound) and the script model (which produces N'Ko). Workers expand the codebook by submitting verified entries.

Cross-Script Bridge. 28 transliteration rules mapping Latin Bambara orthography to N'Ko. Six documented bug classes where the mapping fails (tonal ambiguity, vowel length, nasalization, digraph collisions, foreign loan words, and proper nouns). Workers correct bridge failures, and their corrections feed back into bridge rule refinement.

Training Loop. Worker-produced data flows directly into the training pipeline: translation pairs become SFT examples for language models, transcriptions become ground truth for ASR fine-tuning, and cultural reviews become RLHF preference data. The pipeline processes 522 YouTube videos from the @babamamadidiane channel (approximately 500 hours of N'Ko educational content), with each video yielding approximately 55 OCR frames and 60 minutes of segmentable audio.

---

4. Economic Model

4.1 Task Pricing

Task pricing is anchored to comparable markets but adjusted for two factors: the specialized skill required (N'Ko literacy commands a premium over generic labeling) and the economic context of West Africa (where purchasing power parity makes lower absolute amounts meaningful).

These rates place workers at 2-5x the typical African annotation platform rate ($1-3/hr on Appen and Remotasks) while remaining sustainable for the protocol. The premium is justified: N'Ko literacy is a scarce, specialized skill. Generic English labelers are a commodity. N'Ko cultural validators are not.

4.2 Revenue Projections

Revenue is modeled across three growth phases, assuming STX at approximately $0.30 USD and a 30-day month.

Phase 1: 100 Workers (Months 1-6)

• Average tasks per worker per day: 15
• Average task value: 3 STX ($0.90)
• Daily task volume: 1,500
• Daily gross revenue needed: 1,350 STX ($405)
• Monthly revenue needed: 40,500 STX ($12,150)
• Protocol fee (7.5
• Source: Direct contracts with 3-5 language technology companies needing Manding training data

Phase 2: 1,000 Workers (Months 6-18)

• Average tasks per worker per day: 20
• Average task value: 4 STX ($1.20)
• Daily task volume: 20,000
• Daily gross revenue needed: 80,000 STX ($24,000)
• Monthly revenue needed: 2.4M STX ($720,000)
• Protocol fee (7.5
• Source: API access for AI companies, educational technology platforms, government digitization programs

Phase 3: 10,000 Workers (Months 18-36)

• Average tasks per worker per day: 25
• Average task value: 5 STX ($1.50)
• Daily task volume: 250,000
• Daily gross revenue needed: 1.25M STX ($375,000)
• Monthly revenue needed: 37.5M STX ($11.25M)
• Protocol fee (7.5
• Source: Embedded AI services across West African mobile platforms, multilateral development funding, subscription APIs

4.3 Worker Economics

At Phase 1 rates, a worker completing 15 tasks per day earns approximately 45 STX ($13.50/day). In Guinea, where GDP per capita is approximately $1,200/year, this represents meaningful income: $405/month, roughly 4x the national average monthly income.

Workers can: (a) cash out STX through peer-to-peer exchanges or the micro-DEX, (b) convert to mobile money through local OTC providers, or (c) earn access credits that grant free use of N'Ko-powered applications and services.

4.4 Why This Is Not Exploitative

Three structural safeguards prevent the protocol from becoming another extractive data labeling platform:

On-chain transparency. All payments are recorded on Stacks, anchored to Bitcoin. Workers can verify that the protocol distributes the promised percentage of fees. There is no opaque intermediary taking an undisclosed cut.

Worker-owned output. The training data produced by workers is a public good. It trains open models that serve the workers' own communities. A Bambara ASR model trained on worker-produced data is available to every Bambara speaker, including the workers themselves.

Skill appreciation. Unlike generic labeling where more workers means lower rates, N'Ko linguistic skill appreciates in value as the AI ecosystem grows. A worker who develops expertise in cultural review or complex translation commands higher task rates over time. Their skill is not commoditized, it is specialized.

---

5. Cultural Impact

5.1 Kante's Unfinished Revolution

Solomana Kante created N'Ko in 1949 as an act of linguistic sovereignty. The claim that prompted him, that African languages could not be written, was both a colonial assertion and a practical reality: no standardized script existed for Manding languages. Kante spent years engineering a solution with properties that Latin-adapted orthographies could not match.

But Kante's revolution stopped at writing. N'Ko gave Manding speakers the ability to write their languages in a native script. It did not give them the ability to compute in them. Every digital system, from keyboards to search engines to AI, assumed Latin as the default. N'Ko literacy remained a cultural practice, not an economic one.

This protocol extends Kante's revolution into the computational domain. When a worker validates an N'Ko translation, they are not just earning STX. They are building the infrastructure that makes N'Ko a computationally viable script. Every task completed is a data point that teaches machines to understand a script that serves 40 million speakers.

5.2 Language as Infrastructure

The standard framing of language preservation treats it as a charitable cause: save the endangered language before it disappears. This framing is both patronizing and ineffective. Languages die not because people forget them, but because speaking them offers no economic advantage.

The N'Ko Compute Network inverts this framing. N'Ko literacy is not a charity case. It is a scarce computational resource. The protocol creates economic demand for a skill that only N'Ko-literate people possess. The more AI systems need N'Ko data, the more valuable N'Ko literacy becomes.

This is not language preservation as charity. It is language preservation as market-making.

5.3 Mali's Constitutional Moment

In 2023, Mali elevated national languages including Bambara to official status alongside French, with a 2025 five-year plan to combat illiteracy using N'Ko script for Manding varieties. This creates a concrete policy context for the protocol: government-funded literacy programs will produce thousands of new N'Ko readers who could immediately participate as workers. The protocol provides the economic reason to stay literate.

---

6. Comparison to Existing Systems

The fundamental difference: in every other network, the worker could be replaced by a better machine or a cheaper worker. In this network, the worker's irreplaceability IS the value proposition. A GPU cannot validate whether a Bambara proverb is used correctly in context. A crowdsource worker in Manila cannot evaluate N'Ko tonal accuracy. The computation requires the specific human.

---

7. Roadmap

Phase 0: Foundation (Complete)

• EPOCH Protocol: 8 Clarity contracts deployed on Stacks testnet, 38 tests passing, 2 security audits, $21.64 proven on testnet
• N'Ko ASR Pipeline: FarmRadio Whisper encoder, 3,024-entry codebook, 28-rule cross-script bridge, Gemini Flash OCR
• Mesh Coordination: 5-machine compute cluster, Supabase coordination tables, NUMU real-time bus, Evolution World invariants
• Research: Activation profiling paper documenting 2.90x translation tax, 81-layer analysis, LoRA reduction to 0.70x

Phase 1: Worker Onboarding (Q2-Q3 2026)

• Mobile PWA for task delivery with offline capability
• SMS/USSD fallback for feature phone access
• Worker registration and skill verification flow
• First 100 workers recruited from N'Ko literacy networks in Guinea, Mali, and Ivory Coast
• 3-5 initial customers (language technology companies, educational platforms)
• EPOCH contracts deployed to Stacks mainnet

Phase 2: Scale and Quality (Q4 2026 - Q2 2027)

• Consensus mechanism refinement: inter-rater reliability metrics, worker tier system
• ASR model retrained on worker-produced data (target: 20
• API access for AI companies needing Manding training data
• Worker count: 1,000
• Geographic expansion to Senegal, Gambia, Burkina Faso
• Integration with Mali's national literacy programs

Phase 3: Ecosystem (Q3 2027 - Q4 2027)

• N'Ko-native applications built on the improved models (voice search, dictation, translation)
• Worker cooperative structure: governance tokens for long-tenured workers
• Open-source release of all training data and models
• Worker count: 10,000
• Self-sustaining flywheel: app revenue funds workers, worker output improves apps

Phase 4: Replication (2028+)

• Protocol template generalized for other underserved scripts (Tifinagh, Vai, Ethiopic)
• Cross-language bridges: N'Ko-Tifinagh, N'Ko-Arabic, N'Ko-Ethiopic
• The N'Ko Compute Network becomes a reference implementation for Proof-of-Linguistic-Competence

---

8. Risks and Mitigations

Risk: Insufficient initial demand for N'Ko AI services.
Mitigation: Phase 1 targets enterprise customers (language technology companies) who have existing budgets for annotation work. Consumer demand is a Phase 3 requirement, not a Phase 1 one.

Risk: Worker quality variance.
Mitigation: Multi-worker consensus (2-3 validators per task), skill-tiered routing, and progressive trust scores. New workers start with simple validation tasks and graduate to complex creation tasks as their accuracy history builds.

Risk: STX price volatility affecting worker income.
Mitigation: The micro-DEX contract supports immediate conversion to stablecoins. Task pricing can be denominated in USD-equivalent and converted to STX at execution time.

Risk: Connectivity limitations in target regions.
Mitigation: Offline-first PWA architecture, SMS/USSD fallback, and task pre-fetching. The system degrades gracefully: even a worker with only SMS access can perform binary validation tasks.

Risk: Regulatory uncertainty around cryptocurrency in West Africa.
Mitigation: Workers receive STX but can immediately convert to mobile money through existing OTC networks. The protocol does not require workers to hold cryptocurrency; it is a payment rail, not a speculative asset.

---

9. Conclusion

The N'Ko Compute Network is not a theoretical proposal. Its three constituent systems, the EPOCH smart contracts, the N'Ko ASR pipeline, and the mesh coordination layer, are running in production today. What this paper describes is the protocol that connects them into a single flywheel where cultural preservation generates economic value and economic activity accelerates cultural preservation.

The deeper claim is that Proof-of-Linguistic-Competence represents a new category of consensus mechanism. Bitcoin proved that computational work can secure a network. Ethereum proved that economic stake can secure a network. The N'Ko Compute Network proposes that human cultural knowledge, the kind of computation that machines fundamentally cannot perform, can power a network that builds AI infrastructure for communities that the current AI ecosystem ignores.

Solomana Kante gave 40 million people a way to write. This protocol gives them a way to compute.

---

References

1. Diomande, M. (2026). "From Dead Circuits to Living Speech: Activation Profiling and Script-Native ASR for N'Ko." Independent research.

2. Kante, S. (1949). N'Ko script. Unicode block U+07C0-U+07FF, standardized 2006.

3. Wyrod, C. "The Light on the Horizon: N'ko Literacy and Formal Schooling in Guinea." Signs and Society.

4. Stacks Foundation. "Stacks: A Bitcoin Layer for Smart Contracts." stacks.co (2024).

5. Radford, A. et al. (2023). "Robust Speech Recognition via Large-Scale Weak Supervision." OpenAI.

6. Graves, A. et al. (2006). "Connectionist Temporal Classification: Labelling Unsegmented Sequence Data with Recurrent Neural Networks."

7. Doumbouya, M. et al. (2021). Nicolingua N'Ko text corpus.

8. MALIBA-AI (2024). bambara-asr-v3, 45.73

9. FarmRadioInternational. bambara-whisper-asr. Open-weight Whisper fine-tune.

10. Bayelemabaga (2025). 46,976 Bambara-French parallel segments.

11. WMT 2023. N'Ko shared task, 30.83 chrF++ en-to-nko.

12. Mali Constitution (2023). Elevation of national languages to official status.

---

Contact: Mohamed Diomande, [email]
Protocol: github.com/Diomandeee
Network: Stacks (Bitcoin L2)