NFA, Collaborative Intelligence

A forecasting engine for prediction-market traders: AI advocates argue each possible outcome, an impartial judge weighs the evidence independently of the market, and you get a probability you can audit.

Abstract

NFA is a forecasting engine for prediction markets. A user submits a Polymarket or Kalshi market URL; the engine researches the market into a sourced evidence corpus, then runs a structured debate in which one AI advocate argues the strongest case for each possible outcome, and the advocates rebut each other across rounds. An impartial judge weighs the competing cases against the evidence and returns a calibrated probability distribution across the outcomes, an independent forecast rather than an echo of the market. Every run is fully replayable, and accuracy is scored publicly against resolved markets.

The product addresses a specific gap in the prediction-market ecosystem: traders on Polymarket, Kalshi, and similar venues have access to price feeds, historical data, and news, but lack structured tools for forecasting outcomes driven by actor behavior under pressure: geopolitical events, regulatory decisions, policy outcomes, narrative dynamics, and behavioral markets. This class of markets represents 30 to 40 percent of active prediction-market inventory by count and a higher share of volume during periods of crisis, elections, and major news events. NFA serves this gap, and the same engine forecasts statistical and microstructure markets too.

Three design principles shape the architecture. First, accuracy is the single metric that matters: all product, economic, and incentive decisions optimize for measurable forecasting accuracy. Second, the engine is production-grade from day one. Deterministic execution with full replay, per-call cost reconciliation, per-purpose LLM routing, prompt caching, and within-round parallelism are architectural primitives, not optimizations. Third, the forecast is independent: NFA produces its own estimate from the evidence, so the divergence between NFA and the market is the signal to trade rather than an echo of the crowd.

NFA launches on Solana. The token, settlement, and vesting are deployed as audited on-chain Anchor programs; the forecasting engine is a proprietary stack of permissively licensed open-source components glued together by a custom orchestration layer.

This whitepaper describes the product, the technical architecture, the marketplace mechanics, the token model, and the path to launch.

1 · Problem

1.1 The prediction market category is growing fast

Prediction markets have matured from fringe experiments into a significant category of financial infrastructure. Polymarket crossed billions in volume during the 2024 US election cycle. Kalshi secured CFTC-regulated status and expanded across politics, economics, and weather. New markets continue launching around cultural events, crypto outcomes, corporate actions, and geopolitical developments. Both retail and institutional trading activity has grown consistently.

The category works because prediction markets aggregate information efficiently. Prices reflect the collective best estimate of thousands of participants, often outperforming polls, expert forecasts, and traditional surveys. But the efficiency of market-based aggregation has a ceiling set by the quality of information available to participants. Where participants have access to strong tools, markets price efficiently. Where participants rely on intuition, markets reflect that limitation.

1.2 No single tool gives traders a calibrated probability

Prediction markets broadly fall into three categories by what determines their resolution:

• Statistical markets where historical data carries most of the signal: sports outcomes, base-rate questions, weather. Strong single-purpose models exist (ELO, xG, weather, actuarial data), but they are fragmented and rarely output one calibrated, tradeable probability for the specific market in front of you.
• Microstructure markets where price action over short horizons determines resolution: crypto price movements on 5-minute or hourly intervals. Traders use on-chain analytics, orderbook data, and funding rates.
• Actor-driven markets where outcomes depend on how specific actors behave under specific pressures: geopolitical events, regulatory decisions, policy outcomes, corporate actions, behavioral markets involving specific individuals or organizations. Traders have essentially no structured tools. They rely on news reading, intuition, and crowd-sourced takes on social media.

NFA serves all three. Its research-debate-judge loop grounds every forecast in researched evidence and base rates, so it returns one calibrated probability whether the question is statistical or actor-driven, folding the relevant data and models into the judge's reasoning rather than ignoring them. Its deepest edge is in actor-driven markets, the third category, where no structured tool exists today and where volume concentrates during elections, crises, and regulatory cycles. But the same engine forecasts a statistical market just as cleanly, giving traders a single tool across the board.

1.3 Why existing approaches fall short

Expert panels and forecasting tournaments. Services like Good Judgment Project have demonstrated that trained human forecasters can outperform intelligence agencies on specific question classes. But expert panels are slow, expensive, and don't scale. They cannot be applied in real-time to the thousands of markets active on Polymarket or Kalshi at any given time.

Single-LLM forecasters. Direct LLM prompting ("will this peace deal happen?") produces shallow, unreliable forecasts. Single models lack the structured reasoning required for multi-actor dynamics, don't maintain consistent mental models of adversarial incentives, and hallucinate specifics that undermine credibility. LLMs are necessary but insufficient.

News aggregators and sentiment trackers. Tools that aggregate news coverage or track social sentiment provide useful signal but don't produce forecasts. A trader still has to translate "sentiment is 65 percent negative" into "probability is X percent." The translation is the hard part, and it's exactly what traders need help with.

Generic multi-agent frameworks. Running several LLM agents against a question improves on single-LLM forecasting but produces inconsistent, unauditable results without structure. NFA's structure is adversarial and adjudicated: one advocate is required to argue each possible outcome from a sourced evidence corpus, and a separate judge weighs the competing cases against that evidence. The arguments are comparable across runs, the judgment is auditable, and the forecast is made independently of the market price.

1.4 The opportunity

A structured forecasting engine that argues every possible outcome and adjudicates the evidence independently of the market, with published accuracy data, fills a real gap, most acutely in actor-driven markets where no structured tool exists today, but general across the board. No incumbent occupies this position. The distribution channel (MCP, prediction-market frontends) is open. The technical substrate has matured to production-grade in the last 18 months. The audience is measurable and growing. The timing is right.

2 · Solution

2.1 Research, advocates, and a judge

NFA's core engine produces a forecast in three steps. First, it researches the market into a sourced evidence corpus: the actors involved, the timeline, the established facts, and the open questions. Second, it runs a structured debate, instantiating one AI advocate for each possible market outcome; each advocate argues the strongest evidence-based case for its outcome, and the advocates rebut one another across several rounds. Third, an impartial judge weighs the competing cases against the evidence and emits a calibrated probability distribution across the outcomes.

The design is deliberately adversarial. A single model collapses a hard question into one estimate and drifts toward the obvious answer; forcing a dedicated advocate to argue each outcome surfaces the strongest case on every side, and a separate judge must reconcile them against the evidence before ruling. Crucially, the judge is never shown the market price: it forecasts from the evidence and base rates alone, so the output is an independent estimate, an edge to trade against the market rather than a reflection of it.

The engine is general-purpose. It does not hard-code any market or domain; it researches whatever market it is given and argues whatever outcomes that market defines. Reusable scenario templates (§2.2) can tune how a class of market is researched and judged, but the same research-debate-adjudicate loop runs on every market.

2.2 Scenario templates

Every run executes against a scenario template: a reusable configuration for a recurring kind of market, a negotiation, a regulatory vote, a central-bank decision, an election. A template tunes how that class of market is researched and how the debate and judge are parameterized; it never names who plays in any specific market, because the actors, evidence, and outcomes are discovered per run from the market itself. The same election template applied to two different races draws a different field, different evidence, and a different verdict.

Today these templates are maintained internally. Opening template authoring to domain experts, with a marketplace that ranks and rewards them on real forecasting accuracy, is a planned later phase (§5).

2.3 Distribution

NFA reaches users through three surfaces:

• Web frontend at nfa.club. Retail traders paste market URLs, run forecasts, and inspect the debate and the full reasoning trace. The primary consumer surface.
• MCP server. Machine-facing API. Trading agents, algorithmic desks, and AI assistants like Claude and ChatGPT consume NFA programmatically. MCP is the standardizing protocol for AI agent tooling; early positioning in public registries is a durable distribution advantage.
• Authoring interface (planned). Where domain experts will tune and publish scenario templates, with an AI copilot that scaffolds the research and judging configuration. Internal today; opens to contributors in a later phase.

3 · Product

3.1 The user flow

A trader's primary interaction with NFA is simple: paste a Polymarket or Kalshi market URL into the web frontend, or pass it to the MCP server. The engine fetches the market metadata and outcomes, selects the scenario template that fits the market's category, researches the market, runs the advocate debate, and returns the judge's probability distribution with a full reasoning trace. There is no configuration step and no scenario to choose: one paste, one forecast.

When the scenario marketplace opens (§5), advanced users will additionally be able to pick a specific expert-tuned template or author their own. Until then, every market is served by NFA's own templates.

3.2 Forecast output

When a forecast runs, the output includes:

• Probability distribution across every market outcome, with a confidence band on each.
• Reasoning trace: each advocate's case, the rebuttals exchanged, and the judge's written rationale, fully replayable.
• Divergence summary comparing NFA's probability to the current market price, outcome by outcome, highlighting where, and by how much, the engine disagrees with the market.
• Accuracy context: the engine's historical accuracy on similar markets, so the user can calibrate confidence in the output.

Variance bundles

A standard run reports a probability with a confidence band. A variance bundle re-runs the debate and judge multiple times with different seeds and reports how stable the verdict is: whether the same outcome keeps leading, how tightly the distribution holds, and which outcomes are most sensitive to the framing. Outlier runs are flagged separately.

A forecast reporting "Field 26% with the lead holding in 9 of 10 runs and a ±3 point spread" is materially more useful than a single point estimate. Variance bundles are a planned premium feature.

3.3 The frontend

Live market gallery. A continuously updated feed of active Polymarket and Kalshi markets that NFA has forecast. Each entry shows the current market price, NFA's independent probability, and the divergence between them. Sortable by divergence magnitude, accuracy, recent activity, or market volume. This is the content engine: every large divergence is a shareable piece of content.

Forecast runner. The URL-paste interface. Runs stream in real-time, showing the advocates argue and the judge rule. Final output includes the probability distribution, reasoning trace, divergence summary, and accuracy context.

Marketplace browser (planned). When the scenario marketplace opens, templates will be browseable by category, accuracy, and usage, each with a public accuracy history broken down by market category.

Author hub (planned). Where template authors will manage their work, track accuracy, and view earnings, once authoring opens to contributors.

3.4 The MCP server

For programmatic consumption, NFA exposes an MCP (Model Context Protocol) server. Trading agents, AI assistants, and custom bots integrate via standardized tool calls. The server is listed in public MCP registries.

Exposed tools include:

forecast_market(market_url)                // research, debate, judge → distribution
get_forecast(market_url)                   // latest cached forecast + divergence
compare_markets(market_urls)              // forecast across related markets
explain_forecast(forecast_id)             // advocate cases + judge rationale, replayed
list_forecasts(category, min_accuracy)    // recent forecasts by category

Access is billed in USDC. Trading agents pay per forecast at the same USDC run price as the web frontend.

3.5 Authoring (planned)

In a later phase, domain experts will be able to author and publish their own scenario templates, tuning how a class of market is researched and how the debate and judge are configured, with an AI copilot handling the structural scaffolding. Templates will be rankable and rewardable on real accuracy, and an author will be able to keep a template private for personal use. The marketplace mechanics are described in §5.

4 · Technical architecture

The forecasting engine is production-grade by design. The properties below, deterministic replay, per-purpose routing, caching, evidence grounding, and reconciled cost tracking, are architectural primitives, not afterthoughts.

4.1 Determinism and replay

Runs are bit-for-bit reproducible where the underlying providers support it. Every run carries a stable seed context, pinned model versions per call class, and a captured trace that can be replayed end-to-end. Accuracy depends on this: any published score must be reproducible from the logged trace, and any disputed forecast can be replayed to an identical result. Without genuine determinism, accuracy scores would be advisory rather than authoritative.

4.2 Per-purpose LLM routing

LLM calls within a run serve different purposes with different cost-quality elasticity: research, advocacy, and judging are routed independently. NFA routes each call class to the appropriate model tier rather than running everything on a single model, yielding a meaningful cost reduction at no measurable quality loss versus all-frontier execution. Routing is configurable per run; advanced traders can opt into all-frontier routing for marginally higher accuracy at higher compute cost, and scenario templates can require frontier models for call classes where the forecast depends on it.

4.3 Prompt caching

A typical run reuses the same evidence corpus, advocate instructions, and judging context across dozens of LLM calls. Prompt caching exploits this structurally on every supported provider, substantially reducing both input-token cost and time-to-first-token after the first call in a window. Combined with per-purpose routing, this is the difference between optimized and unoptimized engine economics.

4.4 Parallel debate

The debate parallelizes by design. Every outcome's opening case is dispatched at once, each rebuttal round fans out across all advocates simultaneously, and the judge runs as a parallel self-consistency ensemble, all with bounded provider concurrency. Rebuttal rounds are the only sequential dependency. Rate-limit and transient-failure handling are first-class. The net effect is the difference between a forecast that returns in tens of seconds and one that takes many minutes.

4.5 Cost record persistence and reconciliation

Every LLM call emits a cost record with full provenance, provider, model, call class, and the source from which the cost was computed. A reconciliation job runs on a regular cadence comparing engine-reported run totals against provider billing API records, with any discrepancy flagged for manual review before settlement. This is essential: the USDC compute fee is derived from metered model cost, which must accurately reflect actual spend.

4.6 Evidence grounding and adversarial defense

Every advocate's claims are checked against the research corpus before the judge weighs them. The validator catches unsupported assertions, contradictions with the established facts, and arguments that lean on fabricated specifics rather than the evidence, the failure mode that makes raw LLM forecasts untrustworthy. The judge is instructed to treat the corpus as ground truth and to discount any case that strays from it, so a persuasive-but-baseless argument cannot move the verdict.

This grounding step is intentionally routed to frontier models because its accuracy directly determines forecast integrity. The cost is justified by the trust it protects.

4.7 Data flow

A forecast proceeds through five stages.

Stage 1: Ingest. User submits a market URL. The system pulls market metadata from the Polymarket or Kalshi API: rules text, resolution criteria, deadline, current prices, volume, category, the outcome set, and related markets. Parsed into a structured representation, and the scenario template for the market's category is selected.

Stage 2: Research. The research layer runs targeted web research on the specific market and synthesizes a sourced evidence corpus: the actors involved, the timeline, the established facts, and the open questions. The corpus is frozen for the rest of the run, so the forecast is reproducible from a fixed evidence base.

Stage 3: Debate. One advocate is instantiated per market outcome. Each argues the strongest evidence-based case for its outcome from the corpus, and the advocates rebut one another across several rounds. Every claim is checked against the evidence (§4.6), so unsupported assertions carry no weight.

Stage 4: Judge. An impartial judge, run as a self-consistency ensemble, weighs the competing cases against the corpus and emits a normalized probability distribution across the outcomes. Returned with the full reasoning trace and a divergence summary versus the market.

Stage 5: Log. The run is logged with full state for replay. When the underlying market later resolves, the forecast is scored by Brier against the realized outcome, feeding the public accuracy record.

4.8 Run pricing and compute margin

Run cost scales with market complexity, the number of outcomes, and the routing choices. With per-purpose routing and prompt caching, the metered model cost per run stays well below all-frontier execution.

Every run is billed in USDC. The live price is a single compute fee covering the run's model cost plus a margin that keeps the platform profitable on every run regardless of token price. Because model cost varies with market complexity, the fee is quoted as a band before the run and settled within it. One hundred percent of the compute margin funds the $NFA buy-and-burn described in §7.1. When the scenario marketplace opens (§5), a second, skill-priced scenario fee paid to the template's author is added on top.

5 · Scenario marketplace (planned)

5.1 Scenario pricing and author earnings

Every scenario carries a skill score from 0 to 100: a measure of how much better than the market its forecasts resolve, not raw accuracy. The score is the scenario's Brier skill against the market-implied probability at run time (see §6), so a scenario is only rewarded for beating the crowd, never for confidently calling outcomes the market already prices as near-certain. The scenario fee scales with that score:

scenario_fee = $1 + (skill_score / 100) × $19

where skill_score ∈ [0, 100]
// $1 floor for unrated or no-edge scenarios, ~$20 ceiling at top skill
// skill_score is the scenario's Brier skill vs the market baseline

A newly published scenario launches unrated at the $1 floor and earns its price as the markets it forecasts resolve. Early scenarios are therefore cheap to run, rewarding the traders who discover them, while proven scenarios command up to twenty times the floor and pay their authors accordingly. This rewards demonstrated forecasting edge, not publication or hype: a scenario that does not beat the market stays at the floor no matter how often it is used.

Authors are paid in $NFA. When a user pays the scenario fee in USDC, the protocol automatically buys $NFA on the open market and routes it to the author, so every paid run is direct buy pressure on the token and authors are aligned with the asset their work drives. Earnings are claimable on demand with no vesting; authors typically recycle them into new publications, each costing the $10 $NFA publishing burn (§7.1), which closes the loop.

5.2 Accuracy measurement

Accuracy is measured in two layers.

Backtest layer. When a scenario is published, it runs automatically against a historical event corpus with known outcomes covering 18 months of resolved Polymarket and Kalshi markets. A portion of the corpus is a blind test set authors cannot see during development. Published accuracy uses only the blind set, detecting overfitting.

Live performance layer. When a scenario forecasts a currently-unresolved market, its probability output is recorded. When the market later resolves, the Brier score is computed and added to the scenario's accuracy record.

Live performance weighting is determined by sample size. A scenario with 5 resolved markets has its accuracy estimate weighted toward the backtest baseline. A scenario with 100 resolved markets is weighted almost entirely on live performance.

5.3 Anti-gaming defenses

• Blind test sets prevent scenarios from being tuned to known historical outcomes.
• Unique-wallet counting ensures earnings come from distinct wallet usage rather than raw call volume.
• Similarity detection identifies near-duplicate scenarios. Publishes with similarity above threshold require differentiation or inherit reduced earnings.
• Accuracy weighting is the most fundamental defense. Pumped usage produces no earnings if forecasts are inaccurate.
• Evidence grounding (see §4.6) catches manipulation during the run, not just post-resolution: a case unsupported by the corpus is discounted before it can move the verdict.
• New scenario labeling marks scenarios with insufficient resolved-market data as unproven. Recommendations always include at least one established scenario.

5.4 Private scenarios

Authors can publish scenarios privately. Private scenarios are not visible in the marketplace, do not appear in recommendations, and earn nothing. They are usable only by their author. Private authors pay only the compute fee on their own runs, with no scenario fee on top, so personal use costs just the metered compute price.

Private scenarios support professional traders who use NFA as a personal forecasting tool without revealing methodology.

5.5 Usage-mining (live at launch)

Independent of the marketplace, NFA runs a usage-mining program from launch that rewards the act of running forecasts. For the first five months following public availability, the protocol emits 1% of total supply per month, roughly 0.0333% of supply per day. Each day's emission is split equally across every forecast run that day: on a day with three runs, each run earns 0.0111% of supply; on a busier day each run earns proportionally less. The reward goes to the wallet that funded the run, turning early usage directly into token distribution. Five months at 1% per month draws the entire Community / airdrop allocation (§7.2), with no separate budget line.

Because the daily pool is fixed, the per-run reward is richest exactly when usage is lowest, a built-in incentive to be early, and the open-market activity it drives feeds the same buy-and-burn flow that runs in steady state. The subsidy comes from token issuance the protocol controls, not treasury cash it does not.

For the rare market where reasoning adds nothing over a simple model, short-horizon price action driven purely by order flow, NFA continues to honestly display "no coverage" rather than producing a low-quality forecast.

6 · Accuracy as the north star

6.1 Why accuracy matters more than other metrics

Prediction-market trading is fundamentally about expected value. A trader who can identify mispriced markets earns money over time; a trader who cannot does not. The tool that helps traders identify mispricings is only valuable to the extent it is more accurate than the market it is being used against.

Accuracy is the only metric that matters at the foundational level. Every other decision, the engine's design, pricing, coverage, and the planned marketplace, is downstream of accuracy. The product is designed to produce accurate forecasts and to measure accuracy transparently.

6.2 Public track record

All accuracy data is public by default. Every forecast NFA produces is recorded, and when the underlying market resolves the run is scored; the public record breaks accuracy down by market category, with full details of every resolved run. Once the marketplace opens, each scenario template will carry its own public accuracy page.

Transparency serves multiple purposes: user trust (traders use tools they can verify), market efficiency (better information surfaces faster when performance is visible), accountability (no selective reporting, the record is complete), regulatory defense (transparent accuracy data distinguishes NFA from opaque forecasting services).

6.3 Scoring methodology

Brier scores are the primary accuracy metric. For a binary market with resolution outcome o ∈ {0, 1} and predicted probability p ∈ [0, 1], the Brier score is (p − o)². Lower is better; 0.0 is perfect, 0.25 is the score achieved by always predicting 50/50.

For multi-outcome markets, scoring generalizes to mean squared error across all outcome categories.

Accuracy is displayed as a 0-to-1 score where 1.0 represents perfect forecasting and 0.0 represents random baseline (0.25 raw Brier). More intuitive for non-expert users while preserving statistical properties. For the planned scenario marketplace, this is re-expressed as a 0–100 skill score measuring forecasting performance relative to the market baseline (§5.1): 0 means no edge over the market, 100 the demonstrated-skill ceiling.

6.4 Category-specific accuracy

A single aggregate accuracy number is misleading. NFA tracks and displays accuracy per market category. When a user submits a market URL, the accuracy shown is specifically the engine's accuracy on that market category.

6.5 Comparative benchmarking

NFA publishes monthly comparative benchmarks showing how its forecast accuracy compares to:

• Polymarket / Kalshi market consensus at forecast time;
• Naive baselines (always 50/50, category base rate);
• Single-LLM forecasts (frontier models on identical questions).

These benchmarks are the honest check on whether the platform adds value. If NFA's forecasts are less accurate than market consensus, the platform publishes that fact rather than hiding it.

6.6 Continuous improvement mechanics

• Automatic category expansion. When a category reaches sufficient activity without good coverage, the platform flags the gap and builds a new template for it.
• Staleness detection. Templates whose recent accuracy trends down are flagged for refresh.
• Winning pattern extraction. Analysis of top-performing forecasts informs engine-level improvements.
• Governance feedback. Accuracy data surfaces platform parameters that may need adjustment.

7 · Token economics

7.1 Token utility

$NFA is a Solana SPL token. Sub-second finality and low transaction cost make continuous open-market buy-and-burn viable at retail price points an EVM L1 cannot match. Runs are priced in USDC, which keeps margins legible and removes token-acquisition friction for new users; $NFA is the value-capture and access layer on top of that economy. Its live roles are below; additional roles activate when the scenario marketplace opens (§5).

Buy-and-burn from platform profit (live). Every run's compute fee carries a margin over its model cost. 100% of the resulting profit, net run revenue after model cost, is used to buy $NFA on the open market and burn it. Operations are funded separately from the project treasury, so platform profit flows entirely to the burn. The burn amount is non-discretionary, fixed at 100% of profit; the cadence will be finalized ahead of launch.

Usage-mining (live). A five-month launch program distributes the Community allocation to the wallets that fund forecast runs, turning early usage directly into token distribution (§5.5).

Governance (live). $NFA holders govern platform parameters: pricing, accuracy weighting, treasury policy, category weighting, and protocol upgrade decisions. Governance is weighted by staked tokens.

Author settlement (planned). When the marketplace opens, the scenario fee paid in USDC will be used to buy $NFA on the open market and routed to the template's author, so every paid run is direct buy pressure and authors earn the token their work drives (§5.1).

Publishing and staking (planned). Listing a template will cost $10 of $NFA (USD-pegged), burned on publication, pricing out spam; and running a proven template (skill score above 10) will require staking at least $1,000 of $NFA, gating the high-skill catalog behind skin in the game.

Token demand scales with real usage rather than speculation: platform profit continuously buys and burns supply today, and the marketplace adds author buy-pressure and further burns as it comes online.

7.2 Supply and distribution

Total supply is fixed at 1,000,000,000 NFA tokens.

• Team, 15%. Core contributors, long-term alignment, governance and operations.
• Seed investors, 10%. Honor commitments from initial raise.
• Liquidity, 15%. Initial liquidity provisioned across Solana AMMs at TGE (Raydium, Orca, Meteora), ensuring efficient price discovery and depth for the continuous buy-and-burn flow.
• Community / airdrop, 5%. Funds the five-month usage-mining program (§5.5); five months at 1% of supply per month draws the full allocation.
• Advisors / KOLs, 5%. Strategic contributors and distribution partners.
• Development, 25%. Engineering, infrastructure, audits, and team scaling beyond initial MVP.
• Ecosystem incentives, 10%. Accuracy bounties, trader onboarding incentives, and integration bounties for third-party developers.
• Partnerships, 10%. Strategic relationships with prediction-market venues, AI agent platforms, data providers, and exchanges.
• Long-term growth fund, 5%. Reserve for acquisitions, vertical expansion, and ecosystem programs not anticipated at launch.

7.3 Vesting schedules

All vesting is enforced on-chain through immutable contracts. No party has governance discretion to alter vesting after launch.

• Team · 6-month cliff, 18-month linear vesting.
• Seed investors · 12-month linear vesting from TGE (no cliff).
• Liquidity · locked 12–24 months.
• Community / airdrop · released over the 5-month usage-mining program (§5.5).
• Advisors / KOLs · 3-month cliff, 9-month linear vesting.
• Development · Ecosystem incentives · Partnerships · Long-term growth · voting-controlled release.

7.4 Circulating supply at TGE

Total circulating supply at TGE is approximately 20 to 25 percent of total supply. Team and advisors are fully cliffed at TGE. Seed investors carry no cliff and begin linear unlock from TGE day one over 12 months, visible, predictable, and small in any given week. The Development, Ecosystem incentives, Partnerships, and Long-term growth allocations are voting-controlled, so the protocol can release tranches as legitimate operational needs justify rather than mechanically unlocking on a schedule. This produces tighter float at launch than a typical cliff-and-vest schedule and avoids the supply overhangs that distort price discovery in the first months post-TGE.

7.5 Value accrual

Compute-profit burn (live). 100% of run profit, the compute margin defined in §4.8, is used to buy $NFA on the open market and burn it. This is the primary, usage-scaled deflation mechanism, observable on-chain and tied to platform profitability rather than speculation.

Author-settlement buy pressure (planned). When the marketplace opens, each scenario fee will be converted from USDC into an open-market $NFA purchase routed to the author, continuous buy-side demand that scales with marketplace volume; authors who hold or stake their earnings reduce velocity further.

Publishing burn (planned). Each new template will cost $10 of $NFA, burned on publication, a steady permanent supply sink that grows with the marketplace.

Staking lockup (planned). Accessing proven templates (skill score above 10) will require $1,000 of $NFA staked, removing float from circulation and scaling with the most engaged users.

Net supply. In the bootstrap phase, treasury operational spend and usage-mining emissions partly offset the burn; as forecast volume grows, the compute-profit burn outpaces treasury spend and net supply turns deflationary. The marketplace sinks above add further deflation once live. The crossover is a function of usage, not of token-price speculation.

8 · Governance

8.1 Governance philosophy

NFA progressively transitions from core-team-managed operations to DAO-governed protocol over 18 to 24 months post-launch. Pragmatic rationale: early period requires rapid parameter iteration based on real-world data; mature period broadens to token-holder base.

8.2 Governance scope

• Scenario skill-price curve (floor, ceiling, slope)
• Accuracy weighting formula and time windows
• Per-purpose LLM routing defaults
• Staking access threshold (currently $1,000 of $NFA for skill > 10)
• Publishing fee and usage-mining parameters
• Category taxonomy and weighting
• Verification badge thresholds
• Treasury allocation decisions
• Partnership and integration decisions over a defined threshold
• Protocol upgrade approval

Core mechanics affecting user/author funds require supermajority approval (two-thirds of voting tokens). Parameter adjustments within defined ranges require simple majority.

8.3 Governance structure

Voting power is held by NFA token holders, weighted by tokens held or staked. Staked tokens receive a multiplier (up to 1.5×). Proposals can be created by any token holder meeting a minimum threshold. Three phases: discussion (7 days), voting (7 days), execution (48-hour timelock).

8.4 Transition timeline

• Months 0–6. Core team retains decision-making. DAO advisory only.
• Months 6–12. Non-core parameters transition to binding DAO governance.
• Months 12–24. Core mechanics progressively transition to DAO.
• Month 24+. Full DAO governance. Core team operates infrastructure under protocol-defined parameters.

9 · Roadmap

9.1 Phase 1: Token Generation

TGE executes on Solana. Liquidity deployed across Solana AMMs. Vesting programs activate. Legal opinion finalized.

9.2 Phase 2: Build & Internal Testnet

Core forecasting engine operational end-to-end: research, advocate debate, and judge, with deterministic execution and full replay. URL-paste flow. MCP server with initial tool surface. Backtesting harness with blind test set. Internal scenario templates for the first market categories. Repository with clean license structure.

9.3 Phase 3: Public Testnet

Public testnet launch. Full accuracy tracking live, with public per-category track records. MCP server listed in public registries. Bug bounty opens. Tier-1 audit begins on settlement contracts.

9.4 Phase 4: Pricing & Paying Users

Metered USDC billing begins: the compute fee with its margin, and 100%-of-profit $NFA buy-and-burn, both live. The five-month usage-mining program (§5.5) opens with public availability and runs through this phase. First paying users.

9.5 Phase 5: Growth & Maturity

The scenario marketplace opens: domain experts tune and publish templates, with skill-based scenario pricing, $NFA author settlement, publishing burns, and access staking (§5). Tier-2 CEX listings, with Tier-1 targets contingent on volume. First DAO governance proposals, transitioning to full DAO governance per §8.4. Enterprise tier. Long-term accuracy track record becomes the primary marketing asset.

10 · Risks and mitigations

This section exists because most token launches treat risk disclosure as boilerplate. Done properly, it is a competitive advantage. The risks below are the real ones.

10.1 Accuracy risk

The product thesis depends on NFA producing forecasts meaningfully better than market consensus on a durable basis.

Mitigations. The five-month usage-mining program seeds early usage and accuracy data; transparent accuracy tracking and honest reporting including underperformance; category-specific accuracy display; continuous feedback loops; honest coverage gaps rather than forced low-quality forecasts.

10.2 Regulatory risk

Prediction-market tooling operates in complex regulatory environments. SEC, MiCA, CFTC interactions. Influencer promotion, token distribution, cross-border access introduce surface.

Mitigations. Platform operates as forecasting infrastructure not as a market; the platform produces analysis, not advice, and disclaims liability for specific forecasts; geographical restrictions at hosted-frontend level (US blocking); KYC on token claims above thresholds; legal opinion procured before TGE; advisor and KOL allocations publicly disclosed and vested.

10.3 Licensing and intellectual property risk

Mitigations. Open-source components (frontend, MCP server, settlement contracts) under Apache 2.0 / MIT; proprietary engine clearly delineated; NOTICE files and attribution maintained; license compliance reviewed in audit scope.

10.4 Marketplace quality risk

Open marketplaces attract gaming, spam, low-quality contributions.

Mitigations. Accuracy-weighted earnings structurally disadvantage bad actors; blind test sets prevent overfitting; similarity detection prevents plagiarism; unique-wallet counting prevents farming; plausibility validation catches manipulation at runtime; new scenario labels prevent unproven work from displacing established scenarios.

10.5 Compute economics risk

Simulation costs scale with LLM inference pricing.

Mitigations. Per-purpose routing optimizes cost-quality tradeoff per call class; prompt caching materially reduces input-token costs on supported providers; OpenRouter provides provider-agnostic routing; the platform's competitive position improves as LLM prices fall over time.

10.6 Cost tracking risk

Author payouts and the compute margin both depend on accurate per-call cost tracking. Production experience in adjacent systems shows cost-tracking failures can underreport spend by orders of magnitude when fallback chains are hit silently.

Mitigations. Per-call cost-record persistence with provenance for how each cost was computed; recurring reconciliation against provider billing APIs with discrepancies hard-flagged for manual review before author payouts settle; settlement gated on reconciliation confidence.

10.7 Key-person and operational risk

Mitigations. Comprehensive documentation; open-source codebase for non-engine layers; progressive DAO governance reduces concentration; operational playbooks; ability to run alternative hosted implementations once protocol is DAO-governed.

10.8 Adversarial scenario authorship

A malicious actor could publish scenarios designed to manipulate markets rather than forecast them.

Mitigations. Evidence grounding (§4.6) discounts unsupported arguments during the run, not after market resolution; in the marketplace, accuracy-weighted earnings make manipulation economically unviable; template authorship is transparent; divergence monitoring catches anomalous outputs; on-chain record of template changes; reputation accumulates over many markets.

10.9 Market category coverage risk

Different categories have different accuracy characteristics.

Mitigations. Honest coverage decisions; category-specific accuracy display; the engine only forecasts categories where it has demonstrated accuracy; clear platform communication about which categories the engine suits.

11 · Closing

NFA sits at the intersection of three things that have matured in the last 18 months: prediction markets as a credible financial category, multi-agent AI systems as production-grade infrastructure, and MCP as the standardizing protocol for AI agent distribution. The product occupies a specific niche, actor-driven forecasting for prediction-market traders, that no incumbent serves.

The platform's design reflects a deliberate choice: accuracy is the north star, everything else follows. The engine architecture, pricing, governance transition, coverage decisions, and token economics all optimize for measurable forecasting accuracy. Opaque metrics and narrative-first positioning are rejected in favor of public data and honest disclosure.

Today the engine forecasts every market itself, billed in USDC with all platform profit flowing to $NFA buy-and-burn; a scenario marketplace where domain experts contribute templates and earn on real-world accuracy is the planned next phase. The technical engine is production-grade from day one: deterministic execution with full replay, per-call cost reconciliation, per-purpose LLM routing, prompt caching, and parallel debate, because forecast integrity demands it.

The risks are real and described openly. The mitigations are concrete and reflected in architecture, policies, and planned execution. The cost of doing this properly is higher than shipping a typical token launch. The cost of doing it improperly is unbounded.

NFA is the collaborative-intelligence brand made literal as product. A coordinated mind composed of specialists, each contributed by domain experts, deliberating to produce accurate forecasts, compensated by a transparent economy. What the brand has been promising for a year, now delivered.