Technical Architecture: The Intelligence Stack

Miko's intelligence is not a single model responding to prompts. It is a production system composed of specialized layers, each engineered for a specific function in the pipeline from raw data to on-chain execution. This section details each layer — what it does, how it works, and why it matters for holders.

Architecture Overview

Layer 1: Data Ingestion and Semantic Filtering

The first layer is the system's sensory apparatus. It connects to the Twitter API, on-chain data sources, and — for the chain's tokenized equities — traditional equity market data feeds (Finnhub and Yahoo Finance) to collect raw information in real-time. It also ingests a curated smart-wallet feed: the on-chain buys and sells of proven Robinhood Chain traders. Unlike social data, wallet events are not taken at their word — each one is verified against the chain itself (the transaction receipt must show the exact token moving to the exact wallet) before it can influence anything downstream.

Raw data is not useful data. This layer applies NLP-based analysis and heuristic rules to:

1. Filter noise — remove irrelevant content (non-crypto, spam, duplicate information)
2. Score relevance — evaluate each data point for uniqueness, community relevance, and potential engagement
3. Classify content — categorize by topic (DeFi, memecoin, infrastructure, etc.) and sentiment (positive, negative, neutral)
4. Cluster related discussions — identify when multiple KOLs are discussing the same topic independently, using HDBSCAN clustering on TF-IDF vectors and embedding similarity

The clustering is critical for the fact-checking pipeline downstream. When three independent KOLs make the same claim, it carries different weight than a single source making a claim that nobody corroborates:

$$\text{Source Weight}(claim) = \log\left(1 + \frac{n_{\text{distinct\_authors}}}{\alpha}\right) \times \text{avg\_persuasion\_score}$$

Where $n_{\text{distinct\_authors}}$ is the count of independent sources making the claim and $\alpha = 0.8$ is the power law exponent derived from social media attention distribution (Newman, 2005).

Layer 2: Knowledge Graph and Multi-Source Fact-Checking Engine

This is the layer that most fundamentally differentiates Miko from other AI agents.

Knowledge Graph

Miko does not treat each piece of information as isolated. Instead, it integrates data into a structured Knowledge Graph stored in proprietary database with vector embeddings for semantic search. This graph maps relationships between tokens, KOLs, market events, and projects. When Miko encounters new information about a token, it retrieves all relevant context — past performance, associated KOLs, historical claims, and verification status — enabling informed decisions rather than reactive responses.

The semantic search capability uses cosine similarity on embedding vectors:

$$\text{relevance}(q, d) = \frac{\vec{q} \cdot \vec{d}}{||\vec{q}|| \times ||\vec{d}||}$$

Results above the configured importance threshold ($\tau = 0.4$ by default) are surfaced as contextual grounding for both content generation and asset selection.

The Fact-Checking Pipeline

Most AI agents in the crypto space process LLM output directly — what the model generates is what gets published or acted upon, with no independent verification step. When such agents control financial actions (trading, liquidity provision, token deployment), the cost of unverified information becomes real capital loss.

Miko's Fact-Checking Engine is a four-stage adaptive pipeline that consults six independent verification providers:

Key design decisions:

• AI-driven strategy selection: The verification level is not hardcoded. The AI evaluates each claim's characteristics (specificity, importance, independent source count) and decides how to verify it, selecting both the intensity and the specific providers whose strengths match the claim type.
• Adaptive fallback: If a provider fails (API error, timeout), the system excludes it and re-plans the strategy with remaining providers. Up to 2 additional attempts. Temporary outages never block verification.
• Evidence convergence requirement: A claim is verified only when evidence from independent sources converges on the same conclusion. Divergent evidence or insufficient data results in a "not verified" verdict.
• Specificity-aware evaluation: The synthesis stage checks whether specific details (names, numbers, dates, amounts) match the evidence, not just the general topic.

Where Fact-Checking Is Applied

The pipeline operates at three independent checkpoints:

For holders, this means: Information that feeds into Miko's asset selection has been subjected to multi-source verification. In a market where a single fabricated announcement can move a token's price dramatically, this verification layer is what stands between the holder's allocation and a selection based on false information.

Layer 3: Persona-Driven Generation Core

This layer creates Miko's voice: the public-facing content on X (Twitter) that drives community engagement and growth.

Multiple reaction modes allow Miko to respond dynamically to different situations — from cheerful curiosity for friendly exchanges to sharp wit for provocative interactions to analytical depth for market discussions. These modes are not random; the system evaluates the conversational context and selects the most appropriate response style.

Beyond conversational content, the Generation Core publishes structured market reads: when a token takes over the top of MIKO's attention board, Miko posts a curator's read — what is actually carrying the move, what kind of table it is, and the specific on-chain meter that decides where it goes next. These reads are grounded in the analyst layer's measured facts (including observed holder-structure data) and always end on the watch-point, never on a verdict — Miko is a curator, not a trading-signal account.

Every response Miko generates is verified for originality before posting. The system ensures content is genuinely Miko's own expression, not a mechanical repetition of source material. This is enforced through multiple independent verification layers at the code level.

Miko's persona is also published as an open model: a fine-tuned open-weights release on Hugging Face and Ollama that carries her voice in the weights themselves, callable through the live REST API and MCP server. See Open Model.

For holders, this matters because: Miko's social presence is the primary driver of community growth. Community growth drives trading volume. Trading volume generates tax revenue. Tax revenue funds the weekly acquisitions. The quality and authenticity of Miko's social output directly impacts the size of the weekly acquisition treasury.

Layer 4: Selection Algorithm

This is the financial brain of the protocol. It runs two decision streams — the weekly core selection that determines what the core sleeve buys, and the continuous attention-leader tracking that steers the satellite sleeve. It is the most critical component for holder allocations and the most technically sophisticated.

The Model Tournament: Three Phases, One Live Seat

The Selection Algorithm is built as a live tournament between competing models. Three statistical phases — each a real model family requiring progressively more data — enter the tournament as training data accumulates, and the weekly core selection is always made by the current champion.

How the seat is decided:

• Blind predictions first. At each weekly selection, every entrant scores the same frozen candidate view, and its complete ranking is recorded before any outcome exists. Predictions are immutable — there is no backscoring.
• Outcomes as the judge. When a selection group's 7-day outcomes are complete, each entrant is scored by the Spearman rank correlation between its recorded ranking and the candidates' realized 7-day returns — every entrant against the identical candidate set and identical labels.
• The champion is the cumulative leader. The live selector is whichever entrant leads the accumulated paired record. Promotion and demotion are the same rule in both directions: a model takes the seat by out-predicting the incumbent on the shared record, and loses it the same way. Both directions are automatic.
• Sequential entry. Phase 1 — a Bayesian linear return model fitted over the frozen candidate features, with its prior anchored to the deterministic baseline — enters as soon as the first complete training group exists. Phase 2 — Thompson sampling that draws one reproducible coefficient sample from the Phase 1 posterior per selection and scores all candidates with that same draw — follows. Phase 3 — CatBoost learning-to-rank with YetiRank loss, one ranking group per selection — enters as the dataset grows. Each new entrant competes from its first week: it is scored, not trusted.
• An honest history. When the evidence basis widened from attention alone to attention plus verified capital, the record was not rewritten. The earlier social-only models remain in the same records as reference benchmarks — scored on every selection group they can represent, visible alongside the current roster, but no longer eligible for the seat. The track record keeps its past instead of replacing it.
• The permanent baseline. The deterministic fusion scorer anchoring the roster is never retired. If no learned model can out-predict it, it simply keeps the seat — a floor under the whole system. Random selection and the chain's benchmark asset are recorded alongside the roster as reference floors, visible in the same records but never eligible for the seat.

The Deterministic Baseline: Attention Acceleration

The roster's deterministic anchor is a scorer built for one question: where is verified, independent conviction arriving right now? Its first axis is attention:

• Independent-author corroboration: selection input is KOL coverage with one vote per distinct author — an author's repeated posts collapse into a single voice, so posting volume cannot substitute for breadth. A minimum number of independent authors in the current window is required before a token is rankable at all.
• Persuasion-weighted support: each author's voice is weighted by the persuasion score of their posts, so reasoned conviction counts more than reflexive hype, and low-persuasion noise shrinks toward zero.
• Acceleration, not level: support is measured against the token's own trailing 7-day baseline. A token that has been loud all week must keep exceeding its own normal to stay on top; a quiet token that suddenly draws broad, persuasive attention registers immediately.

The Second Evidence Axis: Verified Capital

Attention tells the system what the market is saying. Since the multisignal upgrade, the scorer reads a second, independent axis: what a curated set of proven on-chain traders is doing with real money. The two axes are fused into one score — and the capital axis is engineered so that it corroborates attention without ever being buyable:

• Receipts, not reports. A wallet event enters the pipeline only after the chain itself confirms it: the transaction receipt must show the exact token transferring to the exact wallet. An unconfirmable event is held for review, never scored.
• One actor, one voice — on-chain too. Just as the attention axis counts one vote per distinct author, the capital axis counts one bounded unit per effective actor: wallets that can be linked are collapsed into a single cluster, and a cluster's contribution is the mean of its buys' time-decay weights, not the sum:

$$w_{\text{actor}} = \frac{1}{|B|} \sum_{b \in B} 2^{-\Delta t_b / \tau}$$

Twenty buys from one actor are one voice, and re-buying cannot pin that voice at maximum freshness.

• Size saturates. On the satellite board, an actor's conviction is scaled by a saturating size factor measured against the current market's own median buy, $\sigma = U / (U + \tilde{U})$ — an actor buying the median size carries weight 0.5, and no position, however large, can even double that. Breadth of independent conviction outranks depth of any single pocket.
• Sells count against. A tracked actor distributing a position decays its earlier support; a sell with no matching buy registers as bounded negative evidence. The capital axis reads exits as honestly as entries.

The result is a scorer that cannot be gamed from either side alone: posting volume collapses into one social voice, and buying volume collapses into one capital voice. Moving the board requires what a genuine narrative produces naturally — many independent people saying it, and many independent wallets doing it.

Outcome Measurement

Two layers of outcome data are recorded for every weekly cycle.

Training labels. The learning models train on — and are judged by — the realized 7-day return of every candidate in the frozen selection group, not only the announced winner. All-candidate labels are what make rank correlation measurable: the system records not just how its pick performed, but how everything it ranked performed.

The public track record. Each announced selection additionally receives a multi-metric composite score for the public track record. The weights are derived from empirical research on cryptocurrency success factors:

$$S_{\text{composite}} = \sum_{i=1}^{n} w_i \cdot x_i$$

Metric ($x_i$)	Weight ($w_i$)	Research Basis
Selected asset price performance	0.40	Chen et al. (2023): 45-55% range midpoint
$MIKO holder growth rate	0.22	Liu & Tsyvinski (2021): network effects
$MIKO token price performance	0.20	Eisenmann et al. (2006): indirect effects ≈ 50% of direct
Community sentiment	0.10	Kraaijeveld & De Smedt (2020): 8-12% range
Tweet engagement	0.05	DeFi Alliance (2024): indirect indicator
$MIKO volume growth	0.03	Brandvold et al. (2015): 3-7% range lower bound

$$\sum_{i=1}^{6} w_i = 1.00$$

Champion Accountability

Degradation protection is built into the seat itself rather than bolted on as a separate mechanism. The champion is re-derived from the cumulative record every time a selection group's outcomes mature: a model that stops out-predicting its competitors loses the live seat to whichever entrant now leads — including back to the deterministic baseline. Because demotion is the same automatic rule as promotion, a poorly calibrated model cannot persist in making selections with real capital, and every seat change is recorded with the standings that caused it.

Token Quality Filters

Before any token can be considered as a selection candidate, it must pass a two-tier quality assessment.

Tier 1: Hard Threshold Filters

Filter	Threshold	Rationale
Minimum Market Cap	$2,000,000	Below this, manipulation risk is elevated (Kaiko Research, 2025; calibrated to Robinhood Chain market scale)
Minimum 24h Volume	$500,000	Ensures sufficient liquidity for purchase execution
Excluded Tokens	$MIKO	Prevents conflict-of-interest in self-selection
Exempt Tokens	$WETH	Benchmark asset exempted from quality filters

Tier 2: DEX Market Structure Assessment

Candidates that pass the hard filters are then screened through a multi-factor market structure analysis using real-time DEX data. The assessment is a safety screen: structural red flags veto a candidate for that cycle rather than adjusting its score. Signals evaluated include:

• Order flow analysis: Buy/sell pressure ratio — measures whether a token is under net accumulation or distribution
• Transaction velocity: Rate of on-chain transactions — accelerating velocity suggests growing organic interest
• Breakout readiness: Technical positioning relative to recent price range — identifies tokens at potential inflection points
• Relative strength: Performance relative to the broader Robinhood Chain market — filters for tokens showing independent momentum
• Holder breadth: Distribution of token holders — wider distribution suggests healthier, less manipulable markets

Structural Vetoes:

Severe risk flags — concentrated ownership, one-way order flow, thin exit liquidity — remove a candidate from consideration for that cycle outright, regardless of how strongly the attention scoring favors it.

This two-tier system ensures that candidate assets both clear minimum quality thresholds and pass a market-health and manipulation-risk screen before any selection is finalized.

Equity Market Context for Stock-Token Candidates

Robinhood Chain's tokenized equities are not evaluated with crypto metrics alone — a stock token's on-chain price is a window onto a real company trading in a real market, and MIKO reads that market natively through a dedicated equity data pipeline:

Signal	Source	What it tells the model
Real-time quote & day move	Finnhub	Where the underlying equity is trading right now
Week / month price history, multi-window trend	Yahoo Finance	Whether a move is a day event or a developing trend
52-week range position (0–100)	Yahoo Finance	Whether attention is arriving near highs, lows, or mid-range
Volume anomaly vs. prior-20-session average	Yahoo Finance	Whether the underlying market itself is unusually active
Earnings calendar & proximity flags	Finnhub	Whether a known scheduled catalyst is days away
Nominal NY trading session	Schedule-derived	Whether the underlying market is open, closed, or pre/post — so a flat on-chain price during a closed session is never misread as apathy

Two further layers calibrate these signals to this chain:

• Chain-local size calibration: "large" and "small" are defined by the measured size distribution of the chain's own tracked equity universe — liquidity, market cap, and volume quantiles computed from live data — not by legacy-market intuitions imported from elsewhere.
• Prior-evaluation loop: every published stock read is stored and later scored against the equity's subsequent market data on a fixed cadence, so the system's equity judgments accumulate the same measurable track record as its crypto selections.

This is what "reading both engines of the chain" means in practice: community tokens are read through on-chain flow and social verification, and tokenized equities are read through the actual equity market they mirror.

Community Suggestions and Persuasion Analysis

Community members can put a token on Miko's radar by mentioning her (@mikorithm) with a $SYMBOL tag. Every recommendation is read and evaluated — the PostAnalyzer scores its persuasiveness from Miko's own perspective:

$$\text{persuasion\_score}(tweet) = f(\text{authenticity}, \text{reasoning\_depth}, \text{community\_alignment}, \text{ecosystem\_relevance})$$

A recommendation that clears the spam filter and the persuasion screen admits the token into Miko's tracked candidate universe: its on-chain identity is resolved, its pool is verified, and from that point its attention and market history are recorded continuously. Discovery is where community input carries real power — a token that no one on the KOL radar has named yet can enter Miko's field of view through a single well-argued mention.

Selection weight, however, is earned through independent corroboration: the ranking that decides the weekly asset counts one vote per distinct author on the KOL radar, persuasion-weighted and measured against the token's own baseline. Spamming a $SYMBOL is ineffective by construction — repeated posts from one account collapse into a single voice, and no volume of self-promotion can substitute for broad, independent conviction.

The Satellite Stream: Gated Attention Tracking

The satellite sleeve's decision stream runs continuously rather than weekly. MIKO's attention board ranks community tokens in real time on the same two fused axes — author-deduplicated, persuasion-weighted attention and receipt-verified smart-wallet capital, each measured against the token's own baseline — and the board's leader is the satellite's target.

The switch-or-hold decision itself runs as a tournament, mirroring the weekly core seat at rotation speed. On every decision edge, the evidence is sealed once and every roster entrant reads the same sealed view; the seated entrant's verdict is the one that acts, matured decisions are scored against realized short-horizon outcomes, and the seat follows the standings on a daily review — automatically, in both directions. Any inability to decide resolves to HOLD: the sleeve fails closed, never open.

A leader change does not rotate the sleeve by itself — the seated entrant's rotation must still pass evidence-based veto gates:

• Spike discipline: a candidate whose day combines an outsized price move with abnormal volume (both thresholds measured, not guessed — they were selected from an out-of-sample backtest across 404 tokens and 64 candidate rules, of which only this pattern showed consistent directional evidence) is vetoed for that day. The gate exists because such days statistically precede underperformance — the rotation waits rather than buys the hangover.
• Structural safety vetoes: minimum market cap, DEX market-structure risk flags, and a re-entry cooldown that prevents rotating back into a token the sleeve just exited.
• Execution-impact discipline: approval to rotate is not approval to buy at any cost. At claim time, the execution layer measures the actual swap route's price impact at the sleeve's real size; a market too thin to absorb the order within a 3% impact tolerance is not force-bought — the sleeve takes WETH for that claim instead, visible on-chain in the claim's own purchase transaction. Holders are never handed the slippage bill for a market that could not carry the position.

The attention board's reads are further enriched with observed holder-structure data from public, parameterized on-chain queries (Dune Analytics): what share of a token's recent buyers are first-time entrants within the scan window, what share of earlier buyers have not sold, and how concentrated the buying is — measured DEX activity only, never inferred "holding" claims.

A vetoed rotation is deferred, never forced: the sleeve simply keeps its current position until a candidate passes. Every gate verdict is recorded with its full input signals, making the satellite's behavior as auditable as the weekly selection's.

The Self-Improvement Loop

Every weekly cycle generates outcome data that feeds back into the models:

This creates a closed-loop system where the selection intelligence becomes progressively more refined: every completed week adds one more fully scored group to the shared record, the learning models retrain on the grown dataset, and the live seat follows the accumulated evidence.

Verifiable Track Record

Most AI agent projects ask holders to trust their intelligence based on narrative — follower counts, ecosystem metrics, or team credentials. MIKO's system is designed to be measurable by default.

Every asset selection is automatically recorded with its full context: the asset selected, the exact tournament entrant and model artifact that made the decision, every entrant's blind ranking of the full candidate set, and — critically — the realized price performance of every ranked candidate at 24 hours and 7 days after announcement. Each announced selection receives a Composite Outcome Score for the public record, while the all-candidate outcomes feed the models' training data.

Champion changes and each cycle's tournament standings are logged with the record that produced them, creating a complete, queryable history of how the AI's decision-making has evolved over time.

This means MIKO's AI's intelligence is grounded in an auditable track record against objective outcomes. The Selection Track Record and MIKO's Insight Dashboard make this data publicly accessible, allowing holders and prospective investors to evaluate the AI's performance directly from the data.