How the thesis is tested.

Principles

1. Every number has an address.

   Each evidence item records its source URL, retrieval time, period, transformation and the SHA-256 of the stored provider response.

2. Four labels, never mixed.

   OBSERVED (read from a source), DERIVED (computed here from observed values), REPORTED (stated by a third party), ASSUMED (a scenario choice). Assumptions are never presented as evidence.

3. Missing means missing.

   If a source fails or needs a credential that is not configured, the item is marked unavailable with the reason. Nothing is estimated to fill the gap, and no fixture is substituted in production.

4. Matched windows only.

   Ratios combine series over identical dates from the same provider scope. Flow metrics use the 90 days ending the day before the latest complete day; peer comparisons use one shared 30-day window.

5. Frozen at the start.

   A debate snapshot is written once and never updated. Replays and reports read only that snapshot. A rematch creates a new snapshot and shows the differences.

6. Code decides what is true.

   The AI can only cite snapshot IDs. The server validates every citation and every spoken number before a turn is saved, and applies score penalties itself.

The price ladder

Rungs: $10, $25, $50, $100, 10 years from the data date. Every number is computed by deterministic code from stored responses; no AI writes any number, status or verdict. Ladder version three-digits-2.0.1.

Layer A: what each rung takes

supply_2036          = circulating supply grown at the issuance rule, monthly, for 10 years
market_cap           = rung price × supply_2036
value_needed         = market_cap ÷ cash-flow multiple        (burns + buybacks reaching holders per year)
intents_volume       = (value_needed − burns at today's run rate) ÷ buyback revenue per dollar of volume
growth_needed        = (intents_volume ÷ today's run rate)^(1/10) − 1
dex_share_2036       = intents_volume ÷ (spot DEX run rate × (1 + 20%)^10)
share_multiple       = dex_share_2036 ÷ (Intents run rate ÷ spot DEX run rate today)

Valuation multiples (cash-flow framework)

Market cap divided by the value reaching holders: DefiLlama holders revenue (burns and buybacks) for every asset, the same 90-day run rate, windows ending on the same day (Nov 10, 2026). Read live: Priced like Hyperliquid 19.3x (trailing year 24.8x); Priced like Tron 101x (trailing year 88.8x); NEAR's multiple today 202x (trailing year 497x). NEAR's value is gas burned plus Intents revenue designated for buybacks, the same channels that are projected forward, each with its own driver. The 90-day run rate is the primary basis because NEAR's buyback mechanism started on 2026-02-23; trailing-year multiples are shown for every asset as a check. Applied to NEAR in ten years as assumptions, never forecasts.

Supply

Current rule: 2.5% maximum issuance is the base case. HSP-029: 2.5% falling to 1.6% is shown alongside: Approved by House of Stake on 2026-11-02; linear cut over 24 months after a ~90-day buffer; needs validator adoption and can be vetoed until 2026-11-16.

Buyback revenue per dollar of volume

Intents revenue DefiLlama designates for buybacks divided by Intents volume over the same 90 days (LAD-HOLDER-TAKE-RATE; equal to take rate × buyback share). Held constant: fee competition is the largest single risk, since halving it doubles every volume requirement. Designated buybacks are a classification; NEAR delivered to the buyback wallet is reconstructed separately onchain.

Spot DEX market

DefiLlama's DEX overview mixes categories: spot DEXs, prediction markets, bridges (NEAR Intents is listed as a Bridge), launchpads and trading apps. Its per-protocol breakdown is summed per day by category and only the "Dexs" category is used: $2.88T a year (run rate) against $3.34T for the whole overview. Grown 20% a year in the base case (10% as a low case); it grew 26.3% a year from 2021 to 2025. Intents' ratio to it is a scale reference, not a market share.

Rank

The rung market cap against today's largest 40 crypto assets (CoinGecko), excluding stablecoins and wrapped or tokenized assets. It compares a future value with today's market, so it is a sense of scale, not a forecast rank.

Platform framework

A separate framework, never added to the cash-flow ladder and never used in its tables. Platforms trade at thousands of times the value reaching holders because they are valued on the economy that lives on the chain. For every rung, the page shows the stablecoins and DeFi TVL NEAR would need at the market cap to stablecoins and market cap to TVL ratios of Ethereum, BNB Chain, Solana and Tron (DefiLlama).

Bands: how hard each rung is

Each rung and multiple is placed in a band by how many times today's ratio of Intents volume to spot DEX volume it would need in ten years.

• ≤ 1.5x

  Keeps about today's size relative to spot DEX volume

• ≤ 5x

  Grows up to 5x relative to spot DEX volume

• ≤ 10x

  Grows 5x to 10x relative to spot DEX volume

• > 10x

  Needs more than 10x its size relative to spot DEX volume

Layer B: what we can prove

Reverse view and checkpoints

Implied price in ten years = (today's Intents run rate × (1 + growth)^10 × buyback revenue per dollar of volume + burns at today's run rate) × cash-flow multiple ÷ supply in ten years, for low growth (10% a year), middle growth (20% a year), high growth (40% a year). These are scenarios, so results can fall below today's price. Checkpoints are the yearly volumes on each rung's required growth path at NEAR's own multiple today, so the thesis can be checked every year. The swing factor compares the spread of implied prices across growth scenarios with the spread across the cash-flow multiples; spreads within 1.25x of each other are reported as similar.

The headline sentence on the home page is generated from these numbers by code. The bands and growth scenarios are our judgment, set while building the pipeline and published so anyone can change them in src/lib/research/ladder.ts and rebuild.

Token economics

Every token flow is classified before anything is summed. Version w4-2.0.0.

staking_ratio        = sum(current validators' stake) ÷ total supply            (NEAR RPC)
gross_issuance       = realized net supply growth + burns ÷ market cap
nominal_yield        = gross_issuance × (1 − protocol_reward_rate) ÷ staking_ratio
real_yield_staker    = (1 + nominal_yield) ÷ (1 + net supply growth) − 1
real_yield_holder    = 1 ÷ (1 + net supply growth) − 1                       (does not stake)
circulating_growth   = CoinGecko market cap ÷ price, last day vs a year earlier  (every asset)
holder_yield         = holders revenue, 90-day run rate ÷ market cap         (DefiLlama, every asset)
net_holder_yield     = holder_yield − circulating_growth

Circulating supply from market cap divided by price is one method for every asset, so it includes unlocks that a protocol issuance rate leaves out. One-day changes above 5% are listed as reclassification or unlock events rather than smoothed away. Net holder yield excludes price, so it is a measure of token economics, not a return forecast.

Economy map

Every NEAR product is investigated on its own and placed in one column: observed activity, verified value capture for $NEAR (NEAR delivered to the buyback wallet, or burns), or speculative opportunity, which is never given a dollar value. The fee base is DefiLlama's fees overview filtered to the NEAR chain: it includes NEAR Intents' cross-chain swap fees, which are paid on many chains and are not NEAR L1 gas. L1 activity compares matched 90-day windows a year apart (NearBlocks transactions and active accounts, DefiLlama gas fees). Chain Signatures is read from the v1.signer contract's state over RPC. Near Perps is a front end for Hyperliquid; its builder-code fees accrue to the front end, not to $NEAR.

Competition

NEAR Intents is compared with sixteen cross-chain protocols on the same DefiLlama definitions and the same 90 days: volume where an adapter exists, fees, protocol revenue, holders revenue and take rate. Categories are never added together: only swap venues with public volume enter the volume share, so the share is of measured venues, not of all cross-chain execution. Token valuations use execution protocols with a token (swap venues, bridges, solvers); the median is shown with and without tokens under $25M, as a range, because the peer set is small and thinly traded. Valuing Intents revenue at that range is a conditional comparison, not a fair value of NEAR, whose market cap also prices the L1 and its other products.

Valuation and alternatives

Two revenue bases are computed side by side and never mixed: DefiLlama's designated buybacks, and NEAR delivered to the buyback wallet (valued at market prices on delivery days). On each basis the multiple, the take rate and the projected flow use the same definition and the same 90 days, ending on one documented cutoff; peer multiples do not depend on the basis. NEAR's multiple history only uses windows that fall entirely after the flow began. For the alternatives, each asset is valued with the method that fits it (BTC monetary, platforms on their on-chain economy, Hyperliquid on cash flow, NEAR on both); the growth needed for the same price multiple holds each valuation ratio fixed and lets each asset's ownership dilution persist. That is a model result, not evidence about which asset is likely to deliver it. Risk uses daily DefiLlama prices: volatility, drawdown and correlation with BTC on consecutive days only.

Hypotheses

Ten hypotheses (research/three-digits/SCOPE.md) were written before the data was read, each with the rule that decides it: supported, not supported, mixed, or untested when the data needed is not available. Claims were narrowed after review to what each test measures: H1 tests the concentration of fees attributed to the chain, H2 reports an observed valuation gap, H4 diverging trends, H5 ownership dilution on raw and event-adjusted supply, H6 a fixed-valuation model result, H10 dependence on BTC only. H7 to H9 need BigQuery or the NEAR Intents Explorer API.

Verdict and audit

The investment verdict is fixed prose, finalized after two external review passes. Every lead, trail and characteristic beneath it is generated by a rule from the data in src/lib/research/verdict.ts, so a refresh can never leave a claim the numbers no longer support. Characteristics that can read either way (high turnover, low correlation with BTC) are listed as neither a lead nor a trail. Every correction, its resolution and whether it changed the ladder or the verdict is recorded in research/three-digits/AUDIT-LOG.md.

Evidence rules behind the ladder

Supporting evidence from NEAR's track record and token economics, shown in the evidence section of the thesis. Each sub-question is answered by a fixed rule applied to stored data; no AI is involved. Rules version three-digits-thesis-1.1.0.

Disclosure: these thresholds were written while building the pipeline, after some raw values had been seen. They are not a blind pre-registration. They live in src/lib/research/thesis.ts; change them and run npm run research:rebuild to see how sensitive the verdict is.

The buyback demand target model

The model asks how much measurable NEAR Intents activity would be needed for the revenue it generates to fund token buybacks equal to a chosen share of market capitalization each year. It is a scenario calculation, not an econometric forecast. Formula version nvabe-1.0.0.

annual_buyback_demand_usd = V × f × c × b
buyback_demand_to_mcap    = annual_buyback_demand_usd ÷ M
required_annual_volume    = (y × M) ÷ (f × c × b)
volume_multiple_required  = required_annual_volume ÷ V
required_take_rate        = (y × M) ÷ V      (reverse)

Scenario presets

Set before looking at results. Every adjustment is an assumption.

• Conservative

  Fees compress by half as competitors undercut, NEAR keeps 25% less of fees, half of captured revenue goes to buybacks, and volume does not grow.

  fee −50% · capture −25% · b 50% · growth +0% · y 2%

• Base

  Measured fee rate, capture and buyback share held constant, and the trailing 90-day volume run rate held flat.

  fee −0% · capture −0% · b measured · growth +0% · y 2%

• Optimistic

  Measured fee rate and capture held, all captured revenue goes to buybacks, and volume doubles.

  fee −0% · capture −0% · b 100% · growth +100% · y 2%

Calculator

Starts from the latest stored measurements (90-day window ending Nov 10, 2026). Change any input to test a scenario; edited inputs become ASSUMED.

Three roles, one shared truth set

1. 1The server fetches sources, validates and normalizes them, and freezes one snapshot per debate.
2. 2All ratios and the model are computed by deterministic code outside the language model.
3. 3Bull and Bear receive the same frozen evidence, the same model output and the opponent's prior turns. They have no browsing or tools.
4. 4Each turn is strict JSON (validated with a schema): spoken beats, at most four claims, each with evidence IDs and a fact, inference or assumption label.
5. 5The audit rejects unknown evidence IDs, flags numbers that match no snapshot value (3.5% tolerance), checks length and rebuttal references, and asks for one repair if needed. Remaining problems are stored and shown.
6. 6The Judge receives both transcripts, the evidence, the model output and the audit, and must challenge the model inputs.
7. 7Every turn stores its text hash, model, prompt version (three-digits-prompts-1.3.0), latency, audit and audio reference. Voices read the stored text exactly; they never generate content.

Provider text (labels, adapter methodology) is passed as inert data and sanitized. Users cannot supply URLs or free text to the server: debates accept only the one thesis topic.

Scoring

If penalized totals contradict the Judge's pick, the server overrides it: to the other side if it leads by 5 or more, otherwise to Inconclusive, and records why. The score is a debate argument score. It is not a NEAR investment score, a price forecast or a probability of profit.

Data sources

• CoinGecko

  market data API · public · minimum gap 6.5s between requests

  Documentation (opens in a new tab)

• DefiLlama

  DeFi analytics API · public · minimum gap 0.3s between requests

  Documentation (opens in a new tab)

• NearBlocks

  NEAR explorer API · public · minimum gap 6s between requests

  Documentation (opens in a new tab)

• NEAR JSON-RPC (FastNEAR)

  onchain RPC · public · minimum gap 0.2s between requests

  Documentation (opens in a new tab)

• FastNEAR transactions API

  NEAR transaction indexer · public · minimum gap 0.25s between requests

  Documentation (opens in a new tab)

• NEAR Intents Explorer API

  partner API (JWT) · credential required · minimum gap 5.5s between requests

  Documentation (opens in a new tab)

• Google BigQuery: crypto_near_mainnet_us

  public dataset (your GCP billing) · credential required · minimum gap 0s between requests

  Documentation (opens in a new tab)

Evidence methodology three-digits-evidence-1.1.0. The full request catalog lives in src/lib/data/sources.ts; nothing else in the code chooses a URL.

Reproduce this research

Every number on this site is computed by code from stored source responses. The published snapshot (THESIS.md and canonical.json) and the responses behind it are in the public repository, so anyone can rebuild it offline, with no API keys, no live APIs and no AI model. The live pages recompute from the newest responses stored for the site, so where those are newer, a number here can differ slightly from the published snapshot.

Repository

github.com/wongmayyeong/three-digits-minimum (opens in a new tab)

Published snapshot: observation cutoff

90-day window ends Nov 10, 2026

Published snapshot: responses retrieved

262 responses, Nov 8, 2026, 19:52 UTC to Nov 12, 2026, 15:56 UTC

This site's data

window ends Nov 10, 2026; newest stored response Nov 13, 2026, 04:13 UTC

Model versions

evidence three-digits-evidence-1.1.0 · ladder three-digits-2.0.1 · tokenomics w4-2.0.0 · thesis three-digits-thesis-1.1.0

This site was built from commit

fb1396b (opens in a new tab)

Debate prompts (not needed to rebuild)

three-digits-prompts-1.3.0

Download the data

• Raw responses (opens in a new tab): every public API response the snapshot was built from, one JSON file per request.
• Request manifest (opens in a new tab): for each response, the request URL, the source, the retrieval time and its SHA-256 hash.
• Evidence manifest (opens in a new tab): every evidence item with its value, period, transformation, inputs and the request it came from.
• canonical.json (opens in a new tab) (every published number) and THESIS.md (opens in a new tab) (the written thesis, generated from it).

Rebuild it offline

git clone https://github.com/wongmayyeong/three-digits-minimum.git
cd three-digits-minimum
npm ci
npm run research:rebuild
git status --short artifacts THESIS.md

It worked if the rebuild exits without “Offline rebuild differs from committed output”, and the last command prints nothing: every output was rebuilt byte for byte. Spot checks in the published snapshot: NEAR's own multiple is 212x, and 20% a year Intents growth at that multiple implies $16 in 2036 (canonical.json, ladder.multiples and ladder.implied).

What needs credentials, and what a hash proves

• Not part of the snapshot, so not needed to rebuild: integrator concentration (NEAR Intents Explorer, partner token) and retention cohorts (BigQuery, your own Google Cloud project). npm run research:refresh fetches fresh data and works without keys, at lower rate limits.
• A response hash proves the file you rebuild from is the one that was fetched. It does not prove the source data is right, or that our reading of it is.

Known limitations

• Accounts are not people. Active-account counts include bots, relayers and multiple accounts per person, and falling counts can mean less automation as well as fewer users.
• Volume is not organic demand. Without the partner-only Intents Explorer API, integrator concentration and wash-like activity cannot be measured.
• Third-party definitions. Fee, revenue and buyback figures follow DefiLlama adapter definitions, which can change. Definitions are stored in each snapshot.
• Valuation multiples move with sentiment. Today's multiples are applied to NEAR in ten years as scenarios. The same rung can look easy or extreme depending on the multiple, which is why every column is shown.
• One measured value channel. The ladder counts value reaching holders through Intents buybacks and gas burns. Other possible sources (staking demand, AI agents, NEAR as money) are not quantified because no reliable data supports a number yet; they are listed as unproven, not added.
• Exact data only. NearBlocks returns query-planner estimates instead of counts for busy contracts in older months; those months are flagged and excluded, so some history is shorter than the contract's life.
• BigQuery is optional. Retention cohorts need your own Google Cloud project (research/sql/), so they are not part of the live snapshot.
• Cross-chain comparability. Chain-fee multiples ignore value that fees do not capture, and TVL definitions vary by adapter.
• Subjective scores. Most rubric points are an AI judgement. They are shown with the deterministic audit so readers can disagree.