Decentralised News · AI & Crypto Research Infrastructure · Source review: 19 November 2026
By Heath Muchena
Direct answer. A trustworthy crypto AI search engine should retrieve dated evidence, preserve the source’s actual definitions and show which parts of an answer are observed, calculated or uncertain. A language model can summarise those records, but it cannot create source coverage, fill missing reporting periods or make an old quote current. The essential engineering task is constructing an auditable evidence layer before generating prose.
What Matters
- Combine primary documents, structured blockchain records and market-data feeds without pretending they describe one identical thing.
- Every answer needs an entity, a metric definition, the source’s observation period, the time it was retrieved and the underlying citation.
- When two sources conflict, show the contradiction rather than averaging their answers.
- API authentication, fair access, licensing and historical coverage are parts of the product design, not afterthoughts.
- This article proposes a reference architecture. DN did not build or deploy a new search product for this study.
The Signal: a fresh answer can contain stale evidence
A model might answer in seconds while citing a financial statement for last quarter, a live-looking token price from several minutes ago, and a protocol TVL figure based on a different chain and denomination. The speed of the generated text does not establish freshness of its underlying facts.
The SEC’s EDGAR APIs provide structured filing histories and XBRL financial facts. Dune exposes query-based blockchain analytics, DefiLlama supplies protocol-level metrics, and CoinGecko offers current and historical market data. These are distinct data families with separate timing, entity and access rules. Connecting them requires a schema and decision about what an answer is actually supposed to compare. [1] [3] [4] [5]
DN Alpha Thesis
The strongest advantage in agentic financial research may lie in the evidence object: a dated, source-identifiable claim with its units, scope, uncertainty and permissible reuse rights. Once that object exists, a search interface can answer many questions without repeatedly relying on unsourced summaries.
The defensible moat is not a larger number of model-generated pages. It is the ability to show the lineage of every material number, identify what changes over time and explain exactly when a conclusion no longer follows from the evidence.
1. Start with a question contract
“What is happening in crypto?” is too broad to test. Start with a narrower, falsifiable user request such as “Which protocol’s reported TVL grew fastest in the last 90 days?” or “Does a listed company’s most recent filing show higher cash generation?” A question contract defines the target entities, metrics, horizon, comparison groups, permitted data families and output the system must return if evidence is missing.
Use an explicit null or unavailable state. A query that requires a current on-chain observation should not silently fall back to a blog article from a previous year. A query about financial statements must not replace published cash flow with a price chart or model-estimated revenue unless the answer makes that substitution visible.
2. Give every evidence object seven essential fields
For each material observation, retain the canonical source URL, source identifier or content hash, entity or instrument ID, metric and unit, observation period, retrieval timestamp, and the actual value or quoted rule. Add caveats and licensing conditions as first-class metadata when necessary.
Entity resolution is a separate task. A token ticker can be reused by unrelated contracts; company names can change; a protocol may operate on several chains; an exchange may quote different pairs with the same familiar asset symbol. Preserve addresses, chain IDs, SEC CIKs or other canonical identifiers instead of forcing all entities into ticker strings.
Retain observed values separately from derived figures. If a calculator adds period cash flows or converts token balances into dollars, show the formula, conversion price and its observation timestamp. A reader should be able to reproduce the arithmetic without assuming that the model invented nothing.
3. Treat financial filings as dated claims, not live market data
The SEC describes public APIs for company submissions and financial facts, including XBRL structures. Reported facts have filing dates and reporting periods that can differ substantially. The SEC also describes API update timing as typical rather than guaranteed. [1]
A search engine must distinguish a quarter’s closing asset balance from the date the filing was accepted and from when a user retrieved it. An amended report should not simply overwrite a historical research decision. Store or link the earlier observation when permitted, and mark later corrections as separate evidence versions.
Automated access must respect the SEC’s published fair-access policy and declared user-agent requirements. A system calling thousands of pages in parallel is not justified merely because the information is public. [2]
4. On-chain query output needs its own lineage
Dune queries and protocol dashboards can be immensely useful, but two analysts may produce different answers by selecting different token lists, chains, smart contracts, timezones or deduplication policies. For a reusable answer, retain the query identifier or logic version, parameters, data freshness and output coverage. A screenshot of the latest value alone is not a reproducible research method.
DefiLlama’s API documents protocol TVL and historical-chain endpoints. A user asking about deposits, total assets and TVL may be making three different requests depending on which protocol and measurement conventions apply. A search product should expose the provider’s metric definition and any unresolved mapping rather than convert one label into another without evidence. [4]
Block explorers can corroborate contract events, but event counts do not automatically establish beneficial owner counts, business revenue or profitable users. A chain offers strong evidence of observable state changes; motives and off-chain obligations require different records.
5. Price data has timestamps, methodology and entitlements
CoinGecko documents live and historical market-data products, but their refresh cadence, granularity and rate limits depend on subscription and endpoint. Its November 2026 support article distinguishes the unauthenticated public allowance from a registered Demo plan. Do not treat a marketing-page capacity figure as a universal permission for an account without checking the relevant service. [5] [6]
A quoted token price may aggregate several venues or represent a selected market. A trading agent needs to know whether it is comparing a statistical market reference with a realistically executable order. For research, the source and time of the mark are often sufficient; for execution, available liquidity and fees also matter.
Do not turn a stale-data check into a trading recommendation. A cache that is adequate for a broad research summary may be unfit for order submission, and the user should define the tolerance before evaluating the result.
The Evidence Families Compared
| Source family | Useful evidence | Minimum context | Interpretation limit |
|---|---|---|---|
| SEC EDGAR | Company filings and XBRL facts | CIK, accession, reporting period and amendment state | Not a current share quote |
| Dune | SQL-query-based blockchain analytics | Query ID, parameters, chain, time and result version | Not an independent user-identity census |
| DefiLlama | Protocol and chain metrics | Metric definition, entity/chain and collection date | TVL is not necessarily protocol profit |
| CoinGecko | Market and historical prices | Asset/contract ID, venue aggregation, timestamp and plan | Quoted price is not a guaranteed fill |
| DN analysis | Synthesis and derived scenarios | Source record IDs, formula, assumptions and reviewer | A model’s narrative is not a new independent source |
6. The answer engine should refuse unsupported joins
Suppose a user asks whether an AI-agent token’s reported transaction growth increased the owning company’s recognised revenue. The blockchain may show additional transfers, while a company filing may contain consolidated revenue from many products and a project dashboard may define “revenue” as collected protocol fees. Those sources are not necessarily joinable into the causal conclusion implied by the question.
A defensible engine can respond with the three observations, their periods and the missing connection: ownership or fee accrual, the relevant economic entity and the accounting treatment. That can be more useful than a confident invented answer. The absence of a comparable metric is a research result, not a reason to substitute a generic AI prediction.
7. Store a source receipt with every answer
The system should be able to export a compact receipt containing the user question, resolved entities, retrieval run, source URLs, source access times, metric definitions, transformations, generated answer, and unresolved issues. The receipt itself should have a stable ID and a modification history.
For a publication, the display can offer expandable sources and an evidence download. For an analyst, it may also expose machine-readable JSON with the exact observations. Do not assume that a data licence permits unrestricted redistribution, commercial storage or downstream AI training. Each provider’s terms must be reviewed for the intended workflow before such a feed is deployed.
This architecture allows a correction to propagate to related briefs and calculators, while keeping earlier decisions inspectable. A silent replacement destroys the audit trail.
DN Evidence Freshness Planning Tool
Illustrative source-age budget, not a live data feed or verification of a source. Enter all times in minutes.
Age at answer = age at retrieval + processing time. Overage = maximum of zero and age minus maximum. Headroom = maximum − age. Production scheduling requires its own policy. No source is fetched or judged automatically.
8. Three answers the design should distinguish
Supported and sufficiently fresh
Every material value has a resolved entity, compatible unit, accessible source and age within the declared threshold. The system may generate a direct answer while retaining all evidence links. That is readiness to answer the specified question, not proof of prediction accuracy.
Supported but stale
The observation appears credible but exceeds the question’s freshness tolerance. The system can either refresh it or present the last-known value with a clear timestamp. It should not call it current just because the text was generated now.
Unresolved source conflict
Two authoritative-looking sources disagree because of definitions, reporting dates or other unknowns. Show each observation and its scope. Do not assign an invented midpoint to make the narrative seem more certain.
The Countercase: not every question needs a heavyweight index
A reader checking a single SEC filing can often answer their question through the official filing and a straightforward formula. An elaborate agent pipeline may add cost and failure points without improving the decision. The architecture earns its complexity when questions recur across entities and periods, when corrections matter, or when many answers need to reuse the same verified observations.
Start with a narrow primary-source panel and a small library of explicit answer types. Measure correctness, citation coverage, retrieval cost, failed joins and time to correction before expanding. This article proposes a design, not a deployed benchmark or a claim that one model is universally best.
Methodology, Sources and Limits
This is a documentary architecture study using the official API materials linked below. DN did not issue API requests, configure keys, deploy the index, benchmark search quality or test a model for this article. Source descriptions are provider statements and should be rechecked for changed plans, entitlements or interface behaviour. The numeric tool is original scenario arithmetic, not observed latency.
Usage rights, data quality, operational security and legal requirements vary by provider and country. The design separates historical data from current observations, marked prices from actual executions and regulatory filings from commercial platform claims.
- SEC EDGAR Data APIs — SEC, page reviewed 8 April 2025; submissions, company facts and reporting periods require interpretation; no authentication for public data APIs.
- SEC Webmaster FAQ — SEC, public scripting guidance with declared user agent and 10 requests/second current maximum; availability times are not guarantees.
- Dune Data API — Dune official API description, accessed 19 November 2026; blockchain query data, freshness and plan-dependent limits.
- DefiLlama API Docs — Provider API reference for protocol TVL and other metrics, accessed 19 November 2026; fields and definitions require review.
- CoinGecko API Docs and Plans — CoinGecko official price and historical data descriptions; rates and data entitlements vary by plan.
- CoinGecko Public API Rate Limit — Provider help page updated 3 November 2026; public unauthenticated limits vary, and registered Demo has a different published allowance.
Frequently Asked Questions
Can I build a crypto AI search engine without paying for data?
A narrow prototype can use some public interfaces subject to access policies and licensing. That does not guarantee sufficient historical coverage, production capacity or redistribution rights.
Should the AI trust all crypto API data equally?
No. Each source has a measurement scope, update cadence, entity definition and limitations. Reconcile those before combining them.
Does adding citations guarantee the answer is correct?
No. A citation can point to a real document while a model misreads its period or definition. The derived claim must be checked against the actual record.
Does this article provide a working search engine?
No. It provides a reference architecture, source map and an illustrative freshness calculator. No production data index is claimed.
Continue the Research
See DN’s exchange API research and measurement reconciliation guide for related tooling principles.
Educational research; not investment advice or a live data service.




