APP CONSTRUCTION
Effective October 18, 2026
How Hedera Watch is built, what its automated collectors measure, and where AI models and background agents assist research, review and commentary.
1. How Hedera Watch is built
Hedera Watch is an independently operated, AI-assisted application built and iteratively edited using Base44. Its interface uses React and JavaScript, Tailwind styling, Recharts visualizations and Three.js globe rendering. Base44 supplies application hosting, authentication, stored data, server functions and scheduled workflows. Public Hedera mirror nodes, exchange market feeds, publisher sources and DefiLlama supply the relevant external inputs.
AI assistance in development is not an independent code audit, security certification, statistical validation or legal review. Software, prompts and configurations can contain errors. The operator remains responsible for maintaining the application and reviewing its outputs; this does not mean every item receives human review before publication.
2. Continuous capture is automation, not an all-seeing AI
Network snapshots, revenue, node transaction activity and fixed-cohort and expanding-market RWA activity have configured five-minute collection schedules. Historical-summary maintenance also runs on a five-minute cadence. Collectors fetch public endpoints, paginate records and calculate aggregates using programmed rules. Live browser prices use exchange streams and REST fallbacks; live TPS uses mirror-node readings. These measurements are not numbers invented by a language model.
Schedules are targets, not guarantees of uninterrupted operation or complete monitoring. Requests, sources, capacity limits, pauses and collection backlogs can delay or prevent updates. A scheduled worker runs bounded tasks, not a permanently active observer. Missing or pending history must not be read as zero activity. This page describes the implementation and configured schedules as reviewed October 18, 2026, not a live execution-status dashboard.
3. AI model configuration and disclosure limits
News discovery, review-preview recovery, daily RWA web research and Daily Vibe use Base44’s built-in language-model integration with automatic model selection rather than a pinned model version. Newsie and the Newsstand editorial agent also request the AI Gateway model identifier automatic. The application does not publish the resolved provider/model identity for each run; no particular underlying model or stable model version can therefore be asserted here.
News text plausibility/safety screening and Newsie’s community-fit evaluation explicitly use the TypeSafe AI evaluation model identifier jev through the Base44 AI Gateway. Jev is used for structured rubric-based judgments, not for measuring blockchain activity. Automatic routing, model availability and behavior can change, so an identifier describes configuration rather than a guarantee of an immutable implementation.
4. Newsie: background news-review agent
The configured 15-minute news workflow combines publisher feeds and AI-assisted web search. After source checks, Newsie handles uncertain article, opinion and YouTube candidates using fetched evidence and up to the latest 20 saved manual decisions per category. It can inspect candidates, read editorial feedback, assess community fit and submit routing decisions. Each batch is limited to six candidates and eight agent steps.
Clear passes can publish automatically. A Newsie-assisted release requires the server-enforced evaluation threshold and relevant saved feedback; unresolved cases remain for manual review and explicit unsafe outcomes remain held. Feedback supplies examples at review time, not model retraining. This is text and metadata screening, not independent fact-checking, a review of video footage/audio or a guarantee that a linked destination is safe. Official Hedera releases have a separate five-minute feed check and can bypass editorial classification while retaining source checks; authorized manual approval is also a distinct route.
5. Newsstand: editorial summary agent
After news refresh, the Newsstand editorial agent reviews up to 20 already collected, eligible stories. Its tools are restricted to inspecting supplied story evidence and returning short summaries. It must inspect each edited story first and is bounded to six turns. The server applies accepted summary edits and review timestamps.
This agent cannot discover arbitrary new stories through its supplied tools or independently certify reported facts. Summaries can still omit context, misattribute statements or contain errors. Opinions and predictions belong to their sources, not the site. Administrator-triggered runs may supplement scheduled runs.
6. Other AI-assisted tasks and rule-based monitoring
RWA discovery is configured daily at 06:15 America/New_York. Its bounded web-research call returns at most 10 sourced leads, combined with issuer announcements and on-chain discovery. This is a single research call, not a separate open-ended tool-using agent. Research alone does not authorize admission: exact primary evidence and on-chain checks, or documented administrator review, are required. Neither route verifies legal backing or reserves.
Daily Vibe is configured at 23:45 America/New_York and uses a single AI call over saved headline sentiment and stored price changes to draft a mood score, headline, summary and drivers. It is commentary, not a market measurement or investment signal. The daily numerical recap is scheduled at 23:55 America/New_York. Breakout checks, revenue normalization, historical aggregation, distribution fits, regressions, correlations, volatility and forecast cones use programmed statistical rules rather than language-model predictions. Contact notifications are event-triggered email delivery, not an AI conversational agent.
7. Inputs, boundaries and human oversight
The disclosed AI workflows process public source text and metadata, public market context and, for news review, saved administrator editorial notes and examples. In the current implementation these prompts do not include contact-form submissions, custom visitor-session records or browser-local holdings-calculator amounts. Do not place private credentials or sensitive personal information in editorial notes or submissions.
The disclosed editorial agents have bounded tools for evidence review, routing and summaries, not tools to trade, connect wallets or take custody of assets. They are background editorial processes, not visitor-facing personal advisers. Automated publication does not imply prior human approval, comprehensive coverage or independent verification. Model confidence scores are routing aids, not measured probabilities that a story is true.
8. Limitations, changes and responsibility
AI can hallucinate, misclassify, miss relevant material or repeat misleading source claims. Deterministic collectors and statistical models can also fail or be unsuitable for a changing market. Verify consequential information against original sources; no output is financial, legal or tax advice. The Methodology, Privacy Policy and Terms explain additional data, privacy and reliance limitations.
This disclosure is a maintained description, not an exhaustive model-execution log or a claim that every configured workflow is currently succeeding. Construction, model routing, prompts, agents, sources and schedules may change. Questions or correction requests can be sent through Contact.
Questions? Contact the site operator. See also our Privacy Policy, Terms of Service and Methodology & Data Sources and App Construction & AI Disclosure.