I check whether numbers deserve trust — evidence gates for AI agents that trade, judge-reliability harnesses, and independent re-checks of published results. Hong Kong.
Most claims I run into — a trading record, an LLM-judge score, a "verified" proof, a pricing engine's output — arrive as a single number with no way to tell skill from luck, or a real check from a silent no-op. My open-source work is the tooling I built to find out.
| evidence-deflator | Is that trading record skill, or could it be luck? Check any Hyperliquid wallet, rescore leaderboards for best-of-N selection, audit AI backtests for knowledge-cutoff leakage. Python stdlib, MCP server, agent skill. → evidencedeflator.com |
| assay | Check the judge before you trust the score — an evaluation harness for knowledge-answering support agents. |
| warrant | Suitability as an on-chain object, not a prompt: revocable warrants, preflight and attestations for AI-agent trading. |
| oai-math-recheck | Re-check one result family from openai/math yourself: compile, axiom audit, statement-equivalence, and a negative control. Worked example: family 325, the complete Crouzeix inequality. |
| ql-invariants | Model-free no-arbitrage invariant scans for QuantLib pricing engines. Found three engine issues (#2903, #2904 — fixed upstream, #2905). |
| erdos-699-i3 | Erdős Problem #699 at i = 3: a counterexample needs 4 |
| agent-tool-interop | Which coding-agent CLIs expose a tool schema that a third-party inference backend can actually drive — taxonomy, drivability criterion, compatibility matrix, reference adapter. |
| prophetmap | AI industry-chain transmission map for US equities — 27-layer propagation model with daily scoring. → live |
| gpu-fairprice | Machine-level fair value and price-vs-time-to-rent curves for heterogeneous GPU rental listings. |
| Web3Watch-HK | Tracker for Hong Kong's virtual-asset regulatory and market developments. |
A long-running side thread — I study linguistics, and these are its artifacts.
| ming-vintage-llm | LoRA fine-tune for Classical Chinese — finding: style transfers, world-model doesn't. |
| ming-vintage-scratch-30m | The control arm — a ~30M model trained from scratch on pre-1424 Classical Chinese. |
| cbdb-letter-recipient-linker | Link CBDB letter titles to recipient person IDs, with pre-registered evaluations. |
Smaller tools
- cli-proxy-doctor — why your terminal CLI is slow or broken behind a proxy chain, in four questions
- figqa — design-system QA for Figma
.figfiles with a CI exit code - groundhog — anti-repetition prompt composition for proactive agents
- token-daily — daily Claude Code token usage from local transcripts
- -fyp-research-4150 — CUHK FYP: Hong Kong's crypto trust paradox (N=287 survey + NLP social listening)
Product & research at HighBlock (BitV), a Hong Kong licensed virtual-asset and securities platform. Before that: brand & PR at Sinohope Tech (1611.HK), livestreaming ops at KuCoin, brand research at LBank, event ops at BingX.
Co-author, AI 智能投资手册 — the chapter on decision review and feedback loops.
CUHK — Journalism & Communication, 2026. Exchange at IESEG (Paris & Lille).
- Evidence Deflator → evidencedeflator.com
- Site → beltran12138.github.io
- X → @zkBernard12138
- LinkedIn → bernard-zhao
- Hugging Face → Beltran12138
- Xiaohongshu → 伯纳德zkBernard (Chinese long-form on LLMs and language)
- Email → bernard12138@gmail.com