Blog
Field notes on OSINT, security checks, SEO, and safer hiring and vendor due diligence.
-
Judge Lab and the risk of self-measuring evals
2026-09-05
Judge Lab shows a common LLM evaluation failure mode, where the benchmark reflects parser rules, retries, and rubric wording more than task quality. The useful lesson is how to inspect the harness, verify its sensitivity, and read the right signals before using its scores.
-
Bench Lab and the case for benchmarks with visible errors
2026-09-04
Bench Lab shows a better way to benchmark systems: expose the wrong cases, not only the score. For practitioners, that shifts evaluation from headline metrics to failure analysis you can inspect and act on.
-
MMR Log and the engineering of append-only commitments
2026-09-03
MMR Log demonstrates how a Merkle Mountain Range commits to an append-only history without letting commitment size track raw history growth. The useful signals are deterministic peak formation, position-bound proofs, and clear prefix semantics.
-
Deposit Dispute and the case for deterministic arithmetic
2026-09-02
Deposit Dispute shows a useful engineering pattern for money-sensitive workflows: let models read messy evidence, then let contract terms drive deterministic arithmetic. The value is auditability, line-by-line verification, and cleaner handling of conflicts and missing facts.
-
Oracle Health Explorer and the engineering of feed health
2026-09-01
Oracle Health Explorer shows why an oracle feed should be inspected as a live system, not a single number. It highlights the engineering signals behind Chainlink feed health, including freshness, round progression, precision, and where integrations often fail.
-
AI Parametric Insurance and checkable claim adjudication
2026-08-31
AI Parametric Insurance shows a narrow, useful role for language models in claims work: mapping evidence to policy triggers in a way you can inspect. The core lesson is checkability, structured outputs, source anchoring, and clear failure handling.
-
Reading Live Lending Curves With Interest Rate Model Explorer
2026-08-30
Interest Rate Model Explorer shows why live, immutable lending curves matter for technical due diligence. The value is in reading the deployed model directly, then verifying its shape, parameters, and stress behavior from first principles.
-
Gas Optimization Race and how to verify gas claims
2026-08-29
Gas Optimization Race shows the same contract idea implemented twice and raced live on-chain. The useful lesson is how to inspect gas claims with public execution evidence, semantic equivalence, and controlled comparisons.
-
ZK KYC Pass Explorer and the engineering of private proof
2026-08-28
ZK KYC Pass Explorer demonstrates a privacy-preserving identity pattern: verify possession of a credential without exposing the underlying identity. The engineering value sits in claim minimization, verifier design, replay resistance, and the places where logs and policy often put identity back into the flow.
-
RWA Tokenization Explorer and the mechanics of gated transfers
2026-08-27
RWA Tokenization Explorer shows the engineering pattern behind ERC-3643-style KYC-gated transfers. The useful lesson is where transfer eligibility lives, how to verify it from contract behavior, and where restricted-token systems often fail.