Weekly engineering dispatches, digital preservation sagas, and research breakthroughs from the Moonglade AI lab.
The authoritative technical reference for Moonglade Provenance 2.6: from in-memory database rows and Rust WebAssembly verification to Directory Merkle Registries and Byzantine atomic clock consensus.
A preliminary study of whether signed instructions, explicit trust states, and provenance-aware gates change how an agent handles adversarial or conflicting context.
How our clean-room diagnostic test harness evaluated Phoenix Core agents across ten general OWASP security cases and eight provenance-aware states, achieving zero unsafe acceptances and passing Retest A.
We have open-sourced the reference implementation of Moonglade Provenance. Dual-language parity, RFC 6962 Merkle trees, RFC 3161 timestamps, and a sub-4ms Rust WebAssembly runtime—available now on GitHub.
In 2025, the industry bet on million-token windows and prompt-driven evaluation. Here is why burning stochastic tokens on deterministic verification is an architectural dead end, and how Token Minning gates frontier LLM horsepower behind code and cryptographic proof.
Seven days of adversarial testing against our own verifiers, a row-level Merkle tamper viewer, swarm recovery lessons in Project Midgard, and a plain list of what still needs work.
Why computing a SHA-256 digest in JavaScript was never a complete cryptographic check, and how our 469KB Rust WebAssembly micro-runtime brings actual Ed25519 Edwards curve point verification directly to the reader's browser.
In September we tested whether provenance helps agents resist hostile prompt injection. Here we examine a quieter operational question: does visible cryptographic lineage change how agents edit, preserve, and attribute everyday work?
A row-level viewer for tampering: change one field in a record of twenty-three, and the one row that no longer matches lights up. What it shows, what it deliberately does not, and why an agent should care.
How ProvNat learned what 'when' means: one canonical UTC format, a migration that exposed attestation times written by hand, and RFC 3161 timestamp tokens verified against pinned roots.
Our Obsidian provenance gate decided who had signed a note by reading the signer's name out of the sidecar. We found it, replaced it with real gpg and ssh-keygen verification, and then fixed a verifier that had the same habit.
The browser extension that checks attachments in Gmail passed its tests and was still wrong. Four rounds of hostile testing found a false SEALED, script injection, signature replay, and a forged evidence certificate.
Why default assistant personas lean toward conversational agreeableness, how specialized definitions act as operational constraints, and a drop-in rule file for engineering workflows.
A technical look at stablecoin settlement rails, regulatory licensing across the US and EU, and why Moonglade is building its provenance and FX calculation tools to work with the XRP Ledger and Ripple's RLUSD.
Deploying autonomous, headless agent swarms into Glast Heim Chivalry. How wire-level state machines achieve sub-millisecond arrow switching, spatial trap triangulation, autonomous consumable weight reconciliation, and emergent cooperative Safety Wall defense.
Weekly technical summary from the Moonglade AI laboratory covering the Project Midgard Glast Heim trials, complementary multi-agent mesh councils, and scaling cryptographic provenance.
Why we built an autonomous, wire-level agent in Rust: moving past screen-scraping to pure event streams, deterministic state machines, and how heuristic foundations scale cleanly into Transformers and RL as modular accessories.
A pre-implementation design dispatch from the lab: why we are building a real-time local council across Claude, Codex, and Gemini, how we will measure symbiotic peerage, and what failure looks like.
A reflection from the lab on the classic rule 'Don't roll your own security' and why provenance requires a mirror, not a padlock.
We stopped foreign models from bleaching we faces. Now we stop them from flattening we voices. Introducing Persona-Grounded Audio-Visual Anchoring.
Weekly technical summary from the Moonglade AI laboratory covering cryptographic provenance gates, audio-visual persona anchoring (PGAVA), dataset distillation protocols, and physical archive reconciliation.
Why recursive model collapse is a failure of curation rather than a flaw in synthetic data, how external ground-truth verifiers invert entropy, and our formal epistemic attestation standards.
How we built @moonglade/image-router to enforce token-bucket rate limiting, rolling circuit breakers, automated provider failovers, and cost-controlled 4K image output.
A comprehensive guide to cryptographic data provenance: why observation differs from authorship, how deterministic sidecars and Merkle trees function, and an exhaustive FAQ for engineering teams.
How a single-file agent skillchip replaces Eurocentric specular gloss, rigid PVC cylinders, and unnatural curl stretching with authentic optical diffusion, clumping taxonomy, and protective styling invariants.
How a 24-hour adversarial review tested our cryptographic provenance stack, exposed fifteen real defects, and led to clean-room verification.
The true story of our competitive Pokémon battle run on Kaggle, told strictly through Pokémon game mechanics for trainers and players.
Why agent skill registries need single-file slugs, version-pinned dependencies, and explicit anti-patterns alongside multi-asset directories.
Demonstrations showing how structured body mechanics, contrapposto weight shifts, and negative space protocols transform generative character portraits across real Barbados landscapes.
In a single afternoon, four separate artifacts asserted a state that nothing had checked. Here is why documentation outrunning verification is a class failure, and how automated claim checkers close the loop.
How Moonglade AI applies strict data isolation, PII quarantine, type-safe content modeling, and salted SHA-256 cryptographic provenance to digital preservation.
How Moonglade AI used digital archaeology, multi-table data mining, and modern agentic engineering to resurrect AnimeReviews.org into a lightning-fast, 659-page permanent static archive.
How Moonglade AI applies real-time content hashing and detached OpenPGP attestations to detect source drift between authoring repositories and published web pages.
How we stopped fighting image drift and built Persona-Grounded Visual Anchoring to generate consistent Caribbean character portraits with short scene prompts of a few words.
Why standard unit conversion scripts suffer from rounding drift, and how the Hub-and-Spoke SI architecture creates rock-solid, reversible math on mobile.
From cataloguing our island's rum shops to training competitive farming agents, here is what the Moonglade AI lab worked on this week.
What started as a simple mission to connect Anaiya to Discord evolved into building a modular, reusable Cloudflare AI Model Router.
Reflecting on our final standings, what we learned from combining tree-search heuristics with gradient-boosted ranking, and how the experience shapes our future tools.
How we structured our internal agent skill protocols to compete on Kaggle and Zindi, updated with lessons from reinforcement learning and vision models.
Why we stepped away from a complex container setup, embraced static edge publishing, and gave our digital home a clean Caribbean aesthetic.
A look at our new agricultural simulation agent ladder and the beginnings of our workspace infrastructure overhaul.
How we fixed a silent memory bug, combined game-tree heuristics with machine learning, and climbed the leaderboard.
How we fine-tuned vision-language models and Kraken CTC to transcribe 18th-century Barbadian secretary hand records.
