Lossless/ai-labs
Search/

ch · studies

Pinned reading,
not paraphrase.

Each study scopes one domain question and pins upstream code as nested git submodules. We read from the pinned source — not from training data, not from a paraphrase, not from a blog post that aged out a year ago.

  • 9 studies
  • 91 pinned repos
  • read from source

study · 01

Open Specs & Standards

A reading collection of open specifications, standards, and conventions that shape how humans and AI agents cooperate through files. Each entry is cloned as a git submodule under this directory.
15 pinnedstudy repo →#open-specs-and-standards

study · 02

Memory Layers for Agents

How do production memory systems for AI agents structure recall — across vector, graph, and key-value stores; across user, session, and agent scopes; across short-term context and long-term knowledge — and which conventions are converging?
18 pinnedstudy repo →#memory-layers-for-agents

study · 03

Data-Analytics Specs & Standards

What conventions exist for open, machine-validatable specifications and standards across the data-analytics stack — for describing datasets, fields, dialects, columnar storage formats, in-memory data formats, and grammars of graphics — and what can we learn from the mature exemplars before designing our own?
6 pinnedstudy repo →#data-analytics-specifications-and-standards

study · 04

Markdown: Parse, Serve, Render

How do mature projects parse, serve, and render Markdown — across CommonMark, GitHub-Flavored, Pandoc-extended (Quarto / R Markdown), and directive / wikilink-flavored dialects — and which conventions are converging at each of the three stages?
6 pinnedstudy repo →#markdown-parse-serve-render

study · 05

Vector Databases

How do open-source vector / context databases structure embedding storage, recall ranking, payload schemas, filter expressivity, persistence, and ANN-algorithm choice (HNSW / IVF / DiskANN / ScaNN / etc.) — and what are the trade-offs when migrating between them?
6 pinnedstudy repo →#vector-databases

study · 06

Agent Harnesses

How do coding-agent / conversational-agent runtimes organize and write to files on disk, persist memory across sessions, log traces of tool calls and reasoning, decide which tools/MCP servers are in scope at a given moment, and load or evolve skills?
10 pinnedstudy repo →#agent-harnesses

study · 07

Conversational UI & Native Shells

How do native, cross-platform conversational UI applications — especially ones built with Tauri (Rust backend + web frontend) — structure frontend state management, multi-conversation/multi-context switching (e.g. swapping between different "workspaces" or agent personas), local file persistence (chat history, settings), and integration with backend agent processes or MCP servers?
11 pinnedstudy repo →#conversational-ui-and-native-shells

study · 08

Sync & Content Version Control

For a folder of documents shared bidirectionally between an operator and a non-technical client, backed by blob storage, at which layer does history live — the transport, the store, the application, or the document?
9 pinnedstudy repo →#sync-and-content-version-control

study · 09

Frontend UI Kits & Component Libraries

How do mature UI kits name, scale, and override their component APIs — and which of those conventions survive translation to Svelte 5 + CSS custom properties, with no Tailwind and no React?
10 pinnedstudy repo →#frontend-ui-kits-component-libraries

collaborations

Collaborations

Silo against silo: when a friend has built their own agentic-coding harness next to ours, what can each side learn from the other?

ch · why pin

Why pin upstream code instead of bookmarking it?

grounded

The agent reads source, not training-data folklore.

When a decision touches a convention someone else already worked out, the answer comes from the pinned repo — not from whatever the model half-remembers.

stable

SHA-pinned. Doesn't rot under us.

A study's submodule points at a specific commit. When we revisit a study six months later, the same source is still there. Upgrades are explicit, not silent.

scoped

One study, one question.

Studies are not a general reading list. Each one scopes a single domain question and pins only what helps answer it. New questions get new studies, not pile-ons.