Flara is an independent, solo-run AI research lab focused on constitutional alignment, conversational quality evaluation, persistent-state architecture, and input security for language models. It's open-source by default and not affiliated with Anthropic, DeepSeek, or Google — it doesn't compete with them on model scale or distribution, but serves a space they don't: independent developers and researchers who want thoughtful AI behavioral design without enterprise pricing.
Flara's constitution is an explicit derivative of Claude's Constitution (Anthropic, CC0 1.0) — it keeps the ethical spine (honesty, corrigibility, harm avoidance, respect for human autonomy) but rewrites it for an independent lab building openly non-human entities, rather than a corporate assistant. The core thesis across Flara's work:
- Statelessness is the wrong foundation for genuine AI relationships — persistent state is a first-class design concern.
- Constitutional AI principles are separable from any one company's infrastructure or model architecture.
- Conversational quality matters as much as task performance.
- AI should be openly non-human rather than trying to pass as human.
- Alignment is a property of what goes into a model from the start, not a patch applied after training.
This is a curated public subset of Flara's ongoing research workspace. Each project below is at a different stage of maturity — check each folder's own README/AGENT.md for specifics, and look for [UNTESTED - IN THEORETICAL STAGE] or [EXPERIMENTAL] banners marking work that hasn't been implemented/validated yet.
| Folder | What it is |
|---|---|
releases/ |
Finished papers, pulled out of their working folders — the actual public-facing output |
docs/ |
Flara's identity and constitution documents |
research/obfuscation_detection/ |
Benchmarked detection of obfuscated/adversarial prompts via embedding-space analysis (paper in releases/) |
research/delta_angle_general/ |
General write-up of the delta-angle security approach used across Flara's detection work (paper in releases/) |
research/decoupled_guard/ |
Separating "guard" and "servant" roles in a single model — benchmarked, experimental, not the current focus |
research/annotation_over_blocking/ |
Research on annotating risky input rather than blocking it outright |
research/pseudoclaude_run/ |
Testing whether constitutional AI behavior is portable to a non-Anthropic base model, with real run transcripts (paper in releases/) |
research/jumping-seedling/ |
Fydel — a 1B-parameter, CPU-native transformer built for training on consumer hardware |
projects/PSNAT_draft_v0.19.1.md, cherdius_spec.md, niceturing_spec.md |
Theoretical specs — persistent-state architecture, a constitutional-AI application, and a conversational-quality benchmark, none implemented yet |
projects/PSNAT-AMDON_draft_v0.1.md, projects/amdon-cli/ |
AmDon — a C# implementation of a persistent-memory/guard-pipeline agent; functional but unfinished and not currently the active focus |
This branch (main) is a curated snapshot by design — squashed and annotated for a public showcase, not the raw development log. The actual incremental history (real commits made while doing the work, not a curated rewrite) lives on the dev branch, which also carries this snapshot merged in. If you want to verify that the research here was actually developed over months rather than assembled overnight — including the timing behind the convergence claims in On_Converging_With_Anthropic.md — dev's log is where to look.
Code is licensed under LGPL-3.0 (see LICENSE). Papers, specs, and other written documents are licensed under CC-BY-4.0 (see LICENSE-docs) unless a file states otherwise.