\|/ias/|\
v1.1 · 2026-04-15 · Apache-2.0 — schemas + tools + individual-protection guide addedThe spiral of learning-to-doing \|/ doing-to-learning.
Eight governance protocols, JSON schemas, generic Python tools, and an individual-protection guide — for anyone processing data with AI agents.
\|/convergence — many learnings narrow to an act of doing (L2D)
iasthe framework — observability at the synthesis point
/|\expansion — one act of doing branches into many learnings (D2L)
IAS is a framework for making human-AI coding sessions observable, measurable, and improvable. It provides reusable protocols and templates that govern Claude Code (and compatible agent) sessions — context management, cost tracking, dependency assimilation, continuous documentation, verifiable state transitions.
Sessions without structure burn through context windows silently, accumulate costs that surface only at billing time, and lose learnings between runs. IAS solves each of these with a named, versioned protocol and a ready-to-adopt template. A two-minute wizard generates a project-specific CLAUDE.md with the appropriate gates turned on, the rest turned off.
Beyond operations, IAS is also a research platform. The state transitions, cost trajectories, and context-pressure curves IAS records across sessions are structured research data — version-controlled artefacts for the mathematical analysis of interacting agent systems with a human in the loop.
v1.1 adds portable regulatory tooling: JSON Schemas for dependency jurisdiction (US CLOUD Act, UK IPA, GDPR Art. 48) and compliance records (GDPR Art. 33, NIS2 Art. 23), a progressive-warning deadline tracker, a dependency-version checker (Gate 6.1), and a confidentiality scanner (Gate 7.5.1). Plus an Individual Protection guide for solo developers and small teams. Apache-2.0, Matrix-free — a JSON file, Python, and a cron job are enough for the regulatory story.
Seven categories of artefact in this release — all Apache-2.0 on Codeberg.
| Category | Ships in ias/ |
|---|---|
| Protocols | 8 standalone specifications under protocol/ |
| CLAUDE.md generator | tools/generate-claudemd.py + templates/CLAUDE.template.md.j2 + 3 worked example configs |
| JSON Schemas | schemas/dependency-jurisdiction-v1.0.json, schemas/compliance-record-v1.0.json |
| Regulatory tools | tools/jurisdiction-audit/, tools/compliance-deadline-tracker/, tools/check-versions.sh, tools/check-confidentiality.sh |
| Templates | CLAUDE.md template, entities.template.json, STATEOFCLAUDE.template.md, supply-chain.template.json, session-costs.template.json, dependency-jurisdiction.example.json, confidentiality-denylist.example.json, incident-response-playbook.md |
| Hooks | hooks/enforce-pretooluse-bash.sh, hooks/enforce-pretooluse-write.sh, rules registry, pipe-tests |
| Guides | docs/individual-protection.md — GDPR Art. 48, practical self-defence for individuals |
Companion protocols — jurisdiction and compliance — ship on foundation-protocols-spec rather than in IAS, because they define event schemas / data formats (the protocol family) while IAS ships the methodology, tools, and guides that use them.
Each protocol has a standalone specification under protocol/ in the Codeberg repo. All eight govern how a human-AI coding session is run.
| Protocol | What it governs | Spec |
|---|---|---|
| Gates | 10 mandatory checkpoints from session start through post-action | gates-v1.0.md |
| DICSTAMACH | Dynamic Interactive Claude State Machine — 4-state context management, consent-gated GC | dicstamach-v1.0.md |
| CONDOC | Continuous documentation — document AS you do, not after | condoc-v1.0.md |
| Cost Tracking | Session cost accounting with inter-commit periods and per-agent attribution | cost-tracking-v1.0.md |
| L2D&D\|/D2L&D | Learning ↔ Doing spiral — never code against unread source | l2d-d2l-v1.0.md |
| Supply Chain | Dependency assimilation tracking — fork, read, learn, record | supply-chain-v1.0.md |
| Spec Quality | Bootstrap test — can a spec regenerate its own template? | spec-quality-v1.0.md |
| Instruction Layers | L1 / L2 / L3 CLAUDE.md layering convention | instruction-layers-v1.0.md |
Everything you need to adopt the protocols in your own project — Apache-2.0, from the Codeberg repository.
Two-minute wizard emits a project-specific CLAUDE.md with the right gates enabled. Three worked examples: python-cli, kotlin-jvm, go-service.
Standalone, versioned specifications for each protocol. Language-agnostic. Adopt them piecemeal or together.
JSON Schemas for dependency-jurisdiction.json and compliance records. Validate your configs before shipping.
Generic Python: jurisdiction-audit produces reports; compliance-deadline-tracker warns progressively ahead of statutory deadlines. No Matrix required.
Example configs, incident-response playbook (Art. 33(3) fill-in-blanks), CLAUDE.md starter, and gate hook scripts.
Cost trajectories, context-pressure curves, and state transitions captured by the protocols are structured research artefacts — fodder for POMDP, information-theoretic, and causal-inference analyses.
Solo developers, researchers, and small teams face the same regulatory questions as enterprises. GDPR Article 48, the US CLOUD Act, NIS2, and related frameworks interact in ways that are often opaque — but the protective side is readily usable once understood. The IAS individual-protection guide walks through the framework in neutral, practical terms and shows how the regulatory tools in this repository give you a timestamped record trail without Matrix, AGPL code, or any particular infrastructure.
Generate a project-specific CLAUDE.md with a wizard, using the IAS template.
pip install jinja2 pyyaml git clone https://codeberg.org/openearth/ias.git cd ias # interactive wizard (~10 questions) python3 tools/generate-claudemd.py --output /path/to/your-project/ # or from a saved config (reproducible) python3 tools/generate-claudemd.py --config templates/examples/python-cli.yaml \ --output /path/to/your-project/
Three worked example configs ship in templates/examples/: python-cli.yaml, kotlin-jvm.yaml, go-service.yaml. Each shows a different gate-selection shape (container build on/off, staged deploy on/off, DICSTAMACH on/off). Copy, edit, run.
foundation.protocols.ai.ias.claude.*
The DICSTAMACH state machine is also formalised as a Matrix event family: 8 primary events (session start/end, domain loaded/unloaded, gc proposed/accepted/declined, pressure threshold) with 2 reusable verify sub-types that allow an independent observer (a gryph-style hook) to confirm state transitions from outside the agent.
The intention layer is emitted by the Claude bridge observing markers in the agent's output; the verify layer is emitted by a gryph-based hook observing Read / Write tool calls. Correlated by Matrix's native m.relates_to using the canonical foundation.protocols.verify.v1 rel_type. Matching primary + verify events indicate a well-behaved session; divergence is an audit signal (hallucination, silent action, transport failure).
IAS captures agent-session state at every protocol checkpoint. The resulting time series is research-grade data for:
The scratch/bootstrap-test-results-*.md files in the repository are the first published analyses under this programme (meta-circular bootstrap applied to governance specifications).
IAS is Apache-2.0 and welcomes contributions. Three ways in:
Run the wizard on one of your own repos. File an issue on Codeberg if anything feels friction-heavy. The repo's issue tracker is the canonical feedback channel.
Add a new protocol, extend an existing one, contribute a language-specific example config. Pull requests via Codeberg follow the contribution path documented in templates/README.md.
The IAS data schema is the foundation for empirical work on human-AI coding interaction. If you are running analyses or want to collaborate on open-science outputs, the issue tracker is the place to start.
\|/ias/|\ — IAS was initiated by Dr. Raazesh Sainudiin under the OpenEarth Network initiative, with development sponsorship from VakeWorks AB. The DICSTAMACH protocol, gate system, CONDOC methodology, and L2D&D\|/D2L&D spiral emerged through iterative collaboration with Claude Code — the framework is itself a product of the interaction it seeks to formalise.
Copyright 2026 The IAS Authors. Licensed under the Apache License, Version 2.0.