Axiom-Grid

Local writing. Explicit control. A focused writing copilot derived from Kairo-Phantom.

Engineering preview, not production-ready. The working slice generates a reviewable suggestion from text you explicitly provide. It does not automatically type into other applications. That feature is gated until target-bound approval and native-platform tests pass.

What is implemented

• Independent Rust preview runtime (axiom-runtime), with no inherited swarm, cloud adapters, plugins, ambient screen capture or injection in its launch path.
• Direct loopback Ollama inference: no proxy use, no redirects, no automatic downloads, no configured cloud fallback.
• Authenticated model-readiness inventory with exact-tag matching and a fail-closed desktop status UI.
• Native IPC bearer authorization; browser-origin and invalid Host requests denied; bounded input/output and one active generation.
• Axiom desktop preview UI with explicit context, literal text rendering and late-result discard.
• Regression coverage for Unicode, session acceptance/cancellation, model identity, CRDT replacement, API security and UI behavior.

Start here

Quickstart · Current handoff · Fork handoff · Launch gates

cd axiom-runtime
cargo test --locked
cargo clippy --locked --all-targets -- -D warnings

An already-installed Ollama model is required for real inference. Local tests use a mock model server, not benchmark evidence.

Quickstart (cold install, Python sidecar)

Use Python 3.11–3.14. Install the focused overlay dependencies, not the inherited requirements.lock (which is neither a complete nor an exact lock):

git clone https://github.com/Kartik24Hulmukh/Axiom-Grid.git && cd Axiom-Grid
python3 -m venv .venv && source .venv/bin/activate
pip install -r docker/requirements-runtime.txt
make serve          # local overlay API on http://127.0.0.1:8765

make run DOC=... Q="..." is the separate grounded-Q&A CLI, not a server. For the focused regression suite, install pytest pytest-asyncio hypothesis httpx ruff psutil and run make pre-push. make test is the kernel/pack suite; it does not mean all inherited tests have been run.

Launch ingestion supports text/Markdown, text-bearing PDF (pdfplumber) and DOCX (python-docx). Scanned PDFs require a separately qualified OCR workflow; XLSX/PPTX are not supported by this overlay ingestion path. PDF boxes enclose page text, not precise field-level highlights; DOCX pagination is estimated. The overlay intentionally uses a deterministic demo gateway. A healthy overlay is not evidence of live model integration or production document Q&A.

Why the sidecar is Python

The launch runtime is Rust, but document intake runs in a Python sidecar on purpose: the OCR and layout engines the product depends on (Docling, pdfplumber, the Office readers) are python-native, and bbox-accurate layout is what makes every suggestion citable. Rust keeps the UI, IPC and security boundary; Python keeps OCR/layout. The boundary is a local loopback contract, so the sidecar can be restarted or sandboxed without touching the runtime.

Platform support

Ghost-typing into third-party applications stays disabled on every platform until target-bound approval and per-platform native CI exist. See CROSS_PLATFORM_REPORT.md for the per-platform evidence matrix.

Concurrency posture (production)

The overlay runs pipelines concurrently under a bounded gate (AXIOM_PIPELINE_CONCURRENCY, default min(32, 4 x cores)). Shared services (ProvenanceLogImpl, MemoryStoreImpl) are thread-safe at class level via kernel/core/threadsafe.py, and SQLite runs in WAL mode with a 5s busy timeout. No global mutex serializes requests.

Scope and provenance

Imported from Kartik24Hulmukh/Kairo-Phantom at 8975743. No upstream changes. MIT license and source history retained; inherited Rust crate names remain compatible during extraction.

Inherited release workflows and launch documents are archived in docs/inherited/. Broad inherited source remains outside the focused executable until dependency-safe pruning is complete. Do not treat inherited reports as Axiom launch evidence. Do not launch the legacy kairo-phantom binary as the Axiom product.

No claim of universal app support, certified zero egress, enterprise compliance, specified model speed, product-market fit or “100x” traction has been verified. Signed event logs cannot by themselves prove absence of network egress.

Production deployment (overlay API)

The hardened FastAPI overlay ships with ops probes and a production container:

# Containerized one-liner
docker build -t axiom-grid -f docker/Dockerfile.overlay .
docker run --rm -p 127.0.0.1:8765:8765 -e AXIOM_API_KEYS="$(python3 -c 'import secrets; print(secrets.token_hex(32))')" axiom-grid

# Ops endpoints
curl localhost:8765/healthz   # liveness
curl localhost:8765/readyz    # real-pipeline readiness (cached, load-safe)
curl localhost:8765/metrics   # runtime counters & latency

Hardening posture (all regression-tested in overlay/tests/test_server.py and the 100x/600-request stress gauntlets): SEC-001 sandboxed path resolution, SEC-002 origin gate, SEC-003 defense-in-depth response headers, SEC-004 bounded extraction pool + payload-size governor, OPS-001 liveness/readiness/metrics.

Scope Boundaries (Kairo Phantom contract)

Axiom-Grid ships the Kairo Phantom document-intelligence core. The scope contract below is enforced by tests/test_scope_discipline.py and mirrored in CONTRIBUTING.md and docs/PUBLIC_ROADMAP.md — change all three together.

Kairo DOES

• READ documents (Word, Excel, PowerPoint, PDF, code, email, design) through the extraction pipeline behind admission control and bounded execution.
• SUGGEST grounded answers with page/line-level citations and a provenance receipt chain (opik_trace_id ↔ receipts.jsonl).
• Route model calls through the Melious gateway with per-route circuit breakers, reasoning-token budgets and dynamic fallbacks.

Kairo Does NOT

• Write, edit, send or execute anything on the user's behalf — the product is READ + SUGGEST ONLY. Every mutation stays a human decision.
• Answer without evidence. No source → no answer: if grounding cannot bind a claim to an extracted span, the response is a refusal, never a guess.
• Persist BYO API keys anywhere except the OS keychain abstraction (scripts/keychain_store.py); config files and logs are scanned for leaks.

Inherited document-pipeline development quickstart

This is the Python document-intelligence test/development path, not the Rust desktop preview launch path above. The Python sidecar integrates Python-native OCR/layout and document parsing libraries; numpy supports local embedding calculations. Use a clean Python 3.12 environment (the inherited embed-anything dependency pins an ONNX Runtime wheel unavailable on Python 3.14). Do not bypass dependency resolution with --no-deps.

python3.12 -m venv .venv
. .venv/bin/activate
python -m pip install -r requirements-test.txt pytest pytest-xdist httpx uvicorn
python -m pytest --import-mode=importlib -q tests/
make run DOC=samples/invoice/sample_invoice_01.txt Q="What is the invoice number?"

The Q&A command requires GNU Make; on Windows use WSL or invoke python scripts/qa_pipeline.py --doc samples/invoice/sample_invoice_01.txt --question "What is the invoice number?". Semantic PDF tests require model2vec and the committed weights in kairo-sidecar/assets/models/potion-base-8M/. Keep KAIRO_REQUIRE_SEMANTIC=1 for semantic validation: hash fallback is not semantic retrieval evidence. The dependency list is not a production lockfile and these commands do not certify that the broad suite is green. See PR #20 for outstanding launch gates.

Platform support boundaries

Python API tests on Linux do not certify native application integration on any platform. No automatic cross-application typing is enabled by this quickstart.