Local Proof Bundle

Frozen historical snapshot (2026-09-23). The counts, hashes and logs in this directory describe commit f439e3c74cc9c36bcf7757d30f793ba5f7632939 only. They are not the current release gate and must not be cited as current test counts. For current evidence see docs/plan-completion-audit.md; raw logs are intentionally not rewritten so the historical bundle stays auditable.

This bundle proves the engineering state of commit f439e3c74cc9c36bcf7757d30f793ba5f7632939 without requiring a public push, clone, external package installation, or external users.

It is a local attestation, not a public/community proof.

Environment

See environment.txt.

Real upstream differential benchmark

The benchmark was rerun on this commit with the real upstream package langgraph==1.2.11.

Commands:

uv sync --directory benchmarks/differential
uv run --directory benchmarks/differential \
  python run_differential.py --refresh

Raw results:

• reactive-benchmark.json
• langgraph-benchmark.json
• differential-benchmark.log
• benchmark-summary.json

Output equality was asserted before timing was accepted.

Results

Workload	Reactive median	LangGraph median	Ratio	Reactive tasks/run	LangGraph tasks/run
chat	0.218 ms	1.147 ms	5.3×	2	2
wide	2.960 ms	3251.446 ms	1098.3×	1	1000
selective	1.847 ms	3180.731 ms	1721.8×	10	1000
nogain	120.178 ms	2750.486 ms	22.9×	1000	1000

The no-gain row is intentionally included: all tasks/nodes genuinely execute.

Tests

Logs:

• reactivegraph-tests.log — 152 passed, 1 skipped
• reactivechain-tests.log — 219 passed
• reactivegraph-sdk-tests.log — 12 passed
• reactivegraph-cli-tests.log — 9 passed
• e2e-tests.log — 43 passed
• deerflow-tests.log — 26 passed
• examples-deerflow-tests.log — 25 passed
• protocol-tests.log — 17 passed
• driver-tests.log — 192 passed

Static verification

• ruff.log — all Python packages passed
• mypy-reactivegraph.log — no issues
• mypy-reactivechain.log — no issues
• mypy-reactivegraph-sdk.log — no issues
• mypy-reactivegraph-cli.log — no issues
• protocol-typecheck.log — passed
• driver-typecheck.log — passed
• typescript-build.log — workspace build passed
• mkdocs-build.log — strict documentation build passed
• docs-links.log — 51 relative links resolved

Build artifacts

• build-reactivegraph.log
• build-reactivechain.log
• build-reactivegraph-sdk.log
• build-reactivegraph-cli.log
• artifacts.txt
• artifacts.sha256

All four Python packages produced a source distribution and wheel. Artifact SHA-256 hashes are recorded in artifacts.sha256.

What this proves

At this commit and environment, ReactiveGraph has:

• working selective execution
• lower wall-clock latency on all four tested workloads
• output parity with upstream LangGraph
• substantially fewer executed tasks on selective workloads
• full local regression/build/docs verification

What this does not prove

It does not prove:

• public package installation
• public CI execution
• external user adoption
• community trust
• cross-machine/cross-platform performance

Those require publication and external users.