Benchmark and launch evidence

This Runtime repository separates engineering verification from research evaluation. Synthetic contract tests prove that code paths compose; they do not prove that a model understands a real person or predicts future behavior.

Public Runtime gates

Question	Method	Pass criterion
Can a fresh store form the complete model contract?	`tests/test_runtime_model_e2e.py`	Point, Line, Face, Volume, Root, receipts, redacted export, and viewer routes all present after two deterministic builds
Can an agent retrieve and verify a fact?	`scripts/sample_demo.py` plus `scripts/verify_sample_mcp.py`	the real streamable HTTP transport lists the required tools, search returns a live entry, and `read_receipt` resolves the same ID/path/content
Does the offline Runtime remain stable?	`PERSOME_LLM_MOCK=1 uv run pytest -m "not macos and not integration"`	all selected tests pass without network or provider credentials
Does the package work outside the checkout?	build wheel, install in a fresh virtualenv, run CLI, inspect bundled resources	CLI starts; Swift sources, Three.js, and PP-OCRv6 weights exist in site-packages
Is committed content publication-safe?	secret, PII, and repository-language scans	zero findings outside the explicit OCR character-data exception

Experience targets

Metric	Target	Interpretation
Median source install	at most 10 minutes	three clean installer runs on a supported Mac; report cache/network conditions
First useful recall	at most 10 minutes	valid capture plus configured semantic provider reaches searchable durable memory within two default five-minute flush windows
Viewer availability	immediate after daemon HTTP start	sparse geometry is allowed and must be labeled degraded

The default cadence supports the recall target by construction, but environment, provider latency, permissions, and evidence quality can delay real data.

v0.2.2 launch observations

Measured on 2026-07-11 on an Apple Silicon Mac running macOS 26.3.1. The host already had uv and its dependency cache; each run used a new PERSOME_INSTALL_HOME, new virtualenv, new CLI directory, no provider key, and the installer-managed Python 3.12 path.

Observation	Runs	Result
Isolated source install	11.716s, 11.896s, 11.926s	median 11.896s; passes the 10-minute target under recorded warm-cache conditions
Complete deterministic synthetic model	`tests/test_runtime_model_e2e.py`	2.37s wall clock; Point through Root plus export/viewer assertions
Streamable HTTP MCP retrieval	16 tools discovered; `search` then `read_receipt`	required model/correction tools present and receipt matched ID, path, and content

These are launch-machine engineering observations, not population latency percentiles. A cold network, first Python download, real provider latency, or missing macOS permission can take longer. Real-person First Useful Recall has a ten-minute operational target but is not represented as a measured benchmark.

What is not measured here

The following belong in a separate persome-bench repository with explicit datasets, consent, baselines, metrics, and licenses:

personal-memory precision and recall;
temporal fact freshness and contradiction resolution;
next-state or next-action prediction;
calibration and abstention;
cross-user generalization;
comparison with screenpipe, Mem0, or provider memory;
real-person longitudinal utility.

Until those artifacts exist, Persome makes no paper-level superiority claim. The README comparison describes product boundaries, not benchmark wins.

Reproduce the synthetic proof

PERSOME_LLM_MOCK=1 uv run pytest tests/test_runtime_model_e2e.py -q
PERSOME_LLM_MOCK=1 uv run python scripts/sample_demo.py --no-open
# In another terminal:
uv run python scripts/verify_sample_mcp.py
uv build

The committed fixtures under tests/fixtures/runtime_model/ contain no real personal data. See Runtime validation for the complete gate.