Phi9 / research & engineering

Run the work

Small working research implementations with source, explicit limits and reproducible outputs.

CONSTRUCTED MODELS · BROWSER EXECUTION

Make the representation visible

Move the two-box split, inspect oscillator phase and expose a binary32 composition failure. Each model states what its inputs mean and what the demonstration does not prove.

Explore the three representations →

RETAINED MEASUREMENTS · NOT RERUN HERE

Read the result that changes the story

Calibration lowers false acquisition in two null conditions. Full sensing still wins on the planted signal. Inspect every condition, including charged work and uncertainty.

Inspect the measured comparison →

Follow the fourteen-chapter working book from intuition to the evidence and its limits.

WORKING OFFLINE WORKFLOW · PYTHON 3.10+ · POSIX/LINUX

One reproducible research release

Run the bounded interpreter, discovery benchmark, controlled extension and artifact pipeline together. The completed release contains 5,856 freshly executed and replayed cases, 75 passing checks, recorded inputs, source pins, traces and a public-safe report.

Download the reproducible system source · Read the runnable guide

Unzip the source package and enter its phi9-release-system directory. Execute and replay a fresh local release:

python3 release.py run --output runs/my-run
python3 release.py replay --output runs/my-run
python3 release.py status --output runs/my-run

Existing output directories are not overwritten. The source-only download omits private execution receipts and development history. The workflow checks component pins and produces a versioned research report. Execution uses trusted local Python code; it is not a security sandbox.

This runs archived program execution and finite simulated experiments—not fresh model inference, a hosted agent, automatic private-account access or an always-on service. No provider calls or new provider charges are part of this release.

Understand the results and remaining failures · Download the lighter public review

PYTHON STANDARD LIBRARY · ARCHIVED MODEL INPUT

Bounded token-program interpreter

Validate a task-local program, transform a vector and inspect the readout. Supports scale, add, dot and readout; unknown instructions reject. Admission checks run before each operation. Budget exhaustion abstains.

Download source, program and tests

PYTHON STANDARD LIBRARY · SIMULATED BENCHMARK

Discovery engine and failure record

Inspect hypothesis elimination, observation costs and why a singleton does not justify acceptance. Includes the revised engine, protocol, raw runs, matrix audit and identifiability check.

Download engine and evidence

Try the transformation

A browser replay of the same four-operation plan. No model call, external data or account access.

First coordinateSecond coordinateCompute budget

Ready: scale 2 → add [1, −1] → dot [1, 1] → readout.

Download the Python implementation below for strict JSON validation and the full rejection tests. This preview uses a fixed allowlisted plan, not arbitrary program execution.

Replay the token program

Unzip the package and enter its token-program-v1 folder:

python3 interpreter.py < codex-program.json
python3 -m unittest -v test_interpreter

The replay uses an archived model-generated program and makes no new model request. The trace records [1, 2] → [2, 4] → [3, 3] → scalar 6.

Verify the discovery implementation

Unzip the discovery package and enter its discovery-revision-v1 folder:

python3 -m unittest -v test_discovery test_revision
python3 discovery.py audit

Read protocol.json before interpreting results. The 12-unit observation cap cannot cover the 13-unit full probe domain; acceptance and post-run correctness are distinct.

The asset shelf

Readable chapter PDF — elementary criterion and proposed experiment.
Program execution report (PDF) — demonstrated synthesis and bounded replay.
Grouped experiment results — 1,440 revised runs.
Asset manifest — file sizes and SHA256 checksums.

Materials are provided for research inspection and evaluation; downloading does not grant a blanket commercial licence. The public subset excludes account/session data, internal outreach, private Geist material and noncommercial third-party datasets.

From implementation to a useful engagement

These are local research tools, not a hosted agent service. There is no promised always-on runner, payment system or automatic access to private accounts. Discuss a scoped evaluation or implementation.