Semantic Graph PoC 0.1

Goal

Test whether a program can be automatically lowered from source syntax into a new symbolic/graph representation, then executed from that representation without reading the original source again.

Experiment A — control flow

Original behavior contains arithmetic, an intermediate local value, a conditional branch, and two returns.

Pipeline:

Python source
  -> Python AST
  -> anonymous slots + semantic operation nodes + branch blocks
  -> JSON semantic graph
  -> graph-only runtime

Test vectors: 6 All matched: True

作者： Neo.K
機構： EveMissLab／一言諾科技有限公司
日期： 2026-10-08（整理日期；原稿未標註日期）
版本： 0.1

Experiment B — real RepoLumen function

Target: repo_semantic/historical_timeline.py::_transition_id

Original semantic behavior:

two ordered inputs
  -> concatenate with NUL separator
  -> UTF-8 encode
  -> SHA-256
  -> hexadecimal digest
  -> first 24 characters
  -> prefix with `hist_`

Automatically produced semantic ops:

arg -> to_str -> const -> arg -> to_str -> concat_many -> utf8_encode -> const -> slot -> sha256 -> hex_digest -> const -> slice -> add

Test vectors: 5 All matched: True

Source-language decoupling check

The canonical graph deliberately removes local/source identifiers.

{
  "branch_graph_contains_original_var_x": false,
  "branch_graph_contains_original_var_y": false,
  "branch_graph_contains_original_var_base": false,
  "transition_graph_contains_from_snapshot_id": false,
  "transition_graph_contains_to_snapshot_id": false,
  "transition_graph_contains_hashlib_name": false,
  "transition_graph_contains_hexdigest_name": false,
  "transition_graph_contains_encode_name": false
}

All values above should be false.

The graph retains only:

• ordered argument positions;
• anonymous local slots;
• constants that affect behavior;
• semantic operations;
• control-flow structure.

Graph-only process

A separate script, runtime_only.py, loads only the JSON graphs and the generic semantic runtime. It does not contain the original functions.

Observed output:

branch graph: 10
RepoLumen transition graph: hist_4aa9461975c271a428d60476

Result

Within the supported subset, the experiment establishes:

1. source syntax is not required after lowering;
2. original variable names are not required;
3. a program can be represented as typed semantic nodes plus control-flow structure;
4. the same graph can be executed by a generic graph runtime;
5. a real RepoLumen function can be canonicalized from source-specific spellings such as .encode("utf-8"), hashlib.sha256(...), and .hexdigest() into utf8_encode, sha256, and hex_digest semantic operations.

This is not yet a proof that arbitrary Python or arbitrary binaries can be losslessly lifted. The difficult next layers are mutation/aliasing, exceptions, generators/coroutines, dynamic dispatch, reflection, external effects, concurrency, C/ABI memory, and runtime-generated code.

Files

• semantic_ir.py — compiler + graph runtime
• experiment.py — compiler/equivalence experiment
• runtime_only.py — graph-only execution
• graph_branch.json — control-flow graph representation
• graph_repolumen_transition.json — RepoLumen function semantic graph
• run_results.json — equivalence results

Stronger isolation / property test

A physically separate runtime directory contains only:

- graph_branch.json
- graph_repolumen_transition.json
- graph_runtime.py
- isolated_runtime_only.py

It contains no compiler and no original function source.

Isolated execution result:

10
hist_4aa9461975c271a428d60476

Random equivalence testing:

• arithmetic/control-flow graph: 10,000 / 10,000 matched
• RepoLumen _transition_id graph: 1,000 / 1,000 matched
• compiler/source present in isolated runtime directory: False

Therefore the semantic representation remains operationally executable after the source-language compiler and original source are absent.