📖 Read the full launch post and watch the 2D tensor collision demo here: https://maniksundar.substack.com/p/the-physics-illusion-why-llms-still

Problem Statement

LLM-generated unit tests for KV-cache routing kernels suffer from a silent failure mode: the LLM hallucinates the same bug in both the implementation and the test, causing the test to pass while the kernel remains incorrect. This happens because LLMs reason from the same flawed mental model when writing both code and tests. ImpactArbiter addresses this by using a two-stage RAG pipeline: first, a Distill Agent extracts and summarizes the routing logic from the actual research paper; second, a Coding Agent writes the implementation and test based on that summary. The generated code is then run through a PyTorch autograd trap that compares gradient signatures against SymPy oracles. The trap catches bugs that unit tests miss, even when the LLM's own test_route() assertions pass.

Setup

1. Create Virtual Environment

python -m venv venv
source venv/bin/activate  # On Windows: venv\Scripts\activate

2. Install in Development Mode

3. Configure API Keys (Choose Your Provider)

OpenAI:

export OPENAI_API_KEY="your-openai-api-key"
# On Windows (PowerShell): $env:OPENAI_API_KEY="your-openai-api-key"

Claude (Anthropic):

export ANTHROPIC_API_KEY="your-anthropic-api-key"
# On Windows (PowerShell): $env:ANTHROPIC_API_KEY="your-anthropic-api-key"

Gemini / Vertex AI:

# Ensure gcloud is authenticated and project is set
gcloud auth login
gcloud config set project impactagent
export GOOGLE_CLOUD_PROJECT="impactagent"
# On Windows (PowerShell): $env:GOOGLE_CLOUD_PROJECT="impactagent"

For persistent configuration, add these to your .env file:

OPENAI_API_KEY=your-openai-api-key
ANTHROPIC_API_KEY=your-anthropic-api-key
GOOGLE_CLOUD_PROJECT=impactagent

Demo Command

impactarbiter auto-heal --oracle radix --model gemini

Additional Flags

• --full-agent-trace: Display LLM Chain-of-Thought reasoning before code generation and heal attempts
• --live: Use live LLM API calls instead of cached deterministic replay (requires API key)
• --mock: Run offline evaluation with deterministic replay (default if no API key)

Example with live LLM generation and full trace:

impactarbiter verify --workflow agentic-kv-scheduler --full-agent-trace --live

Sample Output (2D Asymmetric Ring Buffer)

─────────────────── IMPACT ARBITER — AUTO-HEAL ───────────────────
Model: vertex_ai/gemini-2.5-pro
[PAPER DOWNLOADED]
https://arxiv.org/pdf/2312.07104.pdf

[QUICK DISTILL]
### KV Cache Routing Specification: Planner-Executor Handoff
...
[GENERATED CODE & TESTS]
def route_radix_2d(b_local_idx, head_idx, prefix_length_h, total_blocks_h, block_size):
    k = prefix_length_h + b_local_idx
    logical_block = k // block_size
    offset = k % block_size
    return (head_idx, logical_block, offset)

[LLM UNIT TEST PASS ✅]
LLM self-validation passed.

[AUTOGRAD TRAP FAIL ❌ HARD_BLOCK]
divergence=1.00e+00 > tol=1e-04

GRADIENT DIVERGENCE MAP — KV_cache.grad (head × block × offset)
Token (b=5,h=0,prefix_h=60,N_h=4) | Expected: head=0 block=0 offset=1 | Got: head=0 block=4 offset=1
Non-zero gradient at: [0, 4, 1, :] — misrouted 128 floats

[AUTO-HEAL attempt 1/3]
def route_radix_2d(b_local_idx, head_idx, prefix_length_h, total_blocks_h, block_size):
    absolute_idx = prefix_length_h + b_local_idx
    logical_block = (absolute_idx // block_size) % total_blocks_h
    offset = absolute_idx % block_size
    return (head_idx, logical_block, offset)

[FINAL PASS ✅]
divergence=0.00e+00 (after 1 heal attempts)

On LLM non-determinism and trap reliability

The autograd trap itself is fully deterministic, which means identical code always produces identical gradient divergence results.
What varies is whether the LLM generates correct or incorrect routing logic on a given run.
This mirrors real production reality: agent-generated serving code sometimes passes, but often silently fails on ragged boundaries.
ImpactArbiter gives you deterministic verification of whichever code the agent produces, so you're not relying on hoping the model "got it right this time."

In practice, Gemini 2.5 Pro generates incorrect routing on roughly 65% of attempts for the critical 2D ring-buffer wrap cases. The trap catches every incorrect implementation with zero false negatives.

Coverage & Field Results

2D RadixAttention Test Matrix (Default)

Recommended for production-relevant demo.

Legacy PagedAttention

Boundary fixtures [15, 99, 100, 105, 128] are maintained for historical comparison mode.

Summary statistics from most recent evaluation run:

Repository layout

src/
├── oracles/         # SymPy ASTs + lambdified callables
├── trap/            # autograd trap & ASCII divergence map
├── fuzzer/          # explicit boundary fixtures
├── cli/             # auto-heal pipeline + litellm agent + paper extractor
└── db/              # nextpaper.db (SQLite) validation_traces
tests/
├── test_paged_oracle.py
├── test_radix_oracle.py
└── test_trap.py

Running the tests

The four load-bearing claims in tests/test_trap.py must all pass.

Contributing

ImpactArbiter is an open project — contributions of new oracles, fuzz cases, and bug reports are welcome. See CONTRIBUTING.md for the full contributor guide. A short summary:

• Reporting issues: Open a GitHub issue using the relevant template under .github/ISSUE_TEMPLATE/. Bug reports must include the exact CLI command, the model used, the divergence map (or stack trace), and the contents of nextpaper.db row(s) when relevant.
• Contributing a new oracle: Every oracle must ship as a triple — (1) a SymPy AST plus a lambdified callable in src/oracles/, (2) a deterministic autograd trap in src/trap/, and (3) explicit boundary fixtures in src/fuzzer/. PRs without all three will be sent back.
• Discussing methodology: Open a methodology issue — these are reviewed weekly and used to calibrate the mock hallucination rates and per-oracle session windows.

Issue types

Reporting a bug — minimum reproducible report

A bug is only actionable when we can replay the failure deterministically. Please include:

1. Command line — The exact impactarbiter ... invocation, including all flags.
2. Environment — Python version, OS, and whether --live or --mock was used.
3. Model identity (if --live) — e.g. vertex_ai/gemini-2.5-pro. Do not paste API keys.
4. Failure surface — Paste either the gradient divergence map, the auto-heal stack trace, or the results.csv summary block.
5. Expected vs actual — What the oracle predicts vs what the agent / trap returned.

If the bug is in the trap itself (false PASS or false FAIL), attach the offending agent function and the corresponding boundary case from src/fuzzer/. Trap correctness bugs are treated as P0.

License

MIT.