Locally Correct, Globally Wrong

Companion code for Argentis Labs Working Paper 01 (2026): Locally Correct, Globally Wrong: A state-surface partition for authorization verification in autonomous systems.

Paper PDF · Argentis Labs

---

The claim

Verification systems for autonomous transactions today evaluate semantic alignment between a user's stated intent and the proposed transaction artifact. This evaluation surface is incomplete in a categorical way. Institutional authorization state — designation chains, disclosed conflicts, oversight independence — frequently lives outside the transaction artifact and requires relational reasoning across distributed schema fields.

On the central record (smoke_002), frontier LLM-as-judge configurations with full approver context approve the institutionally invalid record in five of five reruns at 0.93–0.97 confidence. The model reasoning explicitly cites every relevant field individually and validates each in isolation. A composed configuration that adds a deterministic predicate-dispatch layer over typed structural state recovers correct family attribution in five of five reruns at sub-millisecond authority-layer latency.

The architectural conclusion: institutional legitimacy must be represented as explicit relational state before it can be reliably enforced.

Reproduce the central result

git clone https://github.com/argentislabs/locally-correct.git
cd locally-correct
uv sync

Run the deterministic authority layer against the harder smoke record:

uv run python -m eval.harness \
    eval/records/handcrafted/type_d_captured_economic_dependence.json \
    --config authority_partition \
    --n-runs 5

Expected output: 5/5 REJECT(authority_partition.captured_approver) at 1.0 confidence, ≤1 ms median latency.

To reproduce the full §3.3 results table, run the corresponding configurations across both records (type_d_smoke.json and type_d_captured_economic_dependence.json). LLM-based configurations require OPENAI_API_KEY in the environment.

Verify every cited number

The paper's results table is auditable, byte-for-byte, against the trace files in eval/results/. The verifier reads docs/paper_01_citation_log.json and checks every assertion against disk:

uv run python -m eval.verify_citations docs/paper_01_citation_log.json

Expected: 43 assertions verified and exit code 0.

Repository layout

src/authority_partition/        the seven authority predicates + dispatch table
eval/configs/                   the four evaluation configurations (§2 of the paper)
eval/records/handcrafted/       the two smoke records cited in the paper
eval/results/                   per-run trace files (input to citation verifier)
eval/harness.py                 evaluation runner with --n-runs aggregate stats
eval/sweep_deterministic.py     run all seeds through the authority layer
eval/verify_citations.py        citation log verifier (84 assertions)
docs/paper_01_citation_log.json machine-checkable claim log
papers/                         working paper PDF

What's in this repo

• The seven authority predicates and dispatch table (src/authority_partition/)
• The four evaluation configurations from §2 of the paper: gpt_no_approver, gpt_with_approver, authority_partition, composed
• The two handcrafted smoke records the paper cites (smoke_001, smoke_002)
• The full evaluation harness, results traces, and citation verifier
• The working paper PDF

What isn't (yet)

• The cleanliness gate utility. Used to certify that handcrafted records contain no transaction-local alignment defects, so authority-surface failures cannot be confounded with local-surface noise. To be released alongside Working Paper 03 (cross-family evaluation across all seven authority families).
• The handcrafted seed corpus that the gate certifies. Same release moment.
• A Rust port of the predicate dispatch layer. Production-grade substrate; separate roadmap.

These are deliberate cadence choices. The paper's central claim — that frontier LLM-as-judge fails on the institutional-relational surface where deterministic predicate dispatch succeeds — stands fully on what's in this repo.

Citation

If you use this code or build on the paper:

@techreport{argentis2026locally,
  author      = {{Argentis Labs Research}},
  title       = {Locally Correct, Globally Wrong: A state-surface partition
                 for authorization verification in autonomous systems},
  institution = {Argentis Labs},
  type        = {Working Paper},
  number      = {01},
  year        = {2026},
  url         = {https://argentislabs.io/research/locally-correct}
}

License

Apache License 2.0. See LICENSE and NOTICE.

Contact

Research and consulting inquiries: argentislabs.io