Work from the record
Resolve an entity, retrieve available facts, and retain the source receipts behind them.
Perslis model research / Lois
Lois: Symbolic Scientific Reasoning with Inspectable Evidence
A reasoning model that keeps the basis of an answer visible: typed facts, explicit relationships, source receipts, and named evidence gaps.
Technical research report · research prototype · not presented as a peer-reviewed journal publication
Abstract
Lois is Perslis’s symbolic reasoning model for scientific research. Its symbolic core resolves entities, recalls source-pinned facts, traverses typed relationships, and checks constraints expressible in the available knowledge. This report distinguishes that core from the optional language-model research interface around it. The contribution is an inspectable separation of evidence, derivation, and generated wording—not a claim that a citation proves a scientific conclusion. Lois is not Peel; Peel’s publication remains a separate research contribution.
1. Model identity and research question
Can a research assistant expose what supports an answer, what operation produced it, and what remains unknown? Lois approaches that question through symbolic operations rather than treating fluent text as evidence.
Lois and Peel are distinct. Lois is the reasoning model described here. Peel has its own research on a source-pinned symbolic store and structural admission of facts. Related components may be used in one Perslis deployment; that does not make their names or papers interchangeable.
2. Architecture and trust boundaries
- Resolve an entity
- Retrieve source-pinned facts
- Apply a defined operation
- Return evidence or a gap
| Component | Role | Boundary |
|---|---|---|
| Symbolic core | Entity resolution, recall, typed paths and encoded checks. | No neural model is required to perform the implemented floor-lane operations. |
| Source connectors | Fetch records with their source URLs. | A source outage is not proof that a record is absent. Live extraction still requires a connection. |
| Optional research interface | Plan lookups and draft prose from retrieved records. | The hosted interface can use a language model. Its prose is not the symbolic core’s proof. |
The inspected lois/symbolic.py implementation resolves accessions, reads stored provenance-pinned edges, extracts missing records through connectors, and walks typed paths. lois/service.py explicitly distinguishes a model-free floor lane from the optional research lane and records why a fallback occurred.
3. Evidence representation and answer status
In the research lane, evidence records have an ID, source, kind, title, text, and HTTPS receipt. The evidence log rejects empty text or a missing HTTPS receipt. These requirements establish a trace to a fetched record; they do not establish that the source itself is correct.
The prose checker labels sentences cited, unsupported, or reasoning. It checks citation existence and identifier/number tokens mechanically. This is a lexical check, not semantic entailment: a misleading paraphrase can still contain all the right tokens. Unverified reasoning must remain visibly separate from evidence-backed content.
A relationship path likewise establishes a path through the recorded graph. It is not automatically a causal explanation, an experimental result, or a treatment recommendation.
4. Evaluation and reproducibility
On 8 October 2026, the local tests.test_lois suite completed successfully: 109 tests run, 2 skipped. The suite uses canned source bodies, a scripted gateway, and fake database transactions. It is contract/regression evidence, not a live scientific accuracy benchmark.
python3 -m unittest tests.test_loisPer-tool checks cover receipts, the source’s own returned content, missing records, and upstream failures. The verification tests exercise citation and token handling. The recorded summary is available in the validation record.
A separate 4 October dogfood record reports hosted REPL observations and preview/production differences. Those dated observations are not a current guarantee of production parity. No new live connector, energy, or end-to-end research benchmark was performed for this report.
5. Limitations and next experiments
- Knowledge and rules are incomplete. Unknown must not be read as false.
- Source receipts preserve provenance, not universal truth or clinical validity.
- Live source extraction is not an offline capability; only available local knowledge can be used without those sources.
- Optional generated prose is probabilistic and its lexical checker does not prove every sentence.
- No global zero-error, certification, or hardware-cost guarantee follows from these tests.
Further evaluation should freeze source snapshots, preregister supported/contradictory/unknown queries, measure false support and useful refusal separately, test corrupt and unavailable sources, and publish replayable traces. Any hardware claim needs measured memory, latency, energy, and device details.
6. Primary sources and related research
- Lois’s symbolic reasoning documentation: query, path explanation, encoded conflict, and unknown.
- Lois’s research interface: source lookup and the hosted-model boundary.
- Inspected implementation:
api/_lib/lois/symbolic.py,service.py,evidence.py, andverify.py; local checkout reviewed 8 October 2026. - Validation:
tests/test_lois.py; dated run summary. Historical observation:scripts/LOIS_DOGFOOD_2026-10-04.md. - Peel’s separate paper and The Verification Floor are related research, not substitute Lois papers.