1. Research question and hybrid structure

Can interpretation be separated from repeated execution, so a routine business question becomes a reusable, inspectable operation instead of a new probabilistic answer every time?

The hybrid boundary is explicit: learned interpretation proposes a specification; symbolic admission, review, and execution determine what runs. This is not a claim that every stage is weightless. It is a design that removes model calls from the execution path of an already admitted tool.

2. From question to symbolic trace

1. Compose a specification
2. Check the real data
3. Approve and sign
4. Run and inspect

The implemented pipeline grammar is:

rows (filter | join)* [group_by] reducer [top]

Each named step has defined semantics. The documented gate runs ten checks, including grammar, existing columns, data support, determinism, rejection of a wrong answer, abstention without evidence, and grounding under data perturbation. The runtime’s verifier is derived from the pipeline rather than authored by the proposing model.

The trace records the tool, data table, question class, value, derivation, and model-call count. Invalid data can produce REFUSED; an absent table produces NO_EVIDENCE; no matching rows can produce NO_VALUE. These outcomes remain distinct.

3. Meaning still requires review

A procedure can be perfectly deterministic and answer the wrong question. The recorded gate case called “total of paid invoices” filtered open invoices: mechanical checks passed because the procedure was correct as written.

The documented release therefore requires a person to inspect the readback and answer and sign approval. Tools and reviews are bound to digests and signatures; the execution runtime holds verification keys, not admission authority. This boundary makes the model’s proposed interpretation reviewable. It does not make a reviewer infallible or protect a host controlled by an attacker.

4. Existing evidence, not new benchmarks

The 27 September 2026 technical account reports the following historical observations from the pilot. They have not been remeasured for this report.

The existing account also reports one-machine load, memory, and timing measurements and automated checks. Those are dated pilot results, not a benchmark against another product or evidence of a 90% savings guarantee.

5. Limits and evaluation plan

• The published vocabulary handles bounded data procedures, not arbitrary legacy application behavior.
• The documented pilot lacks parameterized tools, relative-date questions, direct database connections, and a self-serve admission gate.
• Arithmetic can be exact while interpretation or source data is wrong.
• The authoring and review cost must be included in any comparison with per-query model calls.
• Model-free local execution does not by itself establish an energy, processor-price, or safety-certification claim.

A broader system-conversion study should report supported system classes, refused mappings, time to admission, semantic-review failures, and trace completeness. Compare total cost over repeated queries on the same tasks, including human review and changes to data schemas.

6. Primary sources and related research

1. How Perslis Symbols works, 27 September 2026: specification vocabulary, gate, trust model, recorded transcripts, dated measurements, and limits.
2. Perslis Floor source and releases: executor and signed release artifacts.
3. Runtime distribution: install and verification instructions.
4. Fail-First Models: related work on learning inside fixed authority. It is not a replacement for this Symbols report.