research · technical · jul 18, 2026
Near-deterministic
by design.
Nobody signs off on a moving target. The framework behind Palm Ark exists to answer one question a generative model cannot: will you say the same thing tomorrow?
12 min read · by the palmdigitalz research group
The problem with brilliance
Ask a frontier LLM to extract the business rules from a 50,000-line COBOL program and it will produce something impressive. Ask it again and it will produce something impressive — and different. Different rule counts, different phrasings, different omissions. For a demo, that's fine. For a regulated migration where the extracted rules become the specification of a bank's core, it is disqualifying.
The failure isn't intelligence; it's epistemology. A generative model produces the most plausible answer, and plausibility is resampled on every run. An auditor doesn't ask whether the answer is plausible. An auditor asks whether it is traceable, complete, and stable — three properties sampling cannot give you.
What "near-deterministic" means
Fully deterministic tools exist: parsers, static analyzers. They never disagree with themselves — and they also never understand anything. They can tell you that paragraph 2400-CALC-PREM moves a value; they cannot tell you it applies a senior-citizen discount.
Palm Ark's framework is built on a simple division of labor: everything that can be computed is computed; the model is only asked questions whose answers can be checked. Deterministic structure carries the load; bounded intelligence fills in meaning; verification closes the loop. We call the result near-deterministic: same input, same output, run after run — with the semantic understanding a parser alone can never produce.
Five stages, one hash
The estate is parsed into a full syntactic model — programs, paragraphs, copybooks, JCL, data definitions. Deterministic. No model involved.
Every candidate rule site is grounded in the parse tree: control flow, data lineage, and the exact source lines it depends on. The model never sees floating text — it sees anchored structure.
Gazelle, our bespoke extraction model, names the business meaning of each grounded site — under constrained decoding, against the anchored evidence, one bounded question at a time.
Every extracted rule is checked back against the source: does the cited code actually imply the stated rule? Rules that fail verification don't ship. Unsupported claims cannot survive this stage.
The final artifact — rules, lineage, evidence — is content-hashed. Run it three times: one hash. That hash is what your reviewers pin to the wall.
three runs · one hash · see palmbench
Why lineage is non-negotiable
Every rule Ark emits carries its full source lineage — the paragraphs, conditions, and data items it was derived from. This isn't a nice-to-have; it's the mechanism that makes the whole framework honest. A rule with lineage is a claim your auditors can click. A rule without lineage is a rumor with good formatting.
Lineage is also what makes the ~95% figure meaningful. Accuracy claims from generative extraction are self-graded — the model asserts, nobody can check. On PalmBench, every scored rule is checkable against the line it came from, which is why we publish the number and invite you to reproduce it.
What this buys you
Audits pass. The extracted specification is stable across runs and evidenced line-by-line — the two things a reviewer actually checks.
Parallel teams agree. When extraction is repeatable, your SI, your internal team, and your regulator are all reading the same artifact — not three samples from a distribution.
The estate becomes a database. Deterministic output means you can diff it, version it, and re-run it after every change window. Extraction stops being an event and becomes infrastructure.
Determinism is demonstrable.
We'll run the framework on your code, in your perimeter — three runs, one hash, in front of your team.
See it on your code