The AI proposes freely (it fluctuates). We clamp it with routine (standard, repeated) and atypical (sudden, person-dependent) rules. Evaluation uses integers/rationals only — no floating point = bit-exact. Because rules are data (S-expressions), an atypical rule can be authored on the spot, deploy-free, and the rule version used for each judgment is pinned into the audit chain. Everything runs inside this tab. Zero upload.
Loading WASM…
1. Rule set(version = content hash. Add one rule → version changes = deploy-free versioning)
rule
kind
guard (S-expr)
effect
Spreadsheet mode(I/O anyone can picture. row = request, column = attribute, the "formula" of the verdict column = the rule set (S-expr). Edit a cell → re-evaluated instantly. Looks like a spreadsheet, but it is lisp-jit bit-exact, zero upload.)
req
category
amount
grade
project
approver
verdict
verdict "formula" = rule set, version
Template (.slimepkg)
2. AI proposal → clamp(try a preset; numbers are "k=v")
3-0. Write in natural language(In production a local Gemma4 turns this into an S-expr, zero upload. Here you can feel the deterministic gates that run after generation. Change 300000 to 3000000 and the gates catch it.)
↓ The S-expr Gemma4 would generate (locally in production; editable here)
3. Author an "atypical" rule on the spot(make a person's tacit judgment explicit as an S-expr. Deploy-free, instant, version changes)
4. Audit log(each judgment hash-chained with version + author = tamper-evident, replayable)
#
proposal
verdict
rule ver / author
hash(rec+prev)
Invariants (bit-exact)
・The clamp layer is integers/rationals/symbols only. A float proposal is folded to a rational at the boundary before evaluation (determinism starts at the folded boundary).
・Same input + same rule version → byte-identical verdict.
・Flexibility is the "author a rule" step; determinism is the "evaluate under a pinned version" step. Pinning the version makes both hold.
・Verify: open DevTools → Network — clamping sends zero requests (only static code is fetched). Note: the clamp only clamps what you wrote. It does not guarantee the rule itself is correct (person-bias is reviewable/rollback-able via version-pin + author).