STANDALONE EXPLORATION · NOT WIRED TO THE TANTULAR PIPELINE

Constrained Gödel-Agent

A self-rewriting loop, sandboxed to one function. It inspects its own policy, scores it against a held eval set, proposes a rewrite, and keeps it only if the score does not regress.
SANDBOXED · CONSTRAINED SELF-REWRITING
1 · SELF_INSPECT
Represent the current policy, score, and history as an inspectable state.
2 · INTERACT
Evaluate the policy against the eval set, producing a score and feedback on the weakest cases.
3 · SELF_UPDATE
Validate a candidate rewrite, monkey-patch it in, re-evaluate, and keep it only if it does not regress.
4 · CONTINUE_IMPROVE
Recursively request the next round of actions, up to a fixed max depth.
CURRENT POLICY
{{ versionLabel }}
{{ scorePct }}%
combined_score, 8 held cases
{{ versionSummary }}
{{ code }}
TASK ENVIRONMENT — 8 HELD CASES, INDONESIAN OFFICE / PUBLIC-SERVICE SUPPORT
{{ row.label }}
{{ row.passMark }}
“{{ row.query }}”
{{ chip.term }} ≠ {{ chip.term }}
{{ row.categoryLabel }} · weight {{ row.weight }}
CONTROLS
Converged at 100% — every held case passes.
A candidate that shortened the fraud-safety branch scored lower and was rolled back automatically. The kept policy is still V4.
HISTORY, MOST RECENT FIRST
round {{ h.round }} · {{ h.version }} {{ h.before }}% → {{ h.after }}%
{{ h.rationale }}
{{ h.verdict }}
No rounds run yet. Click “Run round 1 of 4” to start.
Try the safety gate yourself
SafePolicyLoader — AST validation, simulated
Every candidate is checked before it is ever executed. Edit the code below, or paste your own, and validate it. It must define exactly one function, solve(query, kb), with no imports, no dunder access, and none of the banned builtins.
{{ validationHeadline }}
{{ validationReason }}
import / from…import class / lambda / global / nonlocal __dunder__ access eval / exec / compile / open / getattr / setattr / globals
CURRENT RECIPE — ghifidanukusumo/tantular:latest profile
{{ recipeLabel }}
{{ recipePct }}%
visible gate, 5 real failures
{{ recipeRationale }}
temperature {{ samplingTemp }} presence_penalty {{ samplingPP }} top_k {{ samplingTopK }} 
{{ addendumDisplay }}
VISIBLE GATE — 5 REAL FAILURES FROM SERVING_LATENCY_FINDING.md
{{ row.label }}
{{ row.passMark }}
{{ row.desc }}
{{ row.id }}
PRIVATE HOLDOUT — NOT SHOWN TO THE MUTATOR
{{ holdoutScore }}
{{ row.desc }}
{{ row.passMark }}
Two more deny-listed terms (setting, user), never scored during tuning — the same guardrail as HoldoutDatasetSupportEnvironment. Reveal it once the visible gate converges.
CONTROLS
Dropping the JSON-only clause reintroduced the malformed-output failure. Rolled back automatically.
HISTORY, MOST RECENT FIRST
round {{ h.round }} · {{ h.label }} {{ h.before }} → {{ h.after }}
{{ h.rationale }}
{{ h.verdict }}
No rounds run yet.
A prototype of the pattern in the uploaded Gödel-Agent design document: self-inspection, a scored task environment, constrained self-rewriting, and rollback on regression. Both tabs compute real, deterministic scores in your browser; nothing is sent anywhere.