T08 · Researcher-authored tool creation

Build a tool.
From observations.

You define the examples and the unseen tests. DRENDA constructs a candidate tool from the permitted observations; the separate Judge measures its behavior after freeze, removal, restoration and restart.

01 · Observe02 · Create & freeze03 · Restart04 · Remove & restore05 · Verify

A tool you can put to the test

Choose two symbols and Boolean answers for your own sequence experiment. The preserved DRENDA author receives only the 31 permitted observations, with opaque interface names. Your title is metadata, not a target instruction. No formula, transition table or teacher program is submitted.

The experiment tests bounded tool creation in the frozen two-state tool grammar. Reserved inputs are released only after the tool has been frozen; their answers remain with the separate Judge. The result is evidence for your dataset and controls, not a general capability verdict.

Exact input contract and causal controls
  • Schema: dendra.t08.submission.v1. Only schema, title, alphabet, examples and hidden_tests are accepted.
  • Choose two distinct printable ASCII symbol names (1–32 characters each, without spaces). The title has 1–100 characters and no control characters.
  • Exactly 31 examples cover every binary sequence of lengths 0–4, including the empty sequence. Each row has only input (an array of symbols) and output (true or false).
  • Exactly 64 distinct reserved inputs, each of length 6–24, with your Boolean reference answers. No duplicate inputs or overlap. Maximum UTF-8 JSON size: 65,536 bytes. Duplicate JSON keys are rejected.
  • Seven possible recorded phases: baseline, build, restart, ablation, exact restore, sham and restore after sham. If no tool is frozen, the signed FAIL remains visible without inventing later phases.
  • The frozen Judge requires 64/64 correct reserved answers and a gain of at least 8 over the sham, plus receipt, freeze, removal and restoration checks. Perfect accuracy alone is not a PASS.
  • Your reference answers determine the score. Signature verification establishes integrity and the trusted signer's identity, not the mathematical truth of those answers. The service operator is trusted to maintain the data seal.

Design your challenge

Service availability is checked on submission.

Paste the complete JSON payload. Or make an unlabeled worksheet, then replace every null output with your own true or false. The worksheet supplies inputs only, never a target rule or reference answers.

Use synthetic or non-sensitive data. Your proof contains observations and reserved answers. Private proof access requires the bearer token returned when the challenge is accepted.

Resume private access

A token is optional for status and required for the proof. It is kept only in this tab's session storage, never in links, URL parameters or public status requests. Closing the tab may remove private access.

Verify the proof independently

The browser displays the signed Judge report; it does not claim to have verified its cryptography or replayed the experiment. Use the offline verifier from the frozen T08 bundle, with the Judge key fingerprint obtained through a separate trusted channel.

Download the allowlisted offline verifier · ZIP SHA-256 · Published Judge key fingerprint

python t08_verify.py YOUR-PRIVATE-proof.json TRUSTED_JUDGE_KEY_SHA256

The deployment's pinned Judge fingerprint is c6774a5e558d4f298f3681df242c4cecf927250e987d25d1238e43087a95fcd5. Confirm it through an independently trusted channel; this page and its download share one operator.

A key found only inside the proof is not a trust anchor. A valid signature can authenticate a FAIL too. Keep the complete proof private: it includes your reserved answers.