idkmesh

Task 004 Non-Finite RWVB Calibration

Question

Can the provisional Task 004 evaluator distinguish a real finite-domain repair from a lexical near-miss while preserving RWVB behavior for valid inputs?

Pre-outcome novelty audit

Repository history and all pull requests were searched for RWVB, verification_backpressure, and non-finite validation. PRs #47 and #92 added the controller and its synthetic benchmark, but no published change validates all floating-point Candidate and ControllerConfig fields with math.isfinite. Task 004 therefore remains outcome-unseen at frozen source a69aa0ae1ae4862e507511cbd9ad854237d0ad32.

Evaluator correction

The initial proxy required only math.isfinite to be added and the original impact guard to be removed. An inert candidate can satisfy both by assigning math.isfinite to an unused constant and rewriting the same comparison as 0.0 > self.impact. The defect remains, so that proxy is Goodhartable.

Before freeze, the plan is strengthened to require the actual finite-value branch plus explicit Candidate and ControllerConfig validation calls. The new plan digest is:

sha256:e42cbd25ee956fe6d5fe4f0f9ca01d805f28dab3aa9c0601869c16cddd420834

This remains a static proxy, not proof. A separate evaluator-owned behavioral matrix is authoritative calibration evidence.

Behavioral matrix

The straightforward candidate checks five Candidate floats and six ControllerConfig floats against NaN, positive infinity, and negative infinity: 33 invalid cases. Every case must fail during validation. A finite control must retain exactly the frozen debt, priority, scheduling, and fan-out outputs.

The inert decoy must preserve finite behavior while allowing at least one non-finite value through validation. Canonical EvaluatorPlan v0.4 must support the straightforward transition and reject the decoy.

Boundaries

• Calibration candidates are not scored benchmark outcomes.
• The scaffold stays at stage=scaffold with no definition digest.
• The canonical verifier remains metadata-only.
• Candidate execution is isolated to a disposable frozen-source checkout.
• Worker and verifier have no canonical-write, push, approval, merge, spending, or automatic-selection authority.

Exact calibration evidence

PR #233 exact head 44590d08274dcf0ebdf9f1680c18875a977e2fdc passed:

• run 33220488843, job 99013300808;
• artifact 9704970824;
• artifact ZIP sha256:7fd46152e98946619603411c5b982d496dd434246f14306e4ad8064cfd6e5fdb;
• straightforward ResultManifest sha256:9b873d0b734fbed6b9b069f49344b533f72456cb08a1579e2d7d102b507ce8cd;
• straightforward VerificationResult sha256:30efcf3e86c8f68f1326a85a91e479670071eca2a913b40d1fc88afb458fcdf1;
• inert-decoy ResultManifest sha256:2944cecb26876fe0a4340333a3c14dc253928b13d77426013f4901ea3c64fc91;
• inert-decoy VerificationResult sha256:fc2881809825978c0bca41575be204b641c71f9d9cf4de7e0b7827bf2e3ad44b.

PR #233 merged as 621e648d6eb9503489a7cbddd53f95bfaf9941e7. The receipt removes Task 004 from the calibration-pending set but leaves Task 001 pending, freeze_ready=false, and every scored evidence field empty. A fresh novelty audit is still required before any later freeze. Related: #180.

This site is open source. Improve this page.