Skip to content

About

proof-radar — validity cartography for informal mathematics: per-claim maps of where a result is proven, refuted, empirical, marginal, vacuous, or unprobed, with provenance and coverage.

Topics

Resources

Stars

1 star

Watchers

0 watching

Forks

Latest commit

 

History

1 Commit

Folders and files

Repository files navigation

proof-radar

Validity cartography for informal mathematics — per-claim maps of where a result is proven, refuted, empirically supported, marginal, vacuous, or unprobed, with provenance, margins, and coverage accounting.

House rule: a VERIFIED badge may not assert what the harness cannot rerun.

proof-radar is the middle rung of mathematical validation: adversarial verification at the rigor level papers are actually written — above unaided peer review (uninstrumented, unlogged), below full formalization (months per theorem). It is cartography, not policing: errors are found because boundaries are mapped, and the boundaries are the publishable object.

It is not formal verification, not a substitute for peer review or Lean/mathlib, and not a certainty machine. It emits graded confidence with documented failure modes, or it emits nothing.

Full specification: CHARTER.md.

Two deliverables per paper

Deliverable File What it is
The radar RADAR.md + radar.json the validity map — regions of hypothesis space graded {proven, refuted, empirical, marginal, vacuous, open}, each with provenance, boundary sharpness, margin gauges, coverage
The ledger VERIFICATION.md the audit trail — every check that produced a region, logged, graded, rerunnable

No radar region may exist without ledger rows behind it (Rule 5.1: provenance or nothing). Unprobed space renders as unprobed (Rule 5.2) — the verifier's own blind spots are a first-class layer, the verifier's ker P_C.

The founding incident

In review round three of Keep the Angle, an AI verifier certified a Corollary-1 summation argument as sound. It was not — the argument silently required the multiplicative-uniform form of a limit and a bounded degree ratio. The conclusion was true; the derivation was incomplete; the pass was a false pass on the derivation while correct on the conclusion. Two structural lessons follow, and they are load-bearing:

  1. Conclusion-checks and derivation-checks are different objects — graded on separate axes, always (Rule 4.1). A numerical confirmation of a true conclusion says nothing about the proof.
  2. An embedded verifier is not independent — same model, same context, co-author of the derivation ⇒ correlated errors. Grading runs fresh-context (§6).

That false pass is Baseline Datum #1 of the benchmark (benchmark/).

The checks (C1–C8)

Recomputation · special-case instantiation · numerical falsification harness · type/dimension/units lint · statement↔proof diff · cross-artifact consistency · citation verification · formalization escalation. See checks/. Every ledgered claim gets C1 + C5 at minimum.

Layout

Path What
CHARTER.md full charter (mission, scope, checks, gates, governance)
checks/ the C1–C8 check taxonomy as an operational checklist
schema/radar.schema.json JSON Schema for radar.json
templates/ RADAR.template.md, VERIFICATION.template.md
ledgers/ per-paper VERIFICATION.md ledgers (v0 scope: the trilogy)
benchmark/ the seeded-error benchmark (§8)
nrp/ NRP/Nautilus manifest conventions (§7)

Scope & status

In scope (v0): the Geometric Observation trilogy — Keep the Angle, TurboQuant-Pro, Observation Theory — and their companion artifacts. One VERIFICATION.md per paper. Expansion beyond the trilogy is gated on G2 (Rule 2.1: the project that verifies everything verifies nothing reproducibly).

Status: draft v0.1 — for adoption at G0. Open G0 parameters: final name; benchmark numeric bars; radar.json schema freeze; human-auditor roster; the LLM-verification prior-art scan.

Independence (why you can trust a grade)

The embedded assistant may draft radars and run harnesses; it may not be the sole grader of anything it helped derive (R-IND-5). Grading passes run in fresh context (R-IND-2), cross-model where available (R-IND-3), with ≥10% human spot-audit (R-IND-4). Every ledger row names its verifier (model, version, context freshness).

Governance

PI: A. H. Bond. Stand-down clause: if the gate (G2) fails twice, the project stands down and the postmortem is published — the value was in the adversarial relationship, not the artifact. This clause is the project's own falsifiability.

Licensing

Code Apache-2.0 (LICENSE); documents CC-BY-4.0 (LICENSE-docs); benchmark ground-truth labels sealed until G3.

About

proof-radar — validity cartography for informal mathematics: per-claim maps of where a result is proven, refuted, empirical, marginal, vacuous, or unprobed, with provenance and coverage.

Topics

Resources

Stars

1 star

Watchers

0 watching

Forks

Releases

Packages

Contributors