core/docs/workbench/EVALUATOR.md

CORE Workbench — Evaluator Quickstart

For a technical evaluator with ~10 minutes who wants to test the claims, not read the pitch. CORE is a deterministic cognitive engine; the Workbench is its audit-native UI. The thesis you're here to check: this engine refuses rather than guesses wrong, replays bit-for-bit, and its geometry can't fake coherence.

This document is written to be falsifiable — every claim below points at a surface you can poke or a file you can read. Verify, don't trust.

Run it (one command)

scripts/workbench            # preflight → install → start API + UI, health-gated

Prereqs (the preflight checks them and prints the exact fix if missing): uv, Node ≥ 20, pnpm (corepack enable pnpm). No Rust build required — the backend runs on the pure-Python numpy backend. It prints one URL (http://127.0.0.1:5173); Ctrl+C stops both servers. Everything binds to localhost only; the UI is read-only by construction.

The 10-minute path

Each stop says what it proves — and what it does not.

1. /tour — the determinism narrative. A provider-agnostic, ordered walk: the engine decides a claim, refuses a wrong one, and replays it to the same hash. Each card pulls its "what this proves / does not prove" text from the real demo spec — not marketing copy. Proves: the spine is honest by construction. Not: general capability.

2. /trace + Replay. Pick a turn; read its pipeline stages, grounding, epistemic state, and trace hash. Hit Replay — it re-executes in a sealed fresh runtime and shows hash == hash. Proves: determinism is real and verifiable, not asserted. Not: that the answer is correct — only that it's reproducible.

3. /evals — the wrong=0 ledger. The serving discipline made constant: N correct · N refused · 0 wrong. The zero is load-bearing. Proves: the engine declines rather than emits an unverified answer. Not: high coverage — refusal rate is high on purpose (see scope below).

4. /calibration — the gold-tether arena. Per class: coverage vs the one-sided Wilson floor, the θ ceiling, and a plain "earned PROPOSE / SERVE / neither" verdict. This is where the engine earns the right to guess. Proves: "guessing" is gated by a calibrated statistical floor, not vibes.

5. /logos — Alignment tab (the geometry). The tri-language resonance graph (Hebrew → Greek → English). Open grc_logos_cognition_v1: the Safety tab and edge list flag invalid alignment targets. That is not a bug — the tool is honestly surfacing a real, latent cross-pack data inconsistency that a less-honest UI would have hidden. The geometry telling the truth, live. Proves: the audit surface exposes its own substrate's gaps. Not: that the field is "reasoning" — see scope.

Honest scope (read before you judge capability)

CORE is deliberately narrow and says so. The Workbench is built to prevent overclaiming — let it.

• Verified flagship: propositional entailment. The strongest proven capability is sound + complete deductive entailment, checked against an independent decision procedure. Read CLAIMS.md and the deductive-logic eval lane for the exact, pinned numbers — don't take a README's word.
• GSM8K math is a diagnostic, not the headline. The serving reader refuses most problems rather than risk a wrong answer; that's the wrong=0 discipline, not a capability flex.
• The CL(4,1) field is a coherence-gate, not the reasoner. It localizes and coherence-gates, then hands off to symbolic structure. Anyone pitching it as "geometric reasoning" is overreaching; it is geometric coherence-checking.
• Holonomy proof cards are roadmap. The /logos Holonomy state reads missing_evidence by design — the tri-language "meaning survived the path" proof is authored but not yet computed end-to-end. It is honestly absent, not faked.

Verify, don't trust

• Determinism: Replay any turn; the trace hash must match. Re-run scripts/workbench — same inputs, same hashes.
• wrong=0 is gated, not hoped: scripts/verify_lane_shas.py pins the eval-lane SHAs; scripts/generate_claims.py --check guards CLAIMS.md.
• No hidden mutation: the UI is read-only (proposal mode: none — read-only in every footer); the only write paths are explicit, allowlisted, and human-ratified (/proposals).
• The architecture's invariants are tested: tests/test_architectural_invariants.py and the field invariant (versor_condition < 1e-6) fail loudly when violated.

Where to go deeper

• docs/workbench/ — design system, route map, mastery roadmap.
• CLAUDE.md — the engine's non-negotiable invariants and doctrine.
• docs/specs/runtime_contracts.md — response/telemetry/identity/replay contracts.