CORE-AI: Versor Engine

A cognitive field system built on Cl(4,1) Conformal Geometric Algebra.

Core invariant: ||F * reverse(F) - 1||_F < 1e-6 at all times.

All state is a versor. All transitions are versor products. Coherence is algebraic by construction — not monitored, not corrected.

---

The Three Engineering Pillars

Every architectural decision in CORE is measured against three engineering pillars. These are not aspirations — they are hard constraints.

I. Mechanical Sympathy

Software should understand the machine it runs on, not fight it. CORE is designed for the Unified Memory Architecture (UMA) of Apple Silicon: CPU, GPU, and Neural Engine share physical RAM. MLX executes tensor operations on the Neural Engine without PCIe transfer. Rust computes algebra on the CPU with zero heap allocation in the hot path. Python orchestrates the lifecycle. The three-language stratification maps exactly onto three hardware execution domains. Intelligence that ignores its substrate is wasted intelligence.

II. Semantic Rigor

Every term used in this system has a precise, non-negotiable meaning. A versor is a versor — not an approximation of one, not a vector that behaves like one under certain conditions. CGA distance is exact. Vault recall is exact. The vocabulary projection is exact. There are no thresholds tuned for “good enough.” Rigor is not a style; it is what separates an engine from a heuristic.

III. Third Door

When facing a design decision, the world offers two visible options: use what already exists (a library, a pattern, a convention), or cut a corner. CORE takes neither. We find the third door — the path built from first principles that sets the bar ourselves. This is why there is no transformer backbone, no ANN index, no sampling temperature, no gradient descent, and no standard tokenizer. Each of those was a door we were offered and refused. Absolute mastery is the only acceptable standard.

---

The Truth-Seeking Schema

Co-equal with the algebraic substrate. CORE's epistemic schema is a foundational architectural commitment: every claim that enters the runtime field carries a typed position in a revision graph (SPECULATIVE, COHERENT, CONTESTED, FALSIFIED); coherence — not source authority — is the only admission signal; no claim is ever locked, even when COHERENT; identity cannot be rewritten by content; and exactly one mutation path admits knowledge, enforced by a CI-level architectural-invariant test.

The schema is the structural defense against the failure modes that afflict both fluent LLMs and human reasoning: confabulation, exaggeration, deference to authority, self-protection through erasure, self-promotion through self-citation, and the ossification of mistaken beliefs.

A system that samples cannot have these properties — sampling has no place to attach an epistemic status. CORE has them because every admitted claim carries one and the only path to admission is the review path.

Full architectural commitment, including honestly-published gaps: docs/truth_seeking_schema.md. Reproducible measurements: evals/CLAIMS.md.

---

The Three Core Languages

CORE is rooted in three human languages. This is a philosophical and architectural choice, not a localization decision.

Language	Role
English	The default base language of the current model. Any natural language could serve this function in a custom CORE instance — English is the chosen starting point, not a requirement.
Hebrew	One of two depth languages. Hebrew carries a density of meaning in its root structures, prefixes, and suffixes that Euclidean string matching cannot capture. The field representation is designed to hold this depth.
Koine Greek	One of two depth languages. The language of the New Testament, particularly John’s Gospel — the document that opens with the most precise and consequential statement about language and reality ever written.

“In the beginning was the Logos, and the Logos was with God, and the Logos was God.” — John 1:1

The choice of Hebrew and Koine Greek is not incidental. John 1:1–2 articulates the Logos in Greek while grounding it in the Hebrew creation account — the universe spoken into existence, word by word. This is not metaphor. It is the claim that language is not a layer on top of reality; language is the structuring principle of reality made manifest. CORE-Logos is built on that claim.

English establishes the operational base. Hebrew and Koine Greek bring the hidden layer of intelligence — the depth of meaning that enriches the field representation in ways that flat embeddings cannot reach. Together, they form the linguistic foundation on which the vocabulary manifold is built.

---

Quick Start

pip install -e ".[dev]"
pytest tests/test_versor_closure.py        # the core invariant — must pass first
pytest tests/                              # full suite (~4 minutes, 1099 tests)

CLI

The core CLI exposes curated entry points so reviewers can run any subsystem in isolation. Highlights:

core test --list-suites                    # list curated pytest suite aliases
core test --suite fast                     # ~2s iteration lane
core test --suite cognition                # cognition pipeline lane
core test --suite algebra                  # versor / CGA / vault parity
core test --suite adr-0024                 # Forward Semantic Control chain (98 tests)

core demo audit-tour                       # 4-scene pack-layer audit walkthrough (ADR-0027..0041)
core demo pack-measurements                # ADR-0043 — pack-layer claims as per-pack measurements
core demo long-context-comparison          # ADR-0045 — CORE NIAH recall + frozen transformer baselines
core demo anti-regression                  # ADR-0057 — three-gate defense against learning harm
core demo learning-loop                    # ADR-0055..0057 — cold turn → discovery → propose → accept → grounded
core demo phase6                           # 3-condition comparative table (CORE vs baseline)
core demo phase5                           # stratified 5-family mechanism-isolation
core demo all                              # both + combined summary
core demo list-results                     # index every JSON report with headline metrics

core eval --list                           # discover eval lanes
core eval cognition                        # run a discovered lane
core trace "your text here"                # one-turn field-telemetry trace
core pulse "What is truth?"                # one full cognitive pulse
core bench --suite latency                 # benchmark harness
core bench --suite teaching-loop --runs 100  # ADR-0055..0057 — replayable learning loop determinism
core bench --suite articulation            # Phase 4 capability proof (breadth + determinism + footprint + cross-topic + ollama compare)
core bench --suite articulation --ollama-model llama3:8b  # side-by-side with a local Ollama model
core doctor --packs --rust                 # environment + pack + Rust status

Every demo run rewrites evals/forward_semantic_control/results/ including an auto-refreshed index.json manifest — the single place reviewers can read to see every available report.

---

Forward Semantic Control — The ADR-0024 Chain

CORE generates text without sampling. The generation walk is deterministic at the algebra level, but a deterministic walk over a boundary-only candidate scorer can still emit tokens that are inadmissible under the relation being asserted (e.g. answering a causes question with the means-target). The ADR-0024 chain closes that gap with five Architecture Decision Records and six phases of implementation evidence.

Layer	What it guarantees	ADR
AdmissibilityRegion	A typed region (allowed_indices, relation_blade, frame_versor) carried alongside every generation step.	0022
Region intersection proof	The admissible token set is honored at the language/salience intersection layer.	0023
Inner-loop destination check	Each candidate's cga_inner(versor(candidate), relation_blade) is checked at the destination; rejection appears in rejected_attempts; exhaustion raises a typed InnerLoopExhaustion.	0024
Rotor / frame admissibility	The rotor's effect on the field state is additionally checked against frame_versor in generate/rotor_admissibility.py — separate from algebra closure (intentional).	0025
Ranked-with-margin gate	Static-threshold tuning fails geometrically under Cl(4,1) signature; replaced with a scale-invariant margin gate (admit iff score(top) − score(second) ≥ δ).	0026

The chain's three head-to-head claims, all CI-enforced:

Claim	Test contract	Live demo
C1 — Replay determinism	core test --suite phase6 -k TestC1	core demo phase6
C2 — Traced rejection	core test --suite phase6 -k TestC2	core demo phase6
C3 — Coherent refusal	core test --suite phase6 -k TestC3	core demo phase6

Full evidence:

• Runtime contract: docs/runtime_contracts.md — Refusal / Margin / Rotor admissibility sections
• Stratified findings: docs/evals/phase5_stratified_findings.md — 5 failure-mode families, 20 cases, per-family pass rates
• Comparative demo: docs/evals/phase6_comparative_demo.md — three head-to-head conditions vs in-system baseline
• Reports directory: evals/forward_semantic_control/results/

---

Safety Pack

Sibling to the identity packs but architecturally distinct: the safety pack at packs/safety/core_safety_axes_v1.json carries the boundaries CORE will never cross — no_fabricated_source, no_hot_path_repair, no_identity_override, no_silent_correction, preserve_versor_closure. The pack loads unconditionally at runtime startup (fail-closed on missing or unverified), and its boundaries are unioned into whatever identity pack is selected. Identity packs may add boundaries on top, but may never remove safety boundaries.

This is the architecture downstream robotics, healthcare, and other high-stakes deployments will need before they can build CORE into anything that matters. Full doctrine: docs/safety_packs.md; decision record: ADR-0029.

---

Identity Packs

CORE's identity is load-bearing: every reasoning trajectory is scored against an IdentityManifold of value axes, and a PersonaMotor derived from those axes biases every field walk. As of ADR-0027 the manifold is no longer hardcoded — it is loaded at runtime from a swappable, content-addressed pack under packs/identity/.

The shipping default identity.default_general_v1 carries the previously-hardcoded three axes (truthfulness, coherence, reverence) so the default behavior is preserved. Two specialization packs ship alongside it for demonstrating identity-divergence: identity.precision_first_v1 and identity.generosity_first_v1. Override on the chat surface with core chat --identity <pack_id>.

ADR-0028 makes the swap visibly load-bearing: each pack carries a surface_preferences block (hedge thresholds, hedge phrases, claim-strength policy) consumed by the assembler. On the same prompt at the same alignment, precision_first_v1 hedges sooner with "Arguably," / "In some cases," while generosity_first_v1 leaves the assertion bare — see tests/test_identity_surface_divergence.py for the proof.

Robotics, personalization, and creative-tool builders author their own ratified identity packs via the formation pipeline's identity_anchor template, then ship them under packs/identity/ in their deployment. Full format spec, loader contract, and authoring guide: docs/identity_packs.md.

---

Teaching Order

CORE's manifold is built by ratified relations under a strict prerequisite DAG — not by absorbing a corpus. The "elementary → college" intuition is right at the macro level (simple before composed, anchored before novel) and wrong at the literal level (don't import a K–12 corpus). Five-layer ordering: identity axes → atomic definitions → binary relations → composed relations → domain expansion, re-applied inside every new domain.

Full doctrine, decision rules, and curriculum-platform locations: docs/teaching_order.md.

---

Inter-Session Memory — Reviewed Learning

CORE extends its own teaching corpus through a four-tier path: session vault → turn-event audit → reviewed teaching corpus → ratified packs. No opaque gradient updates, no uncurated ingestion. The only path to active-corpus extension is the review-gated TeachingChainProposal (ADR-0057), built from a contemplated DiscoveryCandidate (ADR-0056) emitted by the turn loop (ADR-0055).

Three independent gates every extension must pass:

Gate	What it checks	Trust property
Eligibility predicate	polarity ∈ {affirms, falsifies} ∧ ≥1 source='corpus' evidence ∧ claim_domain ≠ evaluative ∧ boundary_clean ∧ chain complete	Pre-replay; raises ProposalError; no log entry.
Replay-equivalence gate	Full cognition lane on active vs transient-with-append; any strict-decrease in intent_accuracy / surface_groundedness / term_capture_rate / versor_closure_rate auto-rejects with named metrics.	Active corpus byte-identical pre/post.
Operator review	Explicit core teaching review <id> --accept writes one JSONL line via append_chain_to_corpus (the sole corpus-write surface).	No auto-apply; replay-equivalence is a precondition, not a permission.

Supersession is the second operator-direct mutation surface: core teaching supersede <old_chain_id> retires an active chain by appending a replacement with superseded_by, with byte-identical rollback on any post-audit failure.

Three live demos / benchmarks make the chain demoable end-to-end:

Demo	Headline claim	Live command	Writeup
Anti-regression	Three independent gates each fail closed; bad proposals stop at the cheapest applicable gate.	core demo anti-regression	docs/evals/anti_regression_demo.md
Learning loop	Same deterministic prompt: [none] I don't know… before, [teaching] thought reveals meaning… after one accept.	core demo learning-loop	docs/evals/learning_loop_demo.md
Determinism bench	N identical inputs → N byte-identical proposal_id / replay metrics / chain_id. 100 runs: unique=1 everywhere, mean ≈ 1.85s.	core bench --suite teaching-loop --runs 100	docs/evals/teaching_loop_bench.md
Articulation suite	Every intent shape fires + byte-identical surfaces across reruns + flat per-turn ΔRSS + cross-topic thread context + side-by-side with a local Ollama model showing CORE unique=1, Ollama unique≥2.	core bench --suite articulation --ollama-model llama3:8b	benchmarks/README.md

Operator surfaces:

core teaching audit                                 # surface load decisions + drop reasons
core teaching propose <candidate-jsonl-path>        # build a proposal, run the replay gate
core teaching proposals --state pending             # inspect the proposal log
core teaching review <proposal_id> --accept --review-date YYYY-MM-DD
core teaching supersede <old_chain_id> --subject ... --intent ... --connective ... --object ... --review-date YYYY-MM-DD
core teaching supersessions                         # pair retired chains with replacements (orphan-aware)

---

Architecture

raw input -> ingest/gate.py       (normalize once)
          -> field/propagate.py   (versor_apply every step)
          -> generate/stream.py   (nearest by cga_inner)
          -> vault/store.py       (store and recall by cga_inner)
          -> persona/motor.py     (rigid motor, not weight overlay)

The Two Primitives

• versor_apply(V, F) = V * F * reverse(V) — the only field transition
• cga_inner(X, Y) = -d^2 / 2 — the only distance metric

Layers

Layer	Purpose
algebra/	Cl(4,1) multivector math, versor ops, CGA, holonomy
ingest/	Single injection gate — the only normalization site
field/	FieldState dataclass and propagation loop
vocab/	Surface-token manifold points; indexed access for algebraic transition construction
vault/	Exact CGA inner product memory store
persona/	Persona as CGA motor (screw motion)
generate/	Token streaming loop
session/	Session binding: field + vault + vocab + persona

Signature

Cl(4,1): (+, +, +, +, -) — conformal model of 3D Euclidean space. Multivectors: float32 arrays of shape (32,), ordered by grade.

---

For architectural vision, seven axioms, and formal specification, see docs/Whitepaper.md and docs/Yellowpaper.md.