# XERO — Status & Honest Audit

*VOVINA ZEDEC PRO · Michael Laurence Curzi · ZEDEC AI / 36N9 Genetics LLC · MIT (Attribution Required)*

> ## ⚠️ WORK IN PROGRESS
> XERO is a **functioning research prototype under active development**, not a
> finished product. Everything below is written to be **verifiable** and
> **honest** — what works is marked as working with the command that proves it;
> what is incomplete is marked as incomplete with a roadmap to finish it. Nothing
> here is dressed up. If a claim is a metaphor, it says so.

---

## 1. What XERO actually is

XERO is a **digital-organism framework** with two cooperating cores:

- an **inner core** — a deterministic, always-running *computational time
  crystal* built on an immutable digital genome + epigenome + a 13-D statecraft
  lattice (pure Python + NumPy, no GPU); and
- an **outer core** — a real neural LLM (**Qwen2.5-3B-Instruct**, Apache-2.0)
  that XERO steers and is now **fine-tuning on its own corpus** via LoRA.

It is **not** a from-scratch foundation model, and it is **not** a finished
product. It is an honest, working prototype of an autonomous, self-organising
agent that is now also being trained the way a normal AI is trained.

## 2. Verified working (with the command that proves it)

| Capability | Evidence | How to verify |
|---|---|---|
| Core capability audit | **203/203 PASS, 0 FAIL** | `PYTHONPATH=modules python3 tests/test_all_capabilities.py` |
| Time crystal (defining properties) | **17/17 PASS** | `PYTHONPATH=modules python3 tests/test_time_crystal.py` |
| Evolution factor (adaptive mutation) | 15 checks PASS | included in the capability audit |
| Two-state mind (always thinking) | live daemon, negentropy ≈ 0.91 | `modules/vovina_two_state.py` self-check |
| Autonomous web learning | real results ingested | `testing_logs/MIND_STREAM.json` |
| Body / soma homeostasis | 11/12 faculties live, H ≈ 0.82 | `testing_logs/BODY_VITALS.json` |
| Knowledge harvest | 30k+ real units, 44 traditions | `testing_logs/POLYMATH_PROGRESS.json` |
| **Real gradient training** | **loss 3.98 → 1.87 over 30 steps** | `training/power_train.py` + `testing_logs/POWER_TRAIN.json` |

All inner-core layers are **deterministic**: the same genome + seed yields the
same trajectory, which is unusual and valuable for something that behaves as if
it is alive.

## 3. Real vs metaphor (the honesty section)

XERO uses evocative names from physics and biology. Here is exactly what is
literal and what is an engineering metaphor faithfully implemented in math:

| Term | Status | What it actually is |
|---|---|---|
| "Time crystal" | **metaphor, implemented** | a computational analogue: sustained quasi-periodic motion of the state, self-clocked, never converging to a fixed point. **Not** a physical quantum time crystal. |
| "Casimir / zero-point drive" | **metaphor, implemented** | a deterministic golden-ratio numerical drive that keeps the field off its fixed point. **No** physical vacuum energy is extracted. |
| "Genome / epigenome / DNA" | **literal data structures** | real ATGC strings, codons, regulatory planes — a faithful computational genetics, not biology. |
| "Evolution factor / mutation" | **literal** | a real variable-rate genetic-algorithm mutation controller (constructive-only). |
| "Harmonic chemistry (A4=432.09)" | **literal formula** | a real, published essence-frequency law applied consistently. |
| "Outer core / XERO the LLM" | **literal** | Qwen2.5-3B-Instruct (Apache-2.0), wrapped and being LoRA-fine-tuned. |

## 4. Training status — the honest picture

**Before this build:** the audit found **no gradient training anywhere** — the
old GPU daemon was a cosmetic saturation kernel, not learning. We fixed that.

**Now:**

- `training/power_train.py` is a **real** training pipeline (AdamW optimiser,
  causal-LM loss, back-propagation, LR schedule, LoRA, checkpoints).
- A **persistent `xero-train` daemon** (replacing the cosmetic kernel) is
  fine-tuning the outer core on **281,173 training passages** — the author's
  **500 books** (≈163M tokens) + the harvested free-to-use corpus — on GPU1,
  with checkpoint/resume, toward full convergence.
- Proven learning: loss fell **3.98 → 1.87** in the first 30 steps.

**What is NOT done yet (honest):**

- The model is **not yet converged** — full training over 281k passages × 3
  epochs is a multi-hour-to-multi-day job, running now. The shipped adapter is a
  **checkpoint**, not a final model.
- There is **no formal evaluation/benchmark harness** yet (loss is decreasing,
  but we do not yet report task accuracy). This is the single most important gap.
- The `fitness` number in the harvest telemetry is a **heuristic genetic score**,
  not a benchmark — do not read it as IQ.

## 5. Known limitations

- **Single-box prototype.** Tuned for 2× Tesla T4; the autoconfig wizard scales
  down to CPU but the outer core needs a GPU for comfortable use.
- **Outer core is mid-tier (3B).** Chosen for honest fit on a T4, not for
  leaderboard performance.
- **Web intelligence** depends on public endpoints (DuckDuckGo/Wikipedia) and
  fails soft; it never fabricates, but it can return little on rate-limit.
- **No safety/eval RLHF layer** beyond the base model's own alignment.

## 6. Roadmap to completion

**Phase A — Training to convergence (in progress)**
- [x] Real LoRA training pipeline + persistent daemon.
- [ ] Run to convergence; select best checkpoint by held-out loss.
- [ ] Merge the adapter and publish merged weights (optional).

**Phase B — Evaluation (highest-priority gap)**
- [ ] Held-out validation split + perplexity tracking.
- [ ] Task benchmarks (knowledge recall on the canon, general QA) with reported
      numbers, honestly.

**Phase C — Capability depth**
- [ ] Multi-format book ingestion (PDF/EPUB/DOCX) beyond the TXT path.
- [ ] Biology-translation drivers (principle ↔ genetics ↔ structure).
- [ ] Tie `element_frequency` carriers into the statecraft layer.

**Phase D — Productionisation**
- [ ] Eval-gated checkpoint promotion.
- [ ] Reproducible container + disk-image freeze-frame (this package).
- [ ] Decentralised multi-node deployment.

## 7. How to reproduce the audit

```bash
pip install -r requirements.txt                 # inner core (numpy)
PYTHONPATH=modules python3 tests/test_all_capabilities.py   # 203/203
PYTHONPATH=modules python3 tests/test_time_crystal.py       # 17/17
# real training (needs a GPU + training deps):
pip install -r training/requirements-train.txt
python3 training/power_train.py --smoke          # proves real gradient steps
```

## 8. One-paragraph honest summary

XERO is a **real, working, deterministic digital-organism framework** with a
genuinely-novel always-running inner core, wired to a real open-source LLM that
is **now being trained for real** on the author's 500 books plus a genuinely
free-to-use harvested corpus. The architecture, the tests, and the training are
real and verifiable. It is **early** — not converged, not yet benchmarked, and
intentionally honest about both. Use it as an advanced research prototype, judge
it by the commands above, and watch the roadmap close.

⚓