NeuroCognica

What You Get

For public version 2.0.46 (build serial 70, code-signed) · Last updated 2026-08-31

This is the honest specification of what ChronoSophia2 does today — not a roadmap, not a wish list. Where something is unproven or absent, it says so.

The evidence named in section 2 was witnessed on earlier builds, as each row states. The buyer-install witness was last attempted on internal build 1.4.38, where it blocked at the Windows UAC consent screen. It has not been repeated on any build since, including this one. It is not presented here as completed, and the version this document is pinned to is enforced by desktop/tests/test_customer_doc_claims.py so the claim cannot quietly outlive the build it was made about.

Buy on what is in section 2. Treat sections 3 and 4 as not yet delivered.


1. In the box

2. What is proven to work

These are capabilities the project has actually witnessed end-to-end, not merely wired up.

Capability Evidence
Create tab → staged hero render — type a prompt, get a primary hero PNG with a saved .blend and a sealed manifest Earlier showroom render + UI witnessed 2026-06-21 on v1.2.0; the installed buyer witness remains pending on the current build
Museum Wall and Showroom presentation — chosen by the Display dropdown Render-witnessed; deterministic and test-covered
Material intent enforcement — glass / metal / emissive / colour requests are honoured rather than silently ignored Live-verified (red glass sphere)
Shape guard — nonsense geometry (a goblet as a torus) is corrected before it reaches Blender Live-verified (goblet)
Editable output — every render produces a real .blend with procedural geometry, plus PNG and an .s3v.json manifest Open them in Blender 4.5 without Chronos
Local-first creative operation — no cloud inference, no account, no phone-home for the creative path Updates and optional dependency/model downloads use the configured vendor/feed URLs; prompts, renders, and ledger data are not sent as part of that path
Governance gate — the Sentinel authority + trajectory gate refuses unauthorised protected actions and seals the decision Tested at the API level, enforce-by-default
Harmful-use refusal — every creation request is checked against a published four-category prohibited-use policy before anything is generated, fails closed if the local checker is unavailable, and seals the decision to your ledger. You see it clear the work on every run, not only when it refuses Measured against the live classifier, both error directions, numbers and open findings published in the Sentinel security protocol (available on request; it is not shipped with the installer)
Your creative work is protected from the gate — fantasy violence, weapons, horror, satire, historical, medical and classical-nude subjects are defended by an explicit rule that overrides the classifier's own refusals Same measurement; the false-refusal rate is reported alongside the miss rate rather than hidden

3. The experimental product exposes three pages

ChronoSophia2 ships Create, Gallery, and Lineage.

Lineage is the petri dish, and it fills itself. Every render you finish becomes a cell in it, built from your own local hash-chained ledger. Nothing has to be judged first and nothing has to be added by hand — finishing a render is the lineage event. Create offers your most recent renders so the next one can be grown from one you already made, and what it grew from is sealed into the bundle alongside the original prompt, which is never overwritten.

Your rating is the steering. An overall star, and when something missed, per-axis marks for shape, material, lighting, style and detail, a "wrong thing entirely" flag, and a box for your own words. What you say is carried into the next attempt instead of being recorded and forgotten.

This is generation causality grounded in your own work. It is not image-model weight training, and it does not claim to be.

Changed in 2.0.33 and still true: Lineage previously opened an "Evolution Proof" chamber that was seeded with two built-in demo cells, and the Create picker read from that system rather than from your renders. Both are gone.

4. Not in this release

Do not buy expecting any of these. They are named because pretending they are imminent would be the same lie as shipping them broken:

Two known defects, stated plainly:

5. What you must supply

Requirement Detail
Windows 10/11 x64 Build 18362 or later
16 GB RAM 32 GB recommended
NVIDIA GPU Recommended. A machine with integrated graphics (Iris Xe class) can run Museum Wall and Showroom; a machine with no graphics device at all renders the hero view on the processor instead, more slowly. A measured Cradle Create was about 53 minutes; the canvas step alone has been about 17 minutes. Object mode runs on CPU for vessels and parts with a stated size — a wooden bowl took under a minute on an Iris Xe laptop. The carve inside Object mode (used when the subject is neither a vessel nor dimensioned) does run on CPU - a measured Cradle run was 38 minutes - but it needs three generated views instead of one, and for many subjects the result is not yet a likeness. Treat it as slow and experimental on CPU. Fine-tuning needs 12 GB+ VRAM
~30 GB free disk For the third-party dependency toolchain
Third-party tools Blender 4.5 LTS, ComfyUI, Ollama, Python 3.11, Git and PyTorch — the installer fetches these from their own vendors with your consent, roughly 8–10 GB
AI image checkpoints Not bundled in the installer binary. The setup task can, with your explicit checkbox consent, reuse verified matching files from an existing local ComfyUI model library or fetch the required files: Flux schnell on NVIDIA (~29 GB) or Z-Image-Turbo nvfp4 on CPU (~10 GB), both Apache-2.0. Recipes that need image models refuse to render rather than hand you a blank canvas. See MODEL_LICENSES.md
Patience with a first install The dependency setup is long. This is a local AI toolchain, not a 40 MB app

The one model that installs automatically is qwen2.5:7b-instruct (Apache-2.0), which turns your prompt into a validated scene plan.

6. Two things to know before buying

The installer is code-signed, and Windows shows a verified publisher. The certificate is issued to the individual behind NeuroCognica, so the prompt shows a person's name rather than the brand — expected. Verify the SHA-256 of your download anyway; see VERIFY_DOWNLOAD.md.

Unsigned builds were quarantined on our own machines as Trojan:Win32/Bearfoos.B!ml — an automated machine-learning verdict, not identified malware — on 2026-06-21 and 2026-08-16, and one was hard-blocked by Smart App Control on 2026-08-18, which offers no "run anyway" at all. A signed build was admitted under enforced Smart App Control on 2026-08-22. That is what signing bought, and it is why it was worth doing.

SmartScreen may still say the file "isn't commonly downloaded". That is reputation, not signature, and it fades as installs accumulate.

CPU image generation can take close to an hour. The Create tab uses a 90-minute no-image safety watchdog; it does not promise GPU-like latency.

FLUX.1-dev is not licensed for commercial use of the model. The product default is FLUX.1-schnell (Apache-2.0), which is fine for commercial work. If you switch to dev via CHRONOS_FLUX_VARIANT=dev, that is your licence decision to make with Black Forest Labs.

7. What makes this different

Not "AI generates 3D" — plenty of tools do a version of that. The distinctions:

  1. The output is editable. A real .blend with procedural geometry, not a mesh blob or an image.
  2. Everything is local. Your prompts and work never leave your machine. There is no service to lose access to.
  3. Every artifact carries provenance. Prompt, seed, plan, and policy decisions are sealed in a tamper-evident hash-chained ledger you can verify.
  4. Actions are gated before they execute. Out-of-plan commands are refused and the refusal is recorded, rather than trusted and hoped for.

If none of those matter to your work, this probably is not the tool for you, and we would rather you knew that before paying.

8. Why this document exists

Most software tells you what it might do. This tells you what it does.

Section 2 is the promise. If anything there does not work on hardware that meets section 5, that is a refund — see the Refund Policy, no argument required.

Sections 3 and 4 are the things a demo video would let us quietly skip. They are here because you deserve to decide with them in front of you.