Project Fortress is a patent-pending compute engine that answers a single question:
Is this a photograph of real biological matter, or was it AI-generated?
It does not return a probability. It does not require a trained model or a cloud connection.
It runs a fixed mathematical analysis on the image and returns a number, the same number, every time, on any Apple Silicon machine. That number is cryptographically signed and bound to the source file.
Here is a real result.
Two images of the same biological subject, one genuine photograph, one AI-generated. Project Fortress processes both in under a second.
The genuine image scores 1246.98.
The AI-generated image scores 2191.49.
Real image — 1246.98
AI-generated image — 2191.49
Difference — 944.51 points
The direction of separation, real lower, fake higher, is consistent across all validated categories within the Maillard engine domain. The gap does not narrow under compression. In three of four tested categories, the gap widens under heavy JPEG compression, contradicting the classical prediction for pixel-statistics-based detectors.
The score measures the complexity of spatial frequency patterns in the image patterns that differ fundamentally between photographs of real physical matter and AI-generated imagery.
Run the same images again tomorrow, on a different machine. The scores do not change. Not approximately, exactly. To four decimal places, every time.
That is not a confidence percentage. There is no threshold to debate and no model to retrain.
The result is cryptographically signed and bound to the source file.
Anyone with access to the binary can verify it independently
Most AI detection systems return a confidence score, ie "87% likely to be AI-generated." That score changes depending on the model version, the training data, and the hardware it runs on. It can be contested, retrained, or gamed.
Project Fortress does not work that way. There is no model. There is no training data. There is no confidence interval. The engine applies a fixed mathematical analysis to the spatial frequency structure of the image and returns a deterministic result. The same image produces the same score every time, not approximately, exactly. To four decimal places, across different machines, different sessions, and different copies of the binary.
That reproducibility is the asset. It means the result can be logged, audited, and independently verified by any party with access to the binary. It is suitable for environments where a probabilistic confidence score is not sufficient.
Three specialist engines, each calibrated against a distinct category of biological matter using fixed mathematical weighting constants.
Maillard analyses animal protein, meat and fish.
Mycelium analyses fungal structures.
Chlorophyll analyses botanical and plant matter.
Each engine produces a numeric authenticity score, and every result is cryptographically bound to its source data via a unique Vault_Key
As AI-generated imagery becomes harder to distinguish from reality, authenticating a photograph of real biological material is becoming a genuine challenge, one that existing lab-based and chemical testing methods are slow and expensive to address.
Project Fortress offers a fast, image-based alternative
Each of the three Trinity engines has been tested through repeated, real-world comparisons: genuine photographs against AI-generated reproductions of the same subject, across multiple categories within each engine's domain
Testing has included deliberately diverse and challenging examples, varied specimens, lighting and composition, rather than easy or favourable cases
Results have shown consistent, repeatable separation between real and AI-generated images within each engine's domain, with output verified identical across Apple Silicon M3 and M4 architectures
Extended validation has confirmed compression robustness across three JPEG quality levels (20 of 21 correct) and cross-generator performance against Stable Diffusion, a held-out architecture not used during original calibration (13 of 14 correct)
An independent academic review has been agreed in principle with a leading UK research University.
Independent third-party validation is the next stage of this process.
Project Fortress is designed for deployment contexts where reproducibility and auditability matter more than probabilistic confidence.
Food fraud and delivery platform verification, AI-generated food imagery is being used to support fraudulent refund claims on major delivery platforms. Project Fortress provides fast, image-based triage that does not require laboratory testing or human review.
Defence and intelligence photographic verification of biological material in field or remote environments requires results that are reproducible, auditable, and independent of cloud infrastructure. Project Fortress runs airgapped with no external dependencies.
Regulatory and compliance environments, the EU AI Act requires high-risk AI systems to provide automatic logging and reproducible, auditable outputs from August 2026. Project Fortress is well-positioned to be relevant to this requirement by design.
New biological categories can be calibrated in 48 to 72 hours via a documented standard operating procedure.
The SHA-256 hashes published here constitute the cryptographic fingerprints of the frozen Project Fortress binaries
Cross-Silicon Parity / Apple M3 and M4 / SHA-256 Verified
Extensively tested across multiple biological categories, real and AI-generated comparison
Asset Status: Frozen and Ready for Due Diligence
Airgapped, statically compiled, no external dependencies
Any deviation from these signatures indicates modification from the verified sealed state
Technical documentation and verification evidence available under NDA
Project Fortress is available for technical due diligence under NDA.
Qualified enquiries receive access to the full Tier 1 data room immediately and Tier 2 documentation following NDA countersignature.
To request the data room manifest, white paper, or a non-confidential technical abstract, contact will@atmospherelabs.co.uk or use the form below.
Please include your organisation and the nature of your interest.