Memory should preserve events, not bury them in context.
Systems need a time substrate: what happened, what changed, and what evidence survived contact with reality.
AI needs receipts, not vibes.
Probably AGI is Colby Fox's public trail for the evidence layer AI needs next: memory over time, inference courts, self-audit, and authority that has to be earned.
This is not a lab, a prophecy, or another AGI theater deck. It is one builder publishing enough proof to be judged without dumping the engine room.
Not a lab. One builder's public trail.
Leadership in AI will not come from louder predictions.
It will come from building systems that can prove what happened.
Leadership thesis
AI does not need another prophet.
It needs infrastructure that remembers, tests, audits itself, and refuses authority until the receipts are good enough.
Inference claims need courts, not applause.
Model behavior should face witnesses, controls, contamination checks, and no-authority verdicts before anyone believes the win.
Authority must be earned at the edge.
The default state is locked. Receipts move a system to the next gate; vibes do not.
Public trail
The work has three public names.
These are the visible surfaces of a larger build: memory, inference evidence, and action trails. The public trail shows the shape. The engine room stays bounded.
FDAG
Memory with evidence over time.
Preserve what happened instead of stuffing more history into context.
Explore FDAGFoxEIS
Courts for what happened during inference.
Evidence, controls, and no-authority verdicts separate signal from theater.
Read the explainerYouDAG / PX
Action trails and code understanding.
Work should leave inspectable traces across time, repos, and decisions.
Learn moreBoundary
We publish evidence, not reproduction paths.
Serious people should be able to judge the work. They do not need prompt packs, exact coordinates, kernel details, or a shortcut manual.
Show
- Redacted methods
- Selected logs
- No-authority receipts
- Public-safe artifacts
Do not dump
- Exact coordinates
- Prompt packs
- Kernel details
- Reproduction paths
Evidence drop queued
A Machine That Refuses to Promote Itself
FoxEIS ran a preregistered held-out inference court, found a provisional predictor signal, and still granted no authority.
- 69 / 69
- Held-out cases scored
- 69 / 69
- Controls and router witnesses passed
- 33 / 69
- Clean token/text effects
- none
- Authority granted
Launch 3 in the public sequence: after the founder thesis and FoxEIS explainer.
Start here
Read the trail in order.
Start with the founder thesis, then move into the artifact and boundary pages.
Private demo
For serious people building AI systems.
If you work on AI infrastructure, local inference, evaluation, interpretability, memory, or human-AI collaboration, I can show the private version.