The Blackhole lab: when information is present but cannot be read
Information can be fully present and still unreadable. When does that happen, why, and what does reading it cost? Black-hole information theory is a well-developed physical treatment of that question. The lab studies them as physics, and uses them as a source of tools for keeping collectives of AI agents auditable, which is its long-range aim.
A working frame
The lab works with three axes, as a working hypothesis rather than a theorem. Location: is the information held in the parts of a system, or only in correlations between parts? Capacity: how much can move, limited by light cones in physics, by min-cuts (the narrowest set of links that separates one part of a network from another) in networks, and by bandwidth in agents. Mismatch: how well the reader’s model of the system fits the system.
Bounds that hold in one direction make the frame usable: each says what a reader cannot know, given a stated limit. The deepest open problem is that no law yet joins the three axes.
Scale of the record
- 232 automatic reproduction checks, all passing; the full test suite has 675 tests (2026-10-13).
- 18 pre-registered hypotheses, each with a file that records its prediction before the data.
- 2 real gravitational-wave events analysed from public LIGO data.
- 1 run on quantum hardware (IBM ibm_fez), with a second round pending.
How to read the claims
Claims on these pages carry two tags where the lab’s record assigns them. Confidence: proved, verified (an independent method or a reproduction), observed (one pre-registered method) or suggestive, and pending where nothing has been measured yet. Novelty: known (published by others first), narrow (the core is known and a stated part is the lab’s) or ours. A test that can fail in one direction only certifies its violations; when such a test passes, these pages say it passed, never that something was certified.



