The programme starts on 17 August, and we want you to start figuring out what you'd like to work on during the six days you'll be in Cambridge. Part of this is trying to understand your theory of change. What in the world do you want to affect, and how will you affect it?
For a six-day project your scope should be relatively tight. Having read some of the papers, you should have a sense of what kinds of experiments you'd like to try and the kinds of questions you want to research during your time. Be ambitious but mindful of the finiteness. Good luck!
How to use this unit2 readings + 1 video
Read both, watch the video, then write the sketch below.
OutputA completed project ideation worksheet
Make a copy of the worksheet, fill it in as you read, then share it with gaurav@caish.org and submit the link through the form. First version by Wednesday 12 August, final by Sunday 16 August.
READ 4.1Required20-25 min
You Need a Theory of Victory
Jason Hausenloy, The First Scattering (2026)
Jason Hausenloy separates a theory of change, your action X leads to good outcome Y, from a theory of victory, the full causal story of every step that has to go right, including the ones other people take. His tale of two verification teams contrasts one that burns years on a VC-funded custom security chip with another that ships off-the-shelf hardware taps built around what a US-China treaty would need.
Three mindsets for research on a clock. Truth-seeking (plausible results are usually wrong until you try to break them), prioritisation (a clear north star, checked daily), and moving fast (optimise for information gain per unit time, fail fast). Written for technical AI safety work, but broadly applicable.
Ideas and projects are heavy-tailed. The best project you pick can be far more impactful than your other options. Spending time picking between options is good, but timebox yourself to an hour or two of ideating. Favour projects that let you falsify beliefs quickly and run cheap experiments.
The board below covers Plan A's workstreams only, and the wider field is larger than this course. Aim for something you can scope to six days.
Plan A verification board
Where the inference-only retrofit stands
Plan A is one demanding use case for verification. It is a temporary training pause that still allows inference. Its proposed retrofit copies datacentre traffic, makes workloads reproducible, checks random samples by recomputation, protects the verifier, and closes routes for unverified outputs.
The last of those, making sure nothing leaves unobserved through leftover memory, side channels or covert links, is what the board calls completeness.
That system decomposes into 17 engineering workstreams. Each is marked by how much groundwork already exists, from active work you could build on to open problems you would start cold. This is not a ranking of importance.
Note, August 2026. Lucid have announced an eighteen-month adversarial testbed on an operating cluster, aimed at several of the workstreams marked open below, including recomputation red-teaming, verification reporting, tap installation and physical security. Worth reading if you take one of those on. Their work also covers ground this board does not, such as power-based workload verification, which is a reminder that the board tracks Plan A's retrofit rather than the whole field. The statuses below reflect demonstrated work rather than announced work.
Getting started
Choosing something to work on
Pick a workstream from the board to see where it stands. Most of these are bigger than six days, so if you take one on, choose a slice you could get a result on by Saturday. The five questions below help you scope it into a project.
01Claim
What must a lab, operator or state be able to claim?
02Evidence
What observable artefact could check the claim?
03Adversary
How could the evidence be evaded or tampered with?
04Test
What cheap result would change your view of feasibility?
05Handoff
Who uses the result, and what decision could it change?
4Active
At least one R&D effort exists, usually an early proof of concept. Replication, red-teaming and deployment work are still useful.
1Early testing
Early testing exists, but there is not enough evidence to judge the approach yet.
5Open
No active effort was identified, but the work is expected to be mainly tractable engineering.
7Open, hard
No active effort was identified and the problem looks research-hard. If you take one on, pick a single assumption inside the problem and test that.
Filter by status
01
Network evidence
Optical, network and datacentre systems
No workstreams match this status.
02
Reproducible workloads
ML runtimes, compilers and distributed systems
No workstreams match this status.
03
Partial recomputation
Algorithms, statistics and adversarial testing
No workstreams match this status.
04
Physical assurance
Trusted hardware, inspection and audit systems
No workstreams match this status.
05
Completeness
Memory, side channels and covert communication
No workstreams match this status.
Sketch your project
Make a copy of the project ideation worksheet, work through it as you read, then share it with gaurav@caish.org and submit the link through the submission form. First version by Wednesday 12 August, final by Sunday 16 August. Benjamin Hodgkiss, the facilitator, comments on every sketch before the week starts, and on the first day we use the sketches to settle projects and match you with residents.
