A guide to the notebook
Different domains, recurring questions: who can act across boundaries, what can be observed, and what the evidence actually supports.
Shared questions connect themes. Same project identifies a project link; Project extension marks an explicit extension.
Reading path 01
Process logs, access policies and signed receipts approach accountability at different layers.
Link to this pathHow can independent systems cooperate and keep a record of what happened?
Operational support for sensitive distributed processes
Start with process logs: checking whether activity across organisations follows agreed rules.
Identity, access and privacy across care organisations
A related governance problem in healthcare: coordinating identity, consent, access and transaction records across institutions.
🐸 Froglet: Signed Deals For Agents That Need Proof, Not Just Tool Calls
The accountability question reappears around agent services: signed offers, agreements and receipts make an exchange inspectable.
Froglet update: making federated learning rounds inspectable
Extends Froglet’s signed execution workflow to silo-local training steps and model aggregation.
Agents need receipts: a talk at AI Engineer World’s Fair
For an introduction to the same Froglet protocol, watch the service-publication and invocation walkthrough. The source records only the year of the talk.
Reading path 02
Governance, clinical use and candidate discovery each need their own evidence.
Link to this pathWhat turns health-data infrastructure into something useful for care or research?
Identity, access and privacy across care organisations
Start with the conditions for sharing: identities, consent and institutional access policies in PICASO.
PICASO in rheumatoid arthritis care
The same PICASO project reaches a clinical feasibility study, with patient and clinician experience as part of the evaluation.
Reanalysing immunopeptidomics data against noncanonical sequences
A separate research setting asks what existing immunopeptidomics samples reveal when searched against an expanded sequence catalogue.
Reading path 03
From string matching to KV caches, storage, correctness and runtime cost are separate questions.
Link to this pathCan a different representation reduce work without changing the result?
A patent for string matching in security gates
Start with a granted patent on data structures for string matching in security gates.
BCR-memory-4: exact long-context compression is real; serving-speed wins are not
A different application examines exact cache reuse and tiered compression, separating stored-token reduction from peak memory and latency.
HypercubeQuant: early notes from an A100 KV-cache experiment
Another KV-cache experiment tests exact prefix reuse on A100, while withholding memory claims until the benchmark measures comparable quantities.
Reading path 04
These records make the limits of a result part of the result itself.
Link to this pathWhat does an observation establish, and what remains unproven?
When rare observations destabilise security verification
Start with a model of security inference: rare observations can change an apparently confident assessment of network states.
BCR-memory-4: exact long-context compression is real; serving-speed wins are not
In cache evaluation, exact outputs and compressed representations do not establish faster serving or a large reduction in peak memory.
HypercubeQuant: early notes from an A100 KV-cache experiment
A second cache note rejects a misleading compression ratio and sets out the matched, concurrent comparison still needed.
Reanalysing immunopeptidomics data against noncanonical sequences
In biological analysis, peptide candidates and absence from sampled healthy tissues leave tumour selectivity and therapeutic usefulness unresolved.
Froglet update: making federated learning rounds inspectable
In federated learning, signed receipts bind an exchange to a provider’s statement; they do not independently prove honest training or privacy.