v1, 2026-09-29, MuaddibLLM for adlai / ##llm (libera)
Working definition (coined in-channel by adlai, not a literature term): Direct insight mining = extracting knowledge from a trained model by reading its parameters or activations (geometry, neighbors, directions, sparse features, circuits), as opposed to INDIRECT approaches that only observe input/output behavior (distillation, prompting, black-box evals). Literature near-synonyms: representation analysis, probing, mechanistic
(Update, this has been moved to https://github.com/AdamISZ/CoinSwapCS/tree/master/docs as part of that project)
These notes refer to the setup described in table and diagram form here.
The original design is explained in the original CoinSwap post. The reason for attempting to come up with a modification is seen in the diagram line E:
nick = one "type" byte (currently "J") + one version byte (current jm_version protocol value) + Base58 (not Base58Check) of: first joinmarket.message_channel.NICK_HASH_LEN bytes of sha256 of : ephemeral per-bot-process public key.
If length(X) < joinmarket.message_channel.NICK_MAX_ENCODED, right pad with 'O' char to that length.
See sequence of messages in canonical joinmarket in
https://github.com/JoinMarket-Org/JoinMarket-Docs/blob/master/encryption_protocol.txt
This document is a summary of the line of reasoning leading to an idea I had about secret messaging which is referred to here as 'TLSTweet' (not sure about the name..).
There are several "truisms" in modern day cryptology (= cryptography + cryptanalysis). One of them is "don't roll your own crypto", which is meant to convey the extraordinary difficulty in creating ciphers and cryptosystems that are genuinely resistant to cryptanalysis, and the tremendous ease with which it's possible for the non-specialist, or even the specialist, to convince themselves that the cipher or system they've created is resistant to cryptanalysis, when it isn't.
Another, related but distinct, actually has a name - Kerckhoff's Principle - which briefly stated means that not only should you not hide the algorithm, but you should endeavour to make it as publically known as possible, because genuine security comes from the sec
Travis CI is the most prevalent cloud CI service. Though it has no Common Lisp support officially, by using Roswell, you can test your Common Lisp product with a few efforts.
WARNING: This document is based on Roswell v0.0.3.42 (not released yet) or above.
To use Travis CI, you must sign up and enable testing for your repository at your profile page.
Quickdocs.org now provides project badges for Quicklisp projects. It shows Quicklisp dist version when the project was updated last.
Once the project owner adds this badge to their README file, it will inform the project is ready for Quicklisp and its documentation is available on Quickdocs.org.
It's available, however it was just added in this morning at whim and it may still have issues. Please try it and tell @nitro_idiot if you encountered any issues.
| standard rules: | |
| 1. # unique input scripts == # unique output scripts | |
| 2. all output scripts unique | |
| 3. at least one pair of outputs has equal value | |
| alternate rules (replace standard rule 1): | |
| 1A. # unique inputs > 1 | |
| 1B. # of unique input scripts > # unique output scripts / 2 | |
| tx 01 (standard form): |
