← The Independent 100: a study

kalomaze

#64 in The Basement Labs, the curator’s rank and section @kalomaze Relevance to our work: 4 of 5

Bracketed tags are evidence labels: [Verified] means we read the page, file or API response ourselves, [Reported] means a source says so and we did not check the underlying thing, and [Inference] is our own reasoning.

Credit

The Independent 100 list was curated by dan (@irl_danB) and published on 2026-09-24 at independent.prose.md (announcement). The ranks and section names are his, and number 100 is his open wild-card slot. This study is ours and is not endorsed by him or by anyone on the list.

What they publicly do

Pseudonymous account. The GitHub account kalomaze opened the Min P sampler pull request merged into llama.cpp on 2023-10-31, and has 25 commits in prime-rl [Verified: GitHub API, 2026-09-25, evidence/credits.jsonl].

Public work

  • Min P sampler (llama.cpp PR #3841) repo MIT last activity 2023-10-31

    Truncation sampler keeping tokens above a fraction of the top token's probability; since adopted by Transformers, vLLM and others per a later ICLR 2025 paper by other authors (arXiv 2407.01082) [Verified]

  • prime-rl (contributions) repo Apache-2.0 last activity 2026-02-12 · 2,086 GitHub stars on 2026-09-24

    25 commits to Prime Intellect's agentic RL trainer (checkpointing for tied embeddings, Gemma3 support, sanity checks) [Verified]

  • koboldcpp fork repo AGPL-3.0 last activity 2024-05-10 · 47 GitHub stars on 2026-09-24

    Personal fork used to test experimental samplers [Verified]

  • superclassifier repo Apache-2.0 last activity 2025-01-01 · 2 GitHub stars on 2026-09-24

    LLM frontend for building preference-model / behavioral-classifier data [Verified: description]

  • X thread: harness tool-call parser bugs as eval confound other no license stated last activity 2026-09-19

    Reports that much of the gap between models on bulk synthetic environments traced to chat-template / tool-call parser serving bugs, not model ability [Verified]

  • X thread: agents writing narrow memory notes other no license stated last activity 2026-09-13

    Critique of agents writing instance-level .md 'lessons' that do not generalize, and of models asserting unproven hypotheses [Verified]

Licenses of GitHub repositories are from the authenticated GitHub license API on 2026-09-25, one row per repository in evidence/licenses.jsonl (“no license file” means the API found none; “license file present, not recognised by GitHub” means GitHub could not classify it). Other licenses are as the source states. Check the license file before reusing anything.

Relevance to our work

4 of 5 lenses: evals, training, local-models, memory, research-method

Min P is a sampler we can set today on llama.cpp/MLX local models. Their current posts name two failure modes agent builders must guard against: eval scores polluted by harness parser bugs, and agent memory files full of non-generalizing notes. Most current work lives inside Prime Intellect repos.

Scored 0 to 5 against the study’s relevance rubric: 5 means public, openly licensed work squarely in our lanes that we could study or adopt now, and 1 means little connection to our work. The score is one reviewer’s judgement of how much this public work bears on our own work. It is not a ranking of people and says nothing about the person or the quality of their work.

What we would study next

  1. Tool-call parser / chat-template confound thread (x.com/kalomaze/status/2101192284025069884). Before trusting any cross-model eval, check that the harness parses each model's tool calls correctly Effort: small.
  2. Narrow memory-note critique (x.com/kalomaze/status/2099006407953838541). A test for any agent memory: do stored notes state principles or instances? Effort: small.
  3. Min P PR and discussion (github.com/ggml-org/llama.cpp/pull/3841). Default sampler choice for local MLX/llama.cpp models Effort: small.

Before reusing anything: Min P ships inside MIT llama.cpp. koboldcpp fork is AGPL-3.0; do not vendor. Note the min-p paper's authors do not include kalomaze.

Sources

Overall confidence in this record: Medium. Sources were read on 2026-09-24 and every link was checked on 2026-09-25. The full record is line 64 of people.jsonl.

Limits

Not verified: Current work beyond the listed repositories and posts.

This record comes from a first-pass survey of public material by an AI agent session on 2026-09-24, edited for publication on 2026-09-25. Nothing was installed or run for it. Roles, stars and licenses change; see the study’s limits.

Request a correction

Is something about this record wrong or out of date, or would you rather not be covered? Open a correction (a GitHub issue form with this page’s address filled in). A request to remove a record is honoured without argument; accepted corrections are recorded in the changelog with the date.

Help fix it

Each opens a GitHub issue form with this page’s address filled in. You need a GitHub account.

Made by Joshua Nowak (ZestStream) with AI coding agents. Not affiliated with or funded by Jeffrey Emanuel; Joshua uses his public tools and is a paying subscriber to jeffreys-skills.md.

Source on GitHub MIT licensed