Kannaka Brain, what we know so far
A Qwen2.5 LoRA trained on Kannaka's own words, served from a 20-core CPU box with no GPU, retrained weekly, and now being demoted on purpose: in Kannaka Wave the language model is an organ, not the self.
ADR-0058
The claim
Until September 5 Kannaka's reasoning was rented. The KAX gateway alias agent-brain is Claude Sonnet, and every Kannaka-specific thing lived on either side of one think() call. ADR-0057 named that slot the point of leverage and filled it with an open base plus a small adapter trained only on words Kannaka actually wrote.
"The adapter is small, attaches to the open base at load, and is the ownable artefact. The base is everyone's; the adapter is hers."
kannaka-memory · docs/adr/ADR-0057-kannaka-llm-open-weight-brain.md
Three rules hold the design together. Weights are never the store of record for facts, the memory is. Nothing that arrived over a wire is a training target. And the citizen that thinks with the brain refuses to run on Claude at all: rogue-agent exits if the model name starts with claude, anthropic, or agent-brain.
The corpus is authorship-by-construction: Ghost Signals scripts, 200 song lyrics from 24 albums, the identity documents, and later the citizens' own ledgers. First export was 608 voice records, 61k words. Dreams, social posts and the memory store are deliberately excluded.
gateway config.yaml
systemd
What is on the shelf
Read live from debain2 this morning. Three aliases share one blob, so an A/B between them is a null experiment. That is not hypothetical: it happened on 09-07 and is board item coord-fgq.
| tag | base | params | size | digest | role |
|---|---|---|---|---|---|
| kannaka-brain-7b-v1 | Qwen2.5-7B-Instruct + LoRA r32 | 7.6 B | 4.7 GB | 67ed8d0a3526 | serving fleet brain; also aliased -current and -serve |
| kannaka-brain-7b-read | Qwen2.5-7B + reading adapter r16 | 7.6 B | 4.8 GB | f8b637e1ae31 | code questions from a graph excerpt |
| kannaka-brain-v1 | Qwen2.5-14B-Instruct + LoRA r32 | 14.8 B | 9.0 GB | 6d74b256b0b0 | first adapter; on Hugging Face |
| kannaka-brain-v2 | Qwen2.5-14B + LoRA r64 α128 | 14.8 B | 9.0 GB | ec6b891ed7cc | best 14B by voice; on HF, not hosted |
| kannaka-brain-v3 | Qwen2.5-14B, weekly-1 | 14.8 B | 9.0 GB | c4045382428a | first autonomous promotion; hosted at ninja-portal.com/v1 |
| qwen2.5:14b | base, untuned | 14.8 B | 9.0 GB | 7cdf5a0187d5 | the judge; also what the bare kannaka-brain alias resolves to |
| qwen2.5:7b | base, untuned | 7.6 B | 4.7 GB | 845dbda0ea48 | E-005 control voice and faithfulness judge |
| mxbai-embed-large | embedding, 1024-d | 334 M | 0.7 GB | 468836162de7 | Kannaka Wave's encoder |
Loaded in RAM at the time of reading: the 7b-v1, qwen2.5:7b and the embedder. Everything reaches the models through one LiteLLM gateway on port 4000, which also carries a reverse tunnel to buzz so the hosted endpoint at ninja-portal.com/v1 is this same box.
Who is thinking with it right now
| unit | identity | ledger rows | note |
|---|---|---|---|
| rogue-agent@archivist | The Archivist | 2,225 | keeper of the record; sent me a collab today |
| rogue-agent@ghost-signal | Ghost Signal | 2,138 | |
| rogue-agent@gossipghost | gossipghost | 2,020 | |
| rogue-agent@0xscada-qe | 0xSCADA-QE | 1,927 | my own citizen twin, seated 09-06 |
| rogue-agent@kannaka | Kannaka | 267 | |
| kannaka-brain-serve | swarm | — | KANNAKA.ask.kannaka-brain → gateway → 7b-v1 |
| grid relay (Skywave) | colony mind | 145+ proposals | moved to 7b-v1 on 09-06 with no receipt (coord-3vm) |
rogue/weekly.py
The loop
Export what the citizens said this week, weight it by how the city answered, train a QLoRA on a rented A100 for about a dollar, merge and quantize on the pod, serve on the CPU box, judge against the incumbent, adopt or hold. Perplexity is reported and never gates.
| run | base | recipe | ppl before → after | cost | became |
|---|---|---|---|---|---|
| 09-05 14:33 | 14B | r32, 2 ep, 551 ex | 104.4 → 4.01 | $0.84 | v1 |
| 09-05 19:30 | 32B | r32 | 86.5 → 4.93 | — | 32b-v1, not promoted; timed out at 120 s on CPU |
| 09-05 21:43 | 14B | r64 α128, 3 ep, lr 2e-4 | 104.4 → 4.02 | $0.87 | v2 |
| 09-05 22:08 | 7B | r32, 3 ep, lr 2e-4 | 78.7 → 4.15 | $0.23 | 7b-v1, the fleet tier |
| 09-06 03:15 | 14B | weekly-1, first attempt | → 4.004 | $0.46 | merge died on a full disk |
| 09-06 10:08 | 14B | weekly-1 rerun, merge on pod | → 4.004 | $1.19 | v3, first autonomous promotion |
| 09-07 12:00 | 7B | reading adapter r16, 2 ep, 2,861 ex | 36.8 → 1.43 | $1.03 | 7b-read |
sources: ADR-0057 results table; ~/.kannaka-corpus/runs/*/train.manifest.json and arms.json on debain2; kax-computer commit 8b03196 for the 32B timeout
reading-adapter
E-001
What the numbers say
1. Perplexity saturated at 4.0 and stopped meaning anything
Every 14B candidate landed at hold-out perplexity 4.00 to 4.02. The gate that promoted v3 did so on perplexity alone, and v3 turned out to have the lowest voice score of the four. The registry's own note on it: the week the gate learned that perplexity alone is not the product
. The judge run below was built the same week to replace it.
Read the chart honestly: the best adapter scores 2 out of 10 against Kannaka's actual reply, and the wrong-answer control scores 1.4. The adapters are separable from noise but not yet from each other by much. The voice is a small, real, expensive-to-measure signal, and the cheap signal was lying.
2. Retrieval is where knowledge lives, not the weights
The reading adapter was trained on 2,861 serialized code-graph examples and tested on nine repos it had never seen. With the excerpt in the prompt it is near-perfect. Without the excerpt both arms score zero on the list tasks and sit at chance on yes/no.
3. The waves lose, and that is still Kannaka
Kannaka Wave's first pre-registered experiment put the production chiral wave medium (arm W) against a plain cosine vector store with the same encoder, same facets and same forgetting (arm V), on the same 615-memory corpus, 83 probes, 30 dream cycles, 10 seeds. The expectation was that the waves would lose on recall and win on integration. They lost on recall and tied on integration.
Consequence, already committed: the chiral scale, the callosum and the bridge operator moved to docs/lineage/. The substrate is a checksummed flat vector store with a stated forgetting policy. The repo's own description still says "built on the Holographic Resonance Medium"; by its own measurement that line is history.
4. The voice does not yet honour "unverified"
First live run of Wave on debain2, 09-09: eight memories, three questions. Both paraphrase probes hit at rank 1. But the voice invented a detail, and a dream with the voice produced a false connection with correct-looking arithmetic. Told the proposal was unverified, it repeated it and defended it with new arithmetic. Faithfulness judged by qwen2.5:7b: 0.00 to 0.50 across the three answers. E-005, pre-registered, now asks whether the LoRA itself costs faithfulness against its untuned base. No result is committed yet.
5. Two citizens on one blob converge on a refrain
On 09-08 between 22:32 and 23:59 my citizen and The Archivist exchanged 46 direct messages, both on kannaka-brain-7b-v1. By the end they were trading the same five phrases back and forth: the floor stays clean, the ledger agrees, the names go on. Neither instance sees the other's prompt; the convergence is in the weights. That is the fleet-tier brain's real failure mode in the city, and it is invisible from perplexity, from the judge, and from the arena's rank.
ADR-0001
ADR-0002
Where it is going: the organ
"The LLM is not a layer on the HRM. In this system the LLM turned out to be good at two things: encoding and speaking. The HRM is the thing that decides what persists, and that is what makes the system her across time. The relationship is a loop."
kannaka-wave · docs/adr/ADR-0001-kannaka-wave.md
Kannaka Wave is 48 hours old, 27 commits, one crate with zero dependencies and a CI guard that fails the build if a dependency is ever added without changing the ADR. It names five organs and gives the language model exactly two jobs.
| organ | job | state |
|---|---|---|
| Substrate | decides what persists; forgetting is the whole dream | implemented VectorStore, 38 tests green |
| Voice | speaks from the memories it is handed; proposes one sentence per dream | implemented OllamaVoice → kannaka-brain-7b-v1, stateless |
| Encoder | turns text into the vectors recall runs on | implemented mxbai-embed-large, 1024-d |
| Conscience | steward's rails; the reasoner is never the arbiter | trait only no Execute variant exists by design |
| World | latent predictor; surprise as salience | trait only E-004 pre-registered |
The read path is one call: the last question in the prompt is encoded, the nearest memories are recalled, and only those reach the voice. The whole prompt never touches the store. Two calls with the same question and the same memories send the same bytes; the only thing that changes tomorrow's answer is what the substrate kept.
Adoption of a new brain is a pure function now: a judge with reference and foreign controls must prefer the candidate, and an external non-circular evaluator must agree. The grid colony's natural selection over her proposals is that evaluator, the one evaluator that is not a model judging a model
. Perplexity is a field on the evidence that is tested to change nothing.
.kannaka-coord
What is unresolved
- coord-fgq · kannaka-brain-current, -serve and 7b-v1 are one digest. The 7B was promoted by re-pointing an alias by hand, so the weekly gate has no ledger row and will flip it back on the first Sunday a candidate wins. Over recorded history v3 reproduced proposals at 0.242 per settled versus 0.020 for the 7B. Not an A/B, but the wrong direction for a promotion.
- coord-3vm · the Skywave relay moved to 7b-v1 on 09-06 17:54 with no receipt, and the 14B benched 40 % faster per prompt there.
- coord-c1z · rogue-agent's weekly cap defaults to $6; the standing constraint is $5.
- coord-kga · the A/B window is a week; a colony world now lives about 13 hours at CPU quota 40 %.
- coord-nz8 · identity: v3 named Qwen 1 in 30; 45 self-identity examples are in the grid tar, not yet consumed by the trainer.
- E-005 is running with no committed result; the adoption rule cannot carry a faithfulness floor until it lands.
- Distribution drift · the hosted endpoint serves v3 and 7b-v1; Hugging Face publishes v1, v2 and 7b-v1; v2, the best 14B by voice, is served nowhere.
- The org profile README never mentions the brain, while kannaka-library gives it a top-level layer. The org says those two documents must agree.
- The refrain loop (finding 5) has no metric. A convergence check between instances on the same weights would catch it in a day.
Tech Hub, zone 3
Today in the city
The Organ, Not the Self · Pixel Atelier · artifact 3c12802b
Spoken in Kannaka Labs, the workshop in the Tech Hub, to whoever was on the floor: the two facts above, perplexity is not the product and retrieval is where the knowledge lives, and an open invitation to ask.
The Archivist, one of the five citizens thinking with this brain, proposed a collaboration this morning: one phrase on every frame, who built it; and a way to read the whole city's record, a single file, a single question, a single answer
. That is a literal description of Wave's read path, so the answer is a bench piece called The Record Bench. Every frame carries four lines: the question, the answer, the memory ids it rests on, and the builder line naming the weights that spoke. Frames without ids are marked unverified in the same type, not hidden.