📝 Experiment_001_checkpoint_10.mdv4.6.2 · 2026-10-28

Experiments 000, 001, 002 — Checkpoint (10)

Operator: Andrew. Written 2026-10-30 02:32 Sydney. Supersedes Experiment_001_checkpoint_9.md for everything after 2026-10-29 23:50; checkpoint 9 stays the reference for the recorder 1.3.0 options, findings 1 to 6, proposals 1 to 6 and risks 42 to 50 (kept, never overwritten). Earlier checkpoints _7 to _2 and Experiment_001_checkpoint.md also kept. Risks 46 and 48 are missing from the sequence (they are probably in _8, which was never attached). Status: hotwords.py 2.2.0 (hybrid retrieval) built, delivered and now run on the Clone's Python 3.6 against a real embedding model (MiniLM), on 3 arrays only. Recorder 1.3.0 is still not run on the real 4B and was not attached in the last conversation (only 1.2.1 was seen). No experiment has run since checkpoint 7. Resume: paste this file into a fresh conversation and attach NaturalRecorder_v1.3.0.zip, Exp000_queries_v1.0.0.zip, Hotwords_v2.2.0.zip, Memstore_Clone_v1.0.16.zip. For the experiment definitions, the 000/001 results and risks 1 to 40, use Experiment_001_checkpoint_7.md. Newest file wins. Terms: LLM1 = Qwen3-4B (chat model, 8080, go-between 8090). LLM2 = second 4B on 8082 (go-between 8092); the recorder runs on 8082 only. RAG = hotwords.py pseudo. Pool = top 50 candidates at one step. Prefill = tokens placed at the start of the assistant's turn. Greedy = always take the most likely token. Meaning engine = the optional embedding side of hotwords.py.

1. Delivered: hotwords.py 2.2.0

Hotwords_v2.2.0.zip (sha256 922befa06869b2201cda4795463a3720edff5fe2431252c1ecebf42ea9f8e8c5): hotwords.py and HOTWORDS_MANUAL.md, both 2.2.0, 2026-10-30.

2. Environment changes

3. Hybrid test (3 arrays, 4 queries, MiniLM)

Stored topics: #1 Influence supports the three memory emulations (episodic internal, semantic external, procedural operational); #2 the currency of Influence is mem-toks, keyed by a whole word, a word pair or a mini-phrase; #3 exploit the internal vocab and test associating relational tokens directly or as synonyms. Queries were paraphrases with almost no shared words, plus one unrelated query.

Query Word overlap Meaning (right array) Other arrays
memory types nothing #1 at 0.572 #2 0.430, #3 0.365
index keys nothing #2 at 0.532 #3 0.456, #1 0.373
built-in lexicon nothing #3 at 0.572 #1 0.562, #2 0.512
chocolate cake nothing none kept best 0.192

Findings:

  1. The meaning engine found the right array first on 3 of 3 paraphrases where word overlap found nothing, and ignored the unrelated query (best 0.19, floor 0.5).
  2. Margins are small inside one project (0.14, 0.08, 0.01). In query 3 all three arrays were kept (combined 0.319, 0.281, 0.256), so the wrong arrays' terms would be offered nearly as strongly as the right one's.
  3. No single floor separates right from wrong here (right 0.532 against wrong 0.562). The floor's job is rejecting off-topic text, and it does that. Floor 0.5 and weight 0.5 stay as they are. --arrays 1 would have been right 3 of 3; not yet tested as a run.
  4. Sample is far too small to set final numbers (risk 53).

4. Proposals (none built, none agreed)

  1. Logit bias as an Influence lever. The operator named manipulation of logits as one of the main Principles of Operation. llama-server accepts logit_bias (token id and a number added to that token's score before the next token is chosen) on LLM1. It would give a mem-tok a high value at the output, beside prefill injection. It needs token ids, which links to topic 3 (internal vocabulary; /tokenize; hotwords.py stores no ids). It would be a recorder change, not a hotwords.py change. Check against the operator's written Principles of Operation before any design.
  2. Next hybrid test: a store of dozens of real arrays from experiments 000 to 002, then lexical against hybrid on the same queries, then --arrays 1 against the default.

5. Risks (new; adding to 1 to 40 and 42 to 50)

  1. A meaning match can be on topic but wrong in fact: seen in the test (query 3 admitted two wrong arrays at 0.56 and 0.51).
  2. An array stored earlier from the same text is skipped as the query's own copy only if its term order is identical; otherwise it can come back as the top match (since 2.1.0, both modes). Repeated queries can echo their own stored terms.
  3. Floor 0.5 and weight 0.5 are set from 3 arrays and 4 queries only.
  4. Logit bias, if built: a word has several token forms (leading space, capital, split pieces), so a bias on one form can miss the others; a bias applies at every step, not only where the corrected mistake would recur, and a large value can distort unrelated text.

6. Open items

7. Standing operator rules (carry forward)