Project Influence - Chapter 6 - Checkpoint 1

Date: 2026-10-31 (Sydney) | Another collaboration between Andrew.human and Claude.ai

1. Problem & Resolution

Question: is an "organic" memory mechanism (LLM2 steering LLM1 with mem-toks) viable? How hard can word scores be pushed before replies break?

Done this chapter

• LLM2 baseline dumped: Qwen3-4B-Q4_K_M on Clone (8082), standard prompt (625 tokens), temperature 0, 120 steps, top-50 log-probs. Two runs identical (qwen_llm2_std_120a.json, ..._120b.json).
• Baseline facts: 46 of 120 steps have a top-1 share under 50%; median gap rank1-rank2 is 1.25; median spread rank1-rank50 is 8.81; step 1 is among the hardest.
• Hybrid RAG (hotwords.py v2.2.0 + MiniLM embeddings) tested: the RAG block changes the reply from step 6; mean shift 0.4-1.7 log-prob points (max 5.43). Dump: qwen_llm2_rag_hybrid_120.json.
• Built logit_staircase.py v1.0.0 (LogitStaircase_v1.0.0.zip). Real run on Clone: staircase_20261031-154509.

Staircase result (one prompt, 120 tokens, one run per setting) | | Mem-tok first words (22 distinct ids from 29 candidates) | Control words (22) | |---|---|---| | First reply change | +2 | +5 | | Candidate phrases completed | 0 at all levels | 1 at +5/+6 (blanket, step 119) | | Stray first words | 3 at +2, 6 at +6 | 0 | | Breakdown (repeat/distinct ratio) | none (ratio 0.48-0.52, repeat 1) | none |

Reading:

1. +6 did not break the replies; the push is too weak to damage because most candidate first words sit 5-15 places down (-5 to -12 log-prob).
2. First-word push leaks: the word arrives without its phrase and in the wrong sense (" false" from "false positives" turned step 4 into a "false cause fallacy" topic; " healthy" gave "healthy meals"). It can steer topic, not deliver a phrase.
3. Mem beats control at +2 mostly because mem words sit closer to the top; the control is not matched on closeness. This is not proof the model "understands" mem-toks.
4. The server reports scores BEFORE the bias (confirmed), so "mean shift 0.000" in the summary is uninformative; use the candidate_track logs.
5. One early flip rewrites everything after it (greedy cascade).

2. Risk Analysis

• Static push applies at every step to the same token (common first tokens flagged: "percentage of what" excluded).
• First piece only; capitalised / no-space variants not pushed (risk 54).
• Control not matched on piece shape or closeness (risk 53, plus floor/weight tuned on a tiny sample).
• Single prompt, single run, temperature 0 only: repeatable, not general.
• Hotwords limits: rarity marks subject words as common as the store grows; query array dominated by opening words.
• Unresolved: whether a 4B model can use mem-toks beyond topic steering.

3. Code Change

• NEW logit_staircase.py v1.0.0 (2026-10-31), control_words.txt (40 words), LOGIT_STAIRCASE_MANUAL.md. Stdlib only, Python 3.6 floor; tested 3.11-3.13 and against the Clone's llama-server.
• No other code changed. Unrequested additions (announced): control group, score-reporting self-check, common-token flagging.

Next steps (when we pick it back up)

1. Push the SECOND piece of each phrase too (or the whole phrase path) to see whether phrases can be pulled through.
2. Per-step push: apply the bias only on steps where the candidate is already in the top ten.
3. Try +1.5 / +2.5 steps and the mem_in / mem_new groups separately.
4. Match the control on closeness to the top and piece shape.
5. Later: strengthen the store by usage; LLM1 vs LLM2 comparison.

Operations notes

• Start LLM2 with Qwen34B-LLM2.sh (8082). minilm-l6.sh (8083) is needed only for hybrid RAG runs. Use HOTWORDS_DB=/root/memstore/hw_std.db prefix on every hotwords command.
• Use a new output name each run. Stale/wrong files in ~/memstore: qwen_llm2_std_1.json (0 bytes), pseudo_std_hybrid.json (0 bytes), pseudo_std_hybrid_2.json (wrong store).
• Timing: new prompt read ~154 s; 120 tokens 57-71 s; full staircase ~13 min.