Project Influence - Chapter 6 - Checkpoint 2 (rollback point)

Date: 2026-11-03 (Sydney) | Another collaboration between Andrew.human and Claude.ai

State of everything BEFORE the rewrite through the two agent-memory articles. Restore by unpacking Influence_Ch6_rollback_2.zip; MANIFEST.md5 lists every file.

1. Problem & Resolution

Question: how far can a logit bias (added to token scores at the last step) steer the local Qwen (LLM2: Qwen3-4B-Q4_K_M, llama-server 8082 on Clone) as a "pseudo memory" lever? One fixed standard prompt (md5 d78a0cc67c870050941cdfc719e0a365), temperature 0, 120 tokens, top-50 scores.

Built since checkpoint 1 | Item | Version | What | |---|---|---| | logit_staircase.py | 1.1.0 / 1.1.1 | Step-by-step mode (one token per request): --gate N, --follow LEVEL; own no-push baseline per job with a near-tie check | | | 1.3.0 | --mid LO-HI: at every step push whatever the model ranks LO..HI | | | 1.4.0 | --only WORD: one word or phrase, first piece pushed at every step at big levels | | | 1.5.0 | --reuse-baseline FILE (optional; operator decided each query/reply pair stands alone, so unused) | | swap_words.py | 1.0.0 | Text-side word swapping with a reproducible map (SwapWords_v1.0.0.zip) |

Findings so far (all one prompt, one run per setting, so indicative only)

1. Gentle pushes on plausible words stay fluent. Forced rare words or first pieces hijack the topic and loop.
2. Band width (4-12 vs 4-30) matters little; the level matters.
3. Around +6 on the 4-30 band a newline piece (".\n", unpushed rank 4) flips the reply from lecture-continuation into a quiz/answering mode. The usual metrics (word variety, repeated runs) missed this genre flip.
4. +50 on the 4-30 band: the model picked its 4th-ranked token at all 120 steps. Reply opens with questions ("Now the question: Which two things is survivorbias an explanation about in relation?") and stalls ("Alright. Alright."). No repetition, no content. Baseline: percentages lecture that loops (23% repeated 4-word runs).
5. Single word " missing" (one piece; not in the baseline top 50 at any step): | Level | Result | |---|---| | +3 | identical to baseline | | +6 | diverges at step 4: "Fourth is the missing data. This is a very common one. For example, if you're missing data, you can't really know what's going on." then "So missing data is missing information." repeated; 27 forced steps, but the model's top choice overridden at only 9 (it kept choosing the word itself) | | +10 | "Fourth is missing missing ..." (117 forced) | | +20, +100 | " missing" at all 120 steps (byte-identical) |
6. A pushed word with no job attached is taken literally: the model defines it and repeats it. It follows the word and the sentence pattern but does not link the word to the argument.
7. Greedy cascade: one early flip rewrites everything after it. Near-ties can flip between whole-reply and one-token-per-request generation (steps 63 and 69 seen), so every step-by-step job makes its own no-push baseline.
8. One llama-server slot: two concurrent jobs thrash the cache (about 7000 s per level instead of about 150 s). One job at a time.

2. Risk Analysis

• One prompt, one run per setting, temperature 0: repeatable, not general.
• Fixed push has no stop: once the word is in the text, Attention makes it likelier, so the reply collapses to one word at +10 and above.
• First piece only; wrong-sense leaks (e.g. " bomber" means aircraft, Unabomber, jacket).
• Metrics (variety, repeats) can say "better" while content gets worse (+50 case).
• 4B model at 4-bit: may be unable to use a pushed word beyond following it.
• Predictions so far right about half the time; treat them as hypotheses.

3. Code Change

• No code changed this checkpoint. Delivered earlier: LogitStaircase v1.0.0-v1.5.0 zips, SwapWords v1.0.0.
• Known cosmetic bug (not fixed): plan.md for --mid runs says "mode: whole reply per request" and names the original baseline.

Pending

• Result of --only " wartime aircraft" --levels 2,3,4,5,6,8 (operator to run and attach; not yet received).
• " bomber" bare-word test: prompt "Word: bomber / First thought that comes to mind, in one sentence:" (awaiting "go"). Human baseline from a chat with Claude: operator said Unabomber; Claude first said the Dambusters Lancaster.
• " missing" at +4 and +5 (find where the good start begins).
• War prompt draft (test/war_prompt_draft.txt) awaiting approval; --suppress WORD not built.
• Capped or fading push (word fires a few times then stops): not built; would be listed as an unrequested change first.
• Narrow 4-12 band +6 not run.

Operations notes

• Clone: ~/memstore/LogitStaircase; Qwen LLM2 via Qwen34B-LLM2.sh (8082).
• Run one job at a time; new output name each run (the script refuses existing folders).
• Timing: whole-reply run about 1 min per level; step-by-step about 3 min baseline plus per level.