Project Influence - Chapter 6 - Checkpoint 3

Date: 2026-11-04 (Sydney) | Another collaboration between Andrew.human and Claude.ai

Replaces checkpoints 1 and 2 as the single carry-forward file. Written from files actually read in the Phase 2 / Chapter 2 conversation (listed in section 0). An earlier conversation said a file of this name had been written; no such file existed, so this is the first.

0. What this checkpoint was built from (and what it was not)

Read in full: Influence_Ch6_checkpoint_2.md, Influence_Ch6_rewrite_1.md, the 16-page Phase 1 end / Phase 2 start transcript, InfluenceMem_v1.0.0 (manual, tier files, self-test run), memstore.py v1.1.1, the staircase_20261101-044521 result folder (summaries, plans, run logs), Influence_Ch6_rollback_2.zip (layout, RESTORE.txt, standard_prompt.txt present). Not seen: the two Bala Priya C articles as text (known through rewrite_1 only; article B has 3 diagrams damaged in its .mht), the Clone's real memstore.db, any run of InfluenceMem on the real Qwen. Attach to the next conversation: this file, Influence_Ch6_rewrite_1.md, InfluenceMem_v1.0.1.zip, memstore.py, LogitStaircase_v1.5.0.zip, standard_prompt.txt, qwen_llm2_std_120a.json. The rollback zip (Influence_Ch6_rollback_2.zip) already holds everything older.

1. Problem & Resolution

Project: Influence, a pseudo memory-like LLM2 memory token broker. Aim: a model that does not repeat a corrected mistake. Target architecture (operator, 2026-11-04): Tier 1 memstore FACTS, Tier 2 agent-memory layers (working / episodic / semantic / procedural) after Bala Priya C, Tier 3 Influence enriching the reply. Hold off telling the articles' author until it works.

Chapter 6 findings (one prompt, md5 d78a0cc67c870050941cdfc719e0a365, Qwen3-4B-Q4_K_M, temperature 0, 120 tokens, one run per setting: 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. About +6 on the 4-30 band: a newline piece flips the reply from lecture to quiz mode (variety and repeat metrics missed it).
4. +50 on 4-30: the 4th-ranked token at all 120 steps; question-style filler, no content.
5. " missing" (one piece, never in the baseline top 50): +3 identical to baseline; +6 diverges at step 4 into three sensible sentences then loops ("So missing data is missing information."); +10 "missing" x117; +20 and +100 "missing" at all 120 steps. At +6 only 9 of 27 forced steps overrode the model's own top choice.
6. A pushed word with no job attached is taken literally; the model follows it but does not link it to the argument.
7. Greedy cascade: one early flip rewrites everything after it. Near-ties flip between whole-reply and one-token-per-request generation, so each step-by-step job makes its own baseline.
8. One llama-server slot: one job at a time (two concurrent jobs about 7000 s per level instead of about 150 s).
9. NEW, read from the result folder: " wartime aircraft" (--only, first piece " wartime", id 86852) at +2, +3, +4, +5, +6, +8: byte-identical to the baseline at every level. First differing step none, forced 0, completed 0, stray 0, mean shift 0.000. Run on v1.5.0; the summary says "scores reported before the bias: unknown". None of the operator's four predicted outcomes (aviation in general, the phrase, repetition, a question) occurred, and neither did the author's "something else" in the sense of a visible effect. The levels stopped at +8; " missing" broke through at +6, so the difference between the two words is unexplained and not yet tested above +8.

Built since checkpoint 2 | Item | Version | What | |---|---|---| | InfluenceMem | 1.0.0 | Three tiers: Tier 1 facts adapter, Tier 2 four layers with write policy, trust, provenance, TTL, decay, history, snapshot/rollback, evaluation; Tier 3 bounded, fading push (bias = max_bias x importance x confidence x trust, halves each time the word appears, stops after max-fires), per-push provenance. 35 checks on a stand-in server | | InfluenceMem | 1.0.1 | Tier 1 wired to the real memstore.py database (v1.1.1 schema): load-facts STORE --memstore [--memstore-db PATH]; read-only; approved entries only; key ms.<id>; no terms; importance and confidence stay 1.0. 42 checks, 0 failed; also tested against a database built by memstore.py itself (database file unchanged by the load) |

2. Risk Analysis

• Nothing has run on the real Qwen. InfluenceMem was tested only on a stand-in server whose replies are synthetic.
• Tier 3 pushes nothing from memstore: memstore has no terms field. With memstore as the only Tier 1 source, Tier 3 has nothing to push.
• A superseded or forgotten memstore entry stays valid in Tier 2 until invalidated (no sync built).
• Memstore corrections ("always injected") are not always retrieved by Tier 2, which ranks by word overlap.
• A push cannot carry a fact the model does not have; facts reach the model through Tier 2 text (--inject-text), which makes a new query/reply pair.
• The default push ceiling (+4) and the instruction screen for low-trust writes (a plain pattern list) are guesses, not tuned.
• Sensitive and inferred memstore entries load as approved, with tags.
• One prompt, one run per setting, temperature 0: repeatable, not general. Predictions so far are right about half the time; treat them as hypotheses.

3. Code Change

Delivered: LogitStaircase v1.0.0 to v1.5.0, SwapWords v1.0.0, InfluenceMem v1.0.0 and v1.0.1. Known cosmetic bug (not fixed): plan.md for --mid runs says "mode: whole reply per request" and names the original baseline. Proposed, not built (each would be listed before building): terms for memstore facts (sidecar file by entry id, or the hotwords extractor); sync that closes Tier 2 copies when memstore supersedes or forgets an entry; always-retrieve for memstore corrections; --suppress WORD as the correction mechanism; four-metric report (recall, precision, faithfulness by hand, staleness).

4. Pending

1. First run of InfluenceMem on the Clone (one job at a time, Qwen via Qwen34B-LLM2.sh on 8082): python3 selftest.py --server http://localhost:8082 --term effective, init store, load-facts store --memstore, run store --query-file standard_prompt.txt --dry-run, then run ... --max-bias 3 --gate 30 --max-fires 2 --n 120. A gate of 30 probably leaves the push idle for a word far from the top; --gate 0 is the first run that shows whether the push does anything.
2. " missing" at +4 and +5: not run (no result in the folder).
3. Bare-word " bomber" test ("Word: bomber / First thought that comes to mind, in one sentence:"): not run.
4. Wartime-aircraft levels above +8, or a capped push on that phrase, to test whether it can break through at all.
5. War prompt draft and --suppress: not built.

5. Operations notes

Clone: ~/memstore/LogitStaircase, ~/memstore/ for InfluenceMem; Qwen LLM2 via Qwen34B-LLM2.sh (8082). New output name each run (the scripts refuse existing folders). Ordered shutdowns only. Memstore database default: $MEMSTORE_DB, else $MEMDIR/memstore.db, else /root/memstore/memstore.db.

6. Working rules (standing, from the operator)

Statements over questions; critique before building; list unrequested changes before making them; bump @version and @date, manual version history, version shown at start; deliver a dotted, never-reused zip; release card; show actual output in full; Markdown only; never rename his files; plain words, no unexplained acronyms; the operator decides deployment timing.