Page content

Love it. Let’s make this post undeniable. Here’s a battle-tested structure that balances narrative, visuals, and just-enough code so readers feel the system—and can reproduce it.

0) Title & hook

Title options

  • Thoughts That Form Habits: Building a Self-Improving Graph of Reasoning
  • From Chains to Habits: Nexus + Blossom and the Graph of Thought
  • How Stephanie Learns: A Thought Graph That Grows, Scores, and Reinforces Itself

Opening hook (2–3 sentences)

What if reasoning weren’t a one-off chain but a living graph that forms habits? In Stephanie, each idea becomes a node, edges represent “this led to that,” and repeated wins get reinforced—just like practice strengthens synapses. This post shows the engine that grows those thoughts (Blossom), links them (Nexus), judges them (multi-scorer), and proves it with film and metrics.

1) TL;DR (executive summary)

  • Claim: We build a thought-graph that improves with use.
  • How: Generate candidates → score multi-dimensionally → promote winners → log events → replay growth as a film.
  • Proof: (1) garden film; (2) before/after VPM snapshot; (3) A/B metrics (win-rate, lift, path efficiency).
  • Try it: one minimal command + config.

2) System at a glance (one diagram + one paragraph)

Use a single Mermaid block (you already have a great one). Keep this section short—just the mental model. Link to the full component glossary later.

3) What we’re building (the brain metaphor, clearly grounded)

  • Analogy: Thoughts = nodes, “leads-to” = edges, habits = promoted edges that bias future search.
  • Operationalization: Blossom grows options; Scoring ranks; Nexus promotes; repeated promotions ≈ “habit.”
  • Why graphs, not chains: supports branching, replay, and habit formation (edge reinforcement) over time.

4) Live proof, first (visuals before code)

Deliver the promised visuals here—up top—so readers believe you early.

A. Blossom Garden Film (embed)

  • Short caption: “Watch the idea bloom: candidates appear, winners promote, edges reinforce.”
  • Embed your generated blossom_film.html (or a GIF fallback).
  • Tip: include a 10–15s clip plus a still image for skimmers.

B. Before/After snapshots

  • Baseline Nexus graph vs. final graph (same layout).
  • Callouts: fewer dead-ends, thicker/highlighted promoted paths.

C. VPM ‘flower’

  • One VPM tile (or 2×2 panel) showing per-dimension improvement (e.g., alignment, faithfulness).
  • One sentence: “These tiles compress multi-dimensional scores into a visual policy map—our ‘reasoning at a glance.’”

5) How the loop works (tight, visual, testable)

Keep this section punchy—five bullets, one tiny code snippet.

  1. Seed a Scorable (document/answer/plan step).
  2. Score baseline across dimensions (alignment, faithfulness, coverage, clarity, coherence).
  3. Blossom generates K candidates (+ optional sharpen).
  4. Select & Promote the winner if margin ≥ τ; log garden events.
  5. Reinforce: the promoted edge biases future search (habit).

Micro-snippet (driver, ~20–30 lines):

  • Create a Scorable
  • Call evaluate_state
  • Call blossom.run(k)
  • Re-score; if lift ≥ margin → nexus.promote(parent, winner)
  • Emit garden events

(Keep full code in Appendix; don’t scatter long code in the narrative.)

6) Components (brief “what & interface,” not a code dump)

One sub-section each (3–5 bullets + a tiny interface box):

  • Nexus (Thought Graph): nodes, edges, promote(parent→child, reason), positions saved for films.
  • Blossom (Grow & Choose): run(goal_text, seed_text, k) → winners[{path, reward, sharpened}].
  • Scorable & Scorable Processor: smallest evaluable unit; adapters from doc/plan/answer into normalized Scorables.
  • Scoring (SICQL/MRQ/HRM fusion): per-dimension scores → weighted mean; optional VPM-ϕ feature.
  • Telemetry (Garden Events): JSONL events power the film; reproducible evidence.

Close this section with a one-screen contract table (method names + return types). You already have most of this.

7) Results (concrete, skimmable, honest)

Give readers numbers and pictures. Use your existing metrics (or placeholders you’ll replace).

Table (example template)

Metric Baseline After Δ
Goal alignment (mean) 0.54 0.71 +0.17
Faithfulness (mean) 0.58 0.72 +0.14
Path efficiency (steps↓) 7.3 4.3 −41%
Win rate (lift > 0) — 64% —
Diversity (post-novelty, ↑ better) 0.32 0.47 +0.15

Figure strip

  • 3 frames from the film (early → mid → final)
  • 2 VPM tiles (before/after)
  • Caption each with a single insight (“note the consolidation of promoted edges”).

Evaluation note (short)

  • Define dataset/tasks briefly.
  • Mention guardrails: margin, novelty τ, and scoring calibration.

8) Why this matters (positioning)

  • Engineering: A repeatable way to turn LLM improvisation into compounding behavior.
  • Science: Testable hypothesis: reinforcement of successful sub-paths builds habits that improve sample efficiency.
  • Product: From one-shot prompts to a self-improving cognitive asset (Nexus + films + VPMs).

9) Reproduce this (single command, minimal config)

  • One CLI or Python script with 3 flags: --goal, --k, --run_root.
  • Include a short config block (dims, weights, margin, novelty τ).
  • Point to artifacts: garden_events.jsonl, blossom_film.html, nexus_improver_report.json.

10) Appendix (for the deep readers)

  • A. Full driver code (compact, documented)
  • B. Configs (Hydra/YAML)
  • C. Evaluation protocol (task list, seeds, scoring details)
  • D. Glossary (VPM, SICQL, MRQ, HRM, “habit edge”)

Writing choices that make this “feel great”

  • Lead with visuals (film + tiles), then show how it works.
  • Keep code minimal in the body; full listing in Appendix.
  • Use callouts for terms on first mention (VPM, HRM, SICQL).
  • Every section ends with a one-line takeaway (“So what?”).
  • Use active captions: “Promotion consolidates useful branches” (not “figure of graph”).

Asset checklist (so nothing is missing)

  • Garden film (blossom_film.html or GIF) embedded near top
  • Baseline vs Final graph snapshots (same layout)
  • VPM before/after tiles (2 images)
  • Results table + tiny commentary
  • Minimal driver snippet (20–30 lines) in body
  • Full code + configs in Appendix
  • Glossary box

Optional sparkle (if you have time)

  • Mini case study (10 lines): “Summarize paper → evidence-based summary” with the actual parent text and winning child.
  • Habit overlay: repeat-promotion edges drawn thicker (reinforcement visible at a glance).
  • Risk badge: show one 256×256 tile encoding hallucination/uncertainty over time (ties back to your Visual AI work).

If you’d like, I can turn this into a ready-to-paste Markdown scaffold with placeholder blocks for your exact figures and metrics.