deploy: update Space from deploy_preflight --push

README.md

@@ -22,18 +22,28 @@ tags:
 
 # Microfactory Node: 3D Printer
 
-*The 3D-printing node of the Microfactory — its resident expert is **the Chief Engineer**, a small local model that learns the craft job by job.*
+*The 3D-printing node of the Microfactory. Its resident expert is **the Chief Engineer**, a small local model that learns the craft job by job.*
 
-> My grandfather had a shop — an electrical engineer who tinkered, a home RadioShack with a
+**It predicts where a 3D print will fail before it runs, learns from every job, and runs entirely on your own machine for $0 a month.**
+
+<!-- HERO IMAGE (the strongest single frame; capture at ~1600px wide, save to assets/screenshots/). -->
+<!-- Recommended: the BUILD page mid-reasoning: Chief Engineer O'Brien's precedent read on the left,
+     a predicted failure region called out, the filled LAYER cross-section below. -->
+![Microfactory Node: 3D Printer. The Build page, O'Brien reads precedent and flags where the print will fail, before the nozzle moves.](assets/screenshots/hero-build.png)
+
+`Local Gemma via Ollama` · `two agents: O'Brien proposes, La Forge inspects` · `knowledge that compounds` · `honest about what it doesn't know`
+
+
+> My grandfather had a shop. An electrical engineer who tinkered, a home RadioShack with a
 > machine shop attached. He spent a lifetime accumulating skill that lived in his hands. Then
 > Alzheimer's took him, the shop was sold, and all of it just… went. The Chief Engineer is the
 > opposite of that: a small AI that runs on my own machine and learns the craft of 3D printing
-> the way a shop veteran would — job by job, remembering what worked in which conditions — and
+> the way a shop veteran would, job by job, remembering what worked in which conditions, and
 > tells me where a print will fail *before* it runs.
 
 **Build Small hackathon · Backyard AI track.** A small local **Gemma** model that has accumulated
 expert 3D-printing knowledge across prior jobs and applies it **proactively** to a new one.
-Knowledge **compounds** — job N+1 is better-informed than job N. The proactive part is the
+Knowledge **compounds**: job N+1 is better-informed than job N. The proactive part is the
 *judgment*: the model evaluates what prior jobs transfer to this one, applies precedent when it
 fits, and says *"no close precedent"* when it doesn't.
 
@@ -48,11 +58,11 @@ Job (geometry + material) + Environment (temp, humidity)
   → reflect → distill an environment-keyed lesson → ledger grows → next job is smarter
 ```
 
-The model **never grades its own outcome** — you click *Printed clean / Sagged / Stringing*, or
+The model **never grades its own outcome**: you click *Printed clean / Sagged / Stringing*, or
 run the **simulated world** (a deterministic stand-in for the printer + sensors). Either way the
 outcome comes from outside the model. The honest signal is what makes the compounding real.
 
-**This is more than a lookup.** Two knowledge sources feed every recommendation — *RAG* (retrieved
+**This is more than a lookup.** Two knowledge sources feed every recommendation, *RAG* (retrieved
 prior jobs the LLM reasons over) and a **learned parametric policy** (`learn/policy.py`) that stores
 setting offsets per (material, geometry, *environment-bucket*) and updates from each outcome. Because
 cells are bucketed, a lesson from one humid PETG bridge **generalizes** to the next, not-identical
@@ -83,44 +93,59 @@ Falls back to a deterministic advisor if Ollama is unreachable, so the demo neve
 UI always shows which model ran. Set `CHIEF_ENGINEER_MODEL=gemma4:e2b` if CPU latency is high.
 
 **Resetting demo state.** `data/lessons.jsonl` is now the **durable knowledge base** (12 seed +
-13 ingested lessons — *tracked*, not disposable); the Print loop appends earned/sim lessons to it
+14 ingested lessons, *tracked*, not disposable); the Print loop appends earned/sim lessons to it
 at runtime, and `data/policy.json` is the derived learned policy. For a clean curve between demos,
-**don't delete the ledger** — restore it to the committed baseline and clear only the derived policy:
+**don't delete the ledger**: restore it to the committed baseline and clear only the derived policy:
 
 ```bash
 git checkout -- data/lessons.jsonl     # discard this session's runtime lessons, keep seed + ingested
 rm -f data/policy.json                 # drop the learned policy (re-derives from the loop)
 ```
 
-On the **live Space** (no shell), use the **↺ RESET TO BASELINE** button in the **Review** tab —
+On the **live Space** (no shell), use the **↺ RESET TO BASELINE** button in the **Review** tab:
 it clears this session's accumulated runs + learned policy back to the curated baseline.
 
-## Workspaces — the real print workflow
+## Workspaces: the real print workflow
 
-1. **Studio** — define + preview the job: quick-load 3DBenchy, generate a primitive, or drop a
-   mesh (the engineer *infers* the part class itself — you don't pick it); choose material; the
+1. **Studio**. Define and preview the job: quick-load 3DBenchy, generate a primitive, or drop a
+   mesh (the engineer *infers* the part class itself, you don't pick it); choose material; the
    simulated **environment** (temp/humidity/plate position on a Creality Ender 3 V2) populates.
-2. **Build** — the pre-flight check, **before it prints**: the slicer / layer-by-layer virtual
+2. **Build**. The pre-flight check, **before it prints**: the slicer / layer-by-layer virtual
    print, the engineer's precedent read + reasoning + predicted failure regions, the Spine-validated
    settings + G-code (terminal readout), and a **second opinion** from a separate QA Inspector.
-3. **Print** — run *this* job through the closed loop and watch quality compound fail→clean: the
+3. **Print**. Run *this* job through the closed loop and watch quality compound fail→clean: the
    Engineer proposes → the Spine vetoes → a simulated world prints → the **Inspector grades each
    run** → policy + ledger learn. Simulate one print, or record a real outcome.
-4. **Review** — the compounding made visible: the live ledger (seed → earned → sim), the capability
+4. **Review**. The compounding made visible: the live ledger (seed → earned → sim), the capability
    mesh, and the Inspector's verdict on the whole run.
 
-**The hybrid evaluator (honest by design).** The Engineer — *Chief Engineer O'Brien* — never grades
+**The hybrid evaluator (honest by design).** The Engineer, *Chief Engineer O'Brien*, never grades
 its own work. The deterministic world (`sim/outcome.py`) produces the ground-truth pass/fail; a
 *separate* **QA Inspector** persona, *La Forge* (`core/inspector.py`), reads what O'Brien claimed vs
-what actually happened and writes the verdict — a second opinion in **Build**, a per-print grade in
-**Print**, a run verdict in **Review**. Physics + a skeptical second voice — never the proposer marking its
+what actually happened and writes the verdict: a second opinion in **Build**, a per-print grade in
+**Print**, a run verdict in **Review**. Physics plus a skeptical second voice, never the proposer marking its
 own homework.
 
+<!-- SCREENSHOT: the Print loop: the quality curve climbing fail→clean with La Forge's per-run grades.
+     The compounding payoff in one frame. Save to assets/screenshots/print-loop.png -->
+![The Print loop: quality climbs from failure to clean over a few iterations as the ledger learns, with La Forge grading each run.](assets/screenshots/print-loop.png)
+
+<!-- ───────────────────────────────────────────────────────────────────────────────────────
+MEDIA TODO (capture from the live Space at ~1600px wide; drop PNGs in assets/screenshots/):
+  • hero-build.png:   Build page mid-reasoning: O'Brien's precedent read + a 3D risk flag + LAYER scrubber (the hero, top).
+  • print-loop.png:   Print, the fail→clean quality curve + La Forge's per-iteration grades (above).
+  • second-opinion.png: Build, La Forge's 🔍 second-opinion card disputing a plan (the two-agent moment).
+  • studio.png:       Studio, Benchy loaded, material + simulated environment (optional, for the writeup).
+  • review-ledger.png: Review, the ledger growing seed → earned → sim (optional).
+Demo video + social post go in the Links section below once recorded/published.
+──────────────────────────────────────────────────────────────────────────────────────── -->
+
+
 ## Badges
 
 Off the Grid (local Ollama/Gemma) · Llama Champion (Ollama runs on llama.cpp) ·
 Sharing is Caring ([ledger trace →](https://huggingface.co/datasets/kylebrodeur/chief-engineer-ledger)) · Field Notes (build writeup). Off-Brand if the
-Astrometrics UI lands (see `../DESIGN.md`). **Not** Well-Tuned — fine-tuning is named as frontier.
+Astrometrics UI lands (see `../DESIGN.md`). **Not** Well-Tuned: fine-tuning is named as frontier.
 
 ## What's real vs frontier (honest claims)
 
@@ -129,7 +154,7 @@ Astrometrics UI lands (see `../DESIGN.md`). **Not** Well-Tuned — fine-tuning i
   proactive geometric risk flags, fully local Gemma inference, human-reported + simulated outcomes,
   knowledge ingestion from slicer/firmware configs.
 - **Simulated (the one boundary):** print outcomes, via a deterministic physics-lite stand-in for the
-  printer + sensors (`sim/outcome.py`) — the model never grades its own work.
+  printer + sensors (`sim/outcome.py`): the model never grades its own work.
 - **Frontier (not built):** weight-level fine-tuning on the accumulated ledger, real distributed
   multi-node execution, the physical interfaces (g-code streaming, env sensors, camera defect CV).
 
@@ -144,7 +169,7 @@ Astrometrics UI lands (see `../DESIGN.md`). **Not** Well-Tuned — fine-tuning i
 
 ## License
 
-MIT. **No OrcaSlicer/PrusaSlicer code** (AGPL-3.0) �� trimesh/PySLM/manifold3d/pyclipr only.
+MIT. **No OrcaSlicer/PrusaSlicer code** (AGPL-3.0): trimesh/PySLM/manifold3d/pyclipr only.
 
 ## Source
 

assets/screenshots/README.md

@@ -0,0 +1,17 @@
+# Screenshots (Space card + writeup)
+
+Drop PNGs here, captured from the live Space at ~1600px wide. The Space `README.md` and the
+writeup reference these paths. Until a file exists its image link shows a broken icon, so add
+them before deploying the card for submission.
+
+| File | What to capture | Used by |
+|---|---|---|
+| `hero-build.png` | **Build** page mid-reasoning: O'Brien's precedent read on the left, a predicted failure region called out, the filled LAYER cross-section below. The hero. | README top |
+| `print-loop.png` | **Print**: the quality curve climbing fail→clean with La Forge's per-iteration grades. | README |
+| `second-opinion.png` | **Build**: La Forge's 🔍 second-opinion card disputing a plan (the two-agent moment). | writeup / social |
+| `studio.png` | **Studio**: Benchy loaded, material + simulated environment. | optional |
+| `review-ledger.png` | **Review**: the ledger growing seed → earned → sim. | optional |
+
+Tip: full-screen the browser (drop the tab/URL/taskbar), bump zoom to ~110% so the LCARS
+monospace reads on playback. Use a climbing job (PETG overhang ~30C/65%) so the Print curve
+actually climbs.