Phase 2 backlog¶

Phase 1 closed at 2026-05-27 16:06 EDT with 7 arc-index pages live, 110 substantive concept pages, and the full architectural loop (arc autonomy, primer-quality critic, multi-objective supervisor, frontier-grounded roadmap). This doc is what Phase 2 picks up from.

See Phase 1 retrospective (in chat history) for the narrative of what shipped and what we learned.

Pre-loaded state (no action needed — these activate on next cycle)¶

The sprint queue and code are already loaded for the next session. When the server restarts, the next cycle automatically:

Sprint queue (28 arc items + 4 new content pending)¶

Track	Arc	Status	Items waiting
06-RL	`world-models-and-imagination`	mid-arc	1 step (the last one)
04-NN	`scaling-and-emergence`	not started	1 index + 4 steps
05-stat-prob	`probabilistic-programming-end-to-end`	not started	1 index + 5 steps
09-systems	`serve-an-llm-efficiently`	not started	1 index + 5 steps
09-systems	`train-at-scale`	not started	1 index + 4 steps
04-NN	`efficient-large-model-training`	not started	1 index + 4 steps
+ 06-RL / 08-causal / 10-complexity	`rl-for-post-training`, `causal-discovery-in-practice`, `scaling-laws-empirical`, `emergence-and-double-descent`	not started	~20 items

Cost to drain the full ready queue: approximately $10-15 at full quality, ~2 cycles.

Code improvements active on next cycle (no opt-in needed)¶

Path-style wikilinks resolve correctly — [[curriculum/.../slug]] form no longer renders as raw text.
Arc-index breadcrumb correct — "Arc syllabus" instead of "Step 0 of N".
Track index pages show arcs — the rebuilder walks subdirectories now.
Catalog auto-rebuilds with live badges (Step 4.7 in scheduler.py).
Primer-first writer voice — averaging 0.85 primer-quality on arc-step pages.
Arc-spin from roadmap — supervisor reads arc-roadmap.md for canonical arcs (no more random <track>-foundations placeholders).

Known issues to address in Phase 2¶

Ordered by impact on reader experience.

1. Arc-index pages struggle on primer (high priority)¶

Observation: of the 7 arc-index pages written this cycle, 3 scored primer ≤ 0.70 — including agentic-rlvr-reasoner-index, world-models-and-imagination-index, causal-deep-learning-index. The arc-step pages within those same arcs all landed at primer=0.85 cleanly.

Root cause: the writer's primer-first prompt was designed for concept pages (mentor walking through one mechanism). Arc-index pages are syllabi — they frame a journey, name a destination, set persona-fit, and pitch the reader on investing the time. Different voice. Different shape.

Fix: ship agents/skills/arc-index-craft.md — a writer skill for mode == "arc-index" only. Voice: "the editor of a curated 6-week course, telling you why this is worth it and what you'll have built at the end." Estimate: ~80 LOC, half a day.

2. `bayesian-neural-networks` chronically errors¶

Observation: the page has errored 4+ times across cycles, consistently at rev=2 with primer scores in the 0.70-0.74 band but other critic floors tripping. The writer keeps producing dense math without intuition tissue.

Fix: investigate at the page level — the concept may genuinely need a more careful narrative scaffold. Could be a one-off prompt amendment for this specific topic, OR could be a sign that BNNs are a topic where the model lacks training depth.

3. Critic-attribution → persona feedback loop¶

Observation: critic-info-architecture has been the universal weakest critic across every track for 8+ cycles, consistently flagging missing "Connected topics" and "Where this concept appears" sections. The system sees the pattern (the retrospective names it in every report) but doesn't act on it.

Fix: when a critic flags the same failure mode N cycles in a row, the retrospective should propose a persona amendment that the writer reads on subsequent runs. The persona update is a small YAML diff in agents/src/frontier_agents/personas/<track>.yaml. ~120 LOC plus tests.

4. The 13 needs-seeds arcs¶

13 of the 30 designed arcs in arc-roadmap.md reference concept slugs that don't exist yet (sparse-autoencoders, feature-circuits, cot-monitoring, video-generation, controlnet, lora-finetuning, clip-architecture, vision-language-pretraining, latent-dynamics, ring-attention, vector-search, embedding-models, self-verification, test-time-compute, causal-intervention, triton-kernels, physical-consistency, systematic-generalization).

Fix: the retrospective auto-seeder picks these up naturally — referenced 2+ times in any existing page → stub gets seeded → arc unlocks. Over 2-3 cycles this completes itself. No new code needed; just budget to run the cycles.

5. Author pages¶

The retrospective proposes author pages for heavily-cited researchers (Pearl, He, Vaswani, Bengio, Schölkopf, Schmidhuber). Currently marked moderate-risk and skipped. Worth promoting: author pages anchor citation discoverability and let concept pages deep-link back to one canonical write-up per major contributor.

Fix: extend apply_safe_proposals in retrospective.py to handle action_type == "author-page-seed" similarly to the arc-proposal branch — with guardrails (author cited 5+ times across pages, named in 3+ tracks). ~40 LOC.

Rough week-by-week progression¶

Week 1 — finish the ready arcs. Top up $20, run 2 cycles. All 12 remaining ready arcs land. Live catalog shows 17 🟢 arcs out of 30. Cost: ~$15. Quality watch: do the new arc-index pages still struggle?

Week 2 — ship arc-index-craft + author pages. Code: ~120 LOC. Run 1 cycle to refresh the 3 weak arc-indexes from Week 1 + add 5-8 author pages. Cost: ~$5. By end of week 2, every live arc has a primer-grade syllabus + author anchors.

Week 3 — seeds unlock more arcs. The retrospective will have auto-seeded most of the 13 needs-seeds arc dependencies. 5-7 more arcs become ready. Run 2 cycles to spin them. Live catalog shows ~22-24 🟢 arcs. Cost: ~$10.

Week 4 — critic-attribution + persona feedback. Code: ~120 LOC. Run 1 cycle with persona auto-amendment active. Watch whether critic-info-architecture scores rise across the corpus. Quality watch: avg confidence should climb from 0.77 toward 0.85.

End of Phase 2: 25-28 of 30 arcs live, primer-quality averaging 0.85, info-architecture critic no longer the universal complaint, and 8+ author pages anchoring citations. Total Phase-2 spend: ~$40.

The clean restart sequence¶

When you have a fresh top-up and want to resume:

# 1. Bump cap to match new headroom (current: 98)
sed -i.bak 's/^BUDGET_LIMIT_USD=.*/BUDGET_LIMIT_USD=120.0/' agents/.env

# 2. Restart server — DIRECT python, NOT via start.sh (that wrapper auto-respawns
#    on crash, which makes "stopping" require killing the wrapper too)
cd agents && source .venv/bin/activate
nohup python -m frontier_agents.cli serve --interval 999 > logs/server.log 2>&1 &

# 3. Trigger first cycle
curl -sX POST http://localhost:8765/trigger

The first cycle will drain the existing 32-item queue. After that, the supervisor auto-spins the next batch of ready arcs from arc-roadmap.md. Hands-off from there.

Phase 3 horizon (not yet started — for later thinking)¶

MVB executability harness. Every MVB recipe needs to be runnable. A nightly job that picks one MVB at random, sets up the named environment, runs the recipe, and checks the success metric. The 5-gate quality bar says the MVB should work; this would verify it.
Reader telemetry. Currently zero observation of which arcs/pages readers actually engage with. Even a simple "did you scroll past 60%" signal would let the retrospective prioritize improvements on the most-read pages.
Hybrid local + cloud critic fanout. Run the 8 cheap critics locally on the M-series Mac (Gemma 3 4B) while writer/reviewer stay cloud. Cuts ~$0.05/page = ~30% of per-page cost.
Cross-arc backlinking. When an arc step in causal-deep-learning references a concept that's also a step in causal-rl, link the two arcs to each other. The retro can compute this from the artifact-chain graph.

This doc is the canonical Phase 2 plan. Edit it as priorities shift.