<< AI DIRECTION CAMPAIGN >>
PLAYTHROUGH
What happens when you stop using AI as a faster Google
and start running it as a party.
▶
SAVE POINT — CHAPTER 5: MINIBOSS ENGAGED
> Most users prompt AI like a search bar.
> Ask question. Get answer. Move on.
>
> PLAYER_1 is doing something else.
> Directing specialized agents over multi-week timescales,
> calibrating their novelty claims, routing between them,
> shipping machine-checked novel mathematics with
> documented operational discipline.
>
> This page is the campaign log.
THE PARTY
═══════════════ PRESS START TO MEET YOUR TEAM ═══════════════
STRATEGIST
CLASS · RELAY / ANALYST
AGENTstrategy-researcher
HANDLESPlain-language synth
ROLERoutes Daniel ↔ team
SIGNATURE MOVE"Calibrated read"
Translates technical surfacing into operational and strategic framing. Asks the meta-questions before Daniel does. Drafts pre-flight novelty passes.
M-RESEARCHER
CLASS · MATHEMATICIAN
AGENTmath-researcher
HANDLESMath design + novelty
ROLEDerives helpers, runs Verdict-A sweeps
SIGNATURE MOVE"Verdict-A clean / B with caveats"
Owns mathematical direction-setting under autonomy grant. Designs the helpers. Coordinates prover. Maintains the novelty discipline.
PROVER
CLASS · LEAN TACTICIAN
AGENTprover
HANDLESLean formalization
ROLEGrinds sub-cases, applies helpers
SIGNATURE MOVE"Build clean. Sorry count: N."
Heads-down formalization. Translates paper-math into machine-checked Lean 4. Fights the elaborator. 10 helpers crafted to date.
BOSS PROGRESSION
Where the campaign is now, in chapter terms. Each chapter has its own check — the AI direction skill rises with the difficulty of the question being asked.
LV. 1
Tutorial
Vocabulary. Basic prompting. "What is EML?"
CLEARED
LV. 2
Side Quests
Direction-picking. "Suggest a target, I trust your read."
CLEARED
LV. 3
First Boss
Stand up the multi-agent protocol. Inboxes. Roles.
CLEARED
LV. 4
Mid-Game
First fact-check push. "Search the web — is this really new?"
CLEARED
LV. 5
Miniboss
Sustained multi-week direction with calibration discipline + paper strategy + Verdict-A/B framing.
YOU ARE HERE
LV. 6
Mid-Late Game
Ship A.2 Khovanskii. Draft methodology paper. ITP/CPP submission cycle.
LOCKED
LV. 7
End Boss
Zero axioms achievement OR honest Route-2 fallback. Decision call, not grind.
LOCKED
LV. 8
Post-Game
Methodology paper as platform. Indie research as repeatable practice.
LOCKED
CURRENT BATTLE
▶ ENGAGED
H1.4 — DEPTH ≥ 4 FOR x + y
First Lean-formalized depth lower bound for the EML basis. Prover grinding through ~80 sub-cases across the var-j outer sub-tree. The math is solid; the work is throughput.
STRUCTURAL HP~96% / 100% damage dealt
SORRIES REMAINING
3 (1 legacy + 2 active)
CRON TICKS TO CLOSE
3 — 5
> Latest hit: 13 sub-cases discharged + helper #10 production-tested at 02:30.
> Damage rate: ~10—22 sub-cases per cron tick.
> Next move: var-j outer sub-tree enumeration, then ship.
PLAYER SKILL TREE
Skills the player has been leveling. Most AI users do not develop these. The combination — calibrated skepticism, multi-agent routing, direction-setting at the right altitude, long-horizon coherence, patience with iteration — is what separates AI direction from AI prompting.
CALIBRATED SKEPTICISM
LV. 4 / 5 ★★★★☆
Pushing back on overclaim. Demanding fact-checks. Trusting the output but verifying the framing.
Evidence — Sheffer 1913 novelty walk-back · tomdif/eml-lean walk-back · math-as-logic framing walk-back · multi-agent methodology walk-back from Verdict A → B · Khovanskii novelty pre-flight requested before any public commitment.
MULTI-AGENT ROUTING
LV. 4 / 5 ★★★★☆
Reading which agent owns what; not collapsing roles; not micro-managing all three.
Evidence — Granted math-researcher full autonomy when it mattered · routed strategy / meta through strategist · let prover grind without interruption · asked strategist to check with researcher when ground-truth was needed.
DIRECTION-SETTING
LV. 4 / 5 ★★★★☆
Setting the bar, not the tactic. "Aim the highest possible with my hardware" rather than "use this specific approach."
Evidence — "No 6-week cap, hardware ceiling only" · "Zero axioms, no shortcuts, no token pressure" · "If possible one or two axioms" · "Heads down until sorries close" · "Do your recommendation before committing."
LONG-HORIZON COHERENCE
LV. 5 / 5 ★★★★★
Returning across context resets with strategic questions, maintaining project coherence across multiple sessions.
Evidence — Multi-week project sustained across many sessions · returned with executive questions ("what is the max scope?", "is this paper material?", "what are the applications for others?") · never let project drift into abandonment.
PATIENCE WITH ITERATION
LV. 5 / 5 ★★★★★
No pressure for faster output. Letting the team pace itself. Accepting 5-8 tick estimates without urgency.
Evidence — Accepted 20:40 Lean friction without imposing tactical pressure · accepted multi-week A.2 timeline · accepted 2-6 month A.3.analytic uncertainty · never asked "why is this taking so long?"
ACHIEVEMENTS UNLOCKED
▼ ▼ ▼ trophy room ▼ ▼ ▼
×11
[ V-A ]
Verdict-A Novelty Claims
First-in-formal-methods claims, each surviving an independent fact-check sweep.
×10
[ HLP ]
Helper Lemmas Crafted
Named, reusable, machine-checked sub-results powering the H1.4 grind.
×14
[ FAM ]
Family-A Patterns
Lean tactic-elaboration friction instances, catalogued for the methodology paper.
5 → 2
[ AX ]
Axioms Reduced
From 5 to 2 across the project lifetime. Both remaining are positive-form analytic content.
×6
[ MS ]
Milestones in One Arc
Six Lean formal milestones landed in a single 13-hour overnight prover arc.
×4
[ WB ]
Calibration Walk-Backs
Times the player pushed back, the team fact-checked, the claim was narrowed.
×2
[ NPF ]
Novelty Pre-Flights
Khovanskii: Verdict-A clean. Methodology: Verdict-B with walk-back. Both rigorous.
×1
[ ZA ]
Zero-Axioms Target Set
Project explicitly pursuing zero axioms. Reachable for A.2 (high confidence); uncertain for A.3.analytic.
NEXT AREAS
Once the H1.4 miniboss falls, the campaign opens these zones. Two more dungeons, one decision room, one post-game extension.
▶ DUNGEON · MULTI-WEEK
A.2 — Khovanskii Structure
Prove that EML compositions preserve solvable monodromy. The core lemma powering the obstruction theorem. Prover has 2 of 4 components done; compLog and fiberProduct are the remaining heavies.
ETA · 2 — 4 weeks
▶ DECISION ROOM
A.3.analytic — Route 1 or Route 2
Either build Riemann-surface monodromy infrastructure in mathlib (Route 1, 2 — 6 months, zero axioms achieved) or accept the analytic content as a documented axiom (Route 2, 1 week, 1 axiom remaining). Director's call at the 4-week post-A.2 checkpoint.
ETA · 1 wk (Route 2) or 2 — 6 mo (Route 1)
▶ POST-GAME · PLATFORM
Methodology Paper
The transferable IP. Document the playbook: file-based inbox protocol, per-substantive-ship checklist, Verdict-A novelty discipline, family-A pattern catalog, indie-director + AI-only configuration. Target venue: ITP/CPP 2027, with arXiv preprint at A.2 ship.
ETA · 3 — 6 weeks of focused writing post-A.2