Track A0 — The Klindt POC
This page operationalizes the direction from the David Klindt meeting and the post-call debrief ("David Klindt Post Meeting — Alex + Marco + Vincenzo + Antonio + Tom", 2026-07-07): don't collect blindly, don't wait for a data request — build a small proof of concept with a validation benchmark, then return with something concrete that could be tested or scaled. It precedes the Founding Experiment and de-risks it.
What this proves (and what it doesn't)
The POC is a data-value demo: a world model fails on a dynamics slice; fine-tuned with ~20–30 hours of Playroll data, it succeeds; here are the numbers. It is deliberately not the Founding Experiment's flywheel-value claim (directed beats passive at equal hours). Two different claims, two different buyers of the evidence: the POC reopens the door with David and any lab; the experiment prices the quest engine. Sequence them — the POC is cheaper, faster, and its benchmark harness is reused wholesale by the experiment.
David's "a couple dozen hours is enough" instinct matches the research: value concentrates in small, targeted, labeled seeds (Technical Evidence §A — VPT needed ~2K hours, arguably ~100 for most results), not volume.
The benchmark-first resolution
The debrief's benchmark-vs-samples debate dissolves into the architecture's two tiers (Engine Architecture §1):
- Antonio is right that the benchmark comes first: eval slices are defined and pre-registered before any collection — otherwise the samples define the test and the result is unfalsifiable.
- Vincenzo is right that the quest includes the benchmark: every quest ships with verifier code, so the quest itself is the falsifiable unit.
- Marco's protocol is the annealing pattern (measure base model → fine-tune on our hours → measure again), which is exactly the production data-valuation method this engine standardizes on.
Why Minecraft — testbed, not product
Minecraft is the right POC substrate for reasons the debrief already named, plus one caveat it must not lose:
Right testbed: open pretrained world models exist to fine-tune cheaply (Decart's open Oasis 500M weights, Microsoft's MineWorld, Matrix-Game's Minecraft slice); structured state / G-buffer access makes verifiers and ground truth easy; simple construction/destruction tasks are legible dynamics probes; the surrounding literature (VPT, MineRL, PLAICraft) gives baselines and recipes.
Wrong product: Minecraft is the only game where open action-labeled data is already deep (Competition & Moat §open datasets) — Minecraft data itself is a commodity. We prove the method on Minecraft and sell the method on the cross-game catalogue. No POC result should be pitched as "our Minecraft data is valuable"; the pitch is "our pipeline turns targeted player-hours into measured model deltas — here, on the most commoditized game there is; imagine it on the 19 others."
Protocol, step by step
| # | Step | Detail | Owner (per debrief) | Done when |
|---|---|---|---|---|
| 1 | Survey David's shared papers/repos | Map what harnesses already exist ("is there a repo where we throw our data in?") vs what we must build; extract his lab's eval conventions for dynamics | Antonio (~2 days, committed on the call) | Short memo: chosen base model + harness + gaps |
| 2 | Select base model + eval harness | Default candidates: MineWorld or open Oasis weights (Matrix-Game Minecraft slice as fallback). Criterion: cheapest credible fine-tune with reproducible eval | Antonio + Marco | Base checkpoint runs locally; eval loop reproduces a published number |
| 3 | Define and pre-register the benchmark | Probe candidates P1–P8 in the B-04 survey, plus (per the vault review): P9 geometry-consistency (generated rollouts vs RCON world geometry — the probe that connects to Anto's provisional-IP 3DGS/engine-ground-truth work and feeds the offsite splat artifact) and P10 static/dynamic subject-isolation (control; Anto's recurring theme). Freeze metrics, thresholds, and failure examples before collection. Position P1–P10 as the Minecraft instantiation of Guido's 288-combination cross-engine benchmark and the substrate for David's stress-test benchmark — one benchmark program, three stakeholders, not three efforts | Antonio (design), Marco (review) | Pre-registration doc merged; base-model failure cases documented with numbers |
| 4 | Author 2 quests with verifiers | From the debrief: a construction task (e.g. build a 9×9 wall) and a destruction/collection task. Verifiers via structured state / VLM detection. Deliberately minimal scaffolding — per Tom's point, novelty lives in how players execute (build order, corrections, mistakes); over-guidance would flatten exactly the variance a dynamics benchmark needs (instruction-bias risk §2) | Marco + Vincenzo (quest/DSL), Tom (VLM verifier feasibility) | Quests exist in the requirements DSL; verifiers pass on pilot runs |
| 5 | Collect ~20–30 h with operators | Constrained runs under the locked capture spec (Phase 0 0.1); operators, not open fleet — this is a controlled sample, not a campaign. Mix: quest-directed hours + free-play control hours + an explicit exploration quota (non-goal wandering — David's rat-labyrinth bias point) + hypothesis-test narration on the rider subset ("I think X triggers Y" → observed outcome, as first-class labels — David's offline-RL suggestion, Gordon-corroborated). ⚠ Dependency: Minecraft Java ships outside every supported launcher lane — recorder/detector enablement is unbudgeted engineering work that must precede this step (see catalog overview lane policy). ~27 Minecraft sessions ever recorded, none since April — fresh staffing (~3–4 dedicated operator-days), scheduled around the July 15–20 offsite | Vincenzo (pipeline + recorder enablement), Phil (operators) | Hours banked, verified, provenance-clean |
| 6 | Fine-tune | Same recipe discipline as the founding experiment: identical base checkpoint and compute across arms/ablations; hold out a slice of collected data for validation | Marco + Antonio | Fine-tuned checkpoint(s) + training log |
| 7 | Evaluate against the pre-registered benchmark | Report base-vs-fine-tuned deltas per probe. Honest reporting: include probes that didn't move | Antonio | Results table; every claim traceable to the pre-registration |
| 8 | Package and return to David | Short writeup (benchmark, data card, deltas, what we'd scale) plus the buyer-native data package — David's other named next step from 06-26 was output format ("USB-for-AI"/MCP framing; RLDS/LeRobot-style packaging per the wiki's buyer-native posture). FS-001 elevated (2026-07-08): per the open-platform vision, this format spec is the draft of the platform's dataset packaging standard — the thing that makes "browse it like Hugging Face" real — not just David's deliverable. Dr. R owes FS-001; commission it now. The motion: "you suggested it, we built it, want to test it on your models?" | Alex (relationship), Antonio (content), Vincenzo (packaging) | Sent with the data package; follow-up meeting requested |
Optional sub-experiment: overlay narration
Riding on step 5 at near-zero cost (Tom's suggestion on the call): prompt a subset of operators via overlay to narrate intent while playing ("building the base first"). Measures (a) annoyance/friction of overlay prompts, and (b) whether elicited narration is usable intentionality data — axis A in Quests as a Data Product, the measured +28% ECoT lever, and the data asset reconstruction can't touch. Keep it out of the benchmark arms so it can't confound the dynamics result.
Decision points this POC feeds
- Single-game vs multi-game (Antonio vs Vincenzo on the call): the POC is single-game by design (benchmark risk is the binding constraint — Marco's argument). Vincenzo's diversification concern is honored one level up: if the Minecraft benchmark proves hard to build, the fallback title comes from the Phase 0 title set (0.8), and the Founding Experiment is multi-title anyway.
- Robotics-flavored vs game-gen-flavored quests: David's frame (simple dynamics, constrained tasks) is the robotics-benchmark style; the cross-game catalogue serves the game-generation buyers. Both ride the same quest framework with different requirements — Vincenzo's point, and already the two-leg position. Per the verdict rule, still no manipulation/force quests — constrained navigation/construction dynamics only.
Gate
Track A0 is done when the writeup exists with at least one pre-registered probe showing a clear base→fine-tuned improvement, and it has been sent to David. If no probe moves: treat as tripwire-grade information — interrogate benchmark design first (most likely failure), data spec second, and only then the thesis; the Founding Experiment does not start until A0's post-mortem is written.
Attribution & lineage (vault-verified, 2026-07-08)
The benchmark-first shape is David's own proposal (2026-06-26 main call: "we should make a new benchmark… stress test data… almost like red teaming for world models"), and his verbatim asks — output format, coverage steering, learnable novelty, hypothesis-test structure — are all now in the protocol above. The specifics (Oasis-500M/MineWorld, hour counts, the probe designs) are our synthesis, not his words — and he remains "a warm design-partner offer, not a signed partnership" until terms are agreed. RCON-as-ground-truth is Anto's GTA engine-truth doctrine transplanted to Minecraft; P1–P8 measure generative-model dynamics and do not instantiate his provisional-IP geometry layer (P9 is the probe that genuinely connects to it). The framing discipline from the call stands: fine-tuning demonstrates our data's quality — we are not building competing world models ("if there's one thing in AI you don't want to do, it's compete with DeepMind").
Competitive context (why now)
The competitor David referenced is almost certainly Worldmodeldata — £7M seed announced the day before the call, Cambridge, chaired by Lord Richard Allan (Market & Buyers timeline). Their existence is validation and clock: they license from developers and target 1M hours by 2027; our differentiation is the directed player fleet and measured deltas — which is exactly what A0 starts evidencing.