Minecraft behavioral-vector pilot

Labnote 001: Minecraft behavioral-vector pilot🔗

Question🔗

Can repeated, task-balanced Minecraft sessions produce an interpretable behavioral vector whose within-condition variation is smaller than its uncertainty?

Motivation🔗

Open-ended play exposes choices that short text prompts often compress away: exploration versus exploitation, planning before action, persistence after a setback, resource allocation, and the tendency to revise or abandon a plan. Minecraft offers a controllable environment for measuring those choices across repeated tasks.

The resulting vector is a description of behavior observed under this protocol. It is not a biometric, an identity claim, a personality test, or evidence about behavior outside the measured sessions.

Candidate coordinates🔗

The pilot may evaluate these operational coordinates after the event contract and task battery are frozen:

• exploration breadth;
• plan depth and revision rate;
• resource selectivity;
• persistence and recovery after setbacks;
• construction versus acquisition allocation;
• route efficiency relative to task completion; and
• voluntary marker timing and confidence.

Coordinates that cannot be defined from observable events before collection will be removed rather than inferred afterward.

Collection boundary🔗

• Only the toolkit-owned Voxel Guidance Lab instance may emit study data.
• Recording is explicit per session and visibly indicates its current state.
• The bridge must not collect chat, screenshots, microphone input, keystrokes,

account identifiers, server addresses, or activity outside the game world.

• World and session identifiers are random experiment-local values.
• Raw positions and event streams remain private and are excluded from public

artifacts.

• Human markers are structured experiment events, not unrestricted journal

text.

Staged protocol🔗

Phase 0: contract rehearsal🔗

Validate start/stop behavior, event ordering, timestamps, session separation, marker delivery, crash recovery, and deterministic export. This phase makes no scientific claim and its data cannot enter the pilot analysis.

Phase 1: task battery freeze🔗

Define a small set of matched tasks spanning exploration, acquisition, construction, interruption, and recovery. Freeze task seeds or world fixtures, time limits, allowed assistance, marker prompts, exclusions, and the analysis code before measured play begins.

Phase 2: repeated collection🔗

Collect multiple sessions per task condition. Session order should be rotated when practical. Practice and interrupted sessions remain visible in the ledger but are excluded under predeclared rules.

Phase 3: compilation🔗

Aggregate events into task-normalized coordinates. Report per-session values, within-condition variation, bootstrap intervals over sessions, and leave-one- task-out sensitivity. Do not collapse the vector to a single score.

Advancement gate🔗

The pilot earns interpretation only if:

1. the contract rehearsal has no unexplained event loss or cross-session data;
2. every retained coordinate has a frozen operational definition;
3. at least eight usable measured sessions exist across the task battery;
4. uncertainty and leave-one-task-out sensitivity are reported; and
5. conclusions remain explicitly protocol-bound.

Otherwise the outcome is negative or inconclusive and preserved as such.

Current decision🔗

The version-1 event contract and transparent feature compiler pass their contract tests. The Fabric bridge built and initialized in an isolated live launcher rehearsal on 2026-09-07; recording remained off and no world was opened. A subsequent Accessories/Trinkets compatibility problem in the restored convenience-mod set was corrected before the in-world rehearsal.

First in-world rehearsal🔗

After the mod compatibility issue was corrected, a private in-world rehearsal completed successfully:

• duration: 382.572 seconds;
• 385 contiguous events: start, 382 position samples, one controlled marker,

and a clean stopped boundary;

• event-file permissions: owner read/write only (0600);
• sampling cadence: approximately one position sample per second;
• derived spatial summary: 50 coarse 16-block cells, 808.50 blocks of path

distance, and 0.513 route directness; and

• no raw coordinates or route map were promoted into the research record.

Second in-world rehearsal🔗

A separate rehearsal-002 stream also validated successfully:

• duration: 194.992 seconds;
• 197 contiguous events: start, 195 position samples, and a clean stopped

boundary;

• event-file permissions: owner read/write only (0600);
• derived spatial summary: 7 coarse cells, 223.37 blocks of path distance, and

0.005 route directness; and

• a distinct session identity and task identity, with no event mixing between

the two rehearsals.

The second session intentionally contained no marker. It exposed an ambiguity in the compiler: missing marker latency had been encoded as 1.0. The compiler now emits null for an unobserved first marker, and the session-set validator rejects duplicate session identities.

These are instrumentation results only. Inventory, block-action, damage, and recovery coordinates remain zero because the rehearsal bridge does not emit those events yet. Multi-session separation passed here; termination behavior was tested separately below.

World-exit rehearsal🔗

In rehearsal-003, recording was started and the world was exited without an explicit /vg stop. The bridge received the graceful disconnect event and produced:

• a 5.969-second owner-private stream;
• one start, five position samples, and exactly one clean session end;
• end reason stopped;
• a third distinct session identity with no cross-session events; and
• valid compilation with marker latency left unobserved (null).

This validates automatic closure on ordinary world exit. It does not validate hard-crash behavior: the game remained able to execute its disconnect handler.

Managed debug-crash rehearsal🔗

A second recording reused the human-facing rehearsal-003 task ID and was terminated with Minecraft's F3+C debug crash. It created a new UUID-backed file rather than overwriting the earlier rehearsal. The 39-event stream retained owner-only permissions, contiguous sequencing, and exactly one end boundary.

F3+C also performs enough orderly shutdown to deliver Fabric's disconnect callback. The old bridge therefore labeled the end stopped, which could not distinguish explicit stop, world exit, and managed crash. Future bridge output uses explicit_stop for /vg stop and the cause-neutral disconnected for the callback. Historical stopped streams remain valid. Hard termination without a callback was therefore tested separately.

Hard-termination rehearsal🔗

In rehearsal-004, the Minecraft window was terminated with xkill while recording. The owner-private stream contains one start and 15 contiguous position samples, ending at sequence 15 with no session end. The validator rejects it specifically because its session boundary is incomplete. No repair, usable feature vector, or synthetic crash end is produced, and the four earlier recordings remain unchanged.

Phase 0 is complete. Hard-termination handling is deliberately fail-closed: "recovery" means retaining an auditable incomplete stream while excluding it from compilation, not guessing how or when the process died. Feature compilation is deterministic under the same validated event input.

Full event-surface rehearsal🔗

rehearsal-005 exercised the expanded bridge in one 76.046-second session. Its 120 contiguous events included 76 positions, 25 coarse inventory deltas, 11 world-confirmed block actions, three health-loss observations, one death, one later respawn, one setback marker, and a clean explicit_stop. All event channels validated under the version-1 contract.

The death rehearsal exposed a second analysis confound: inventory removal and recollection around death initially inflated acquisition denominators. The compiler now excludes inventory deltas from a death until a controlled recovered marker. Nineteen such events were excluded in this rehearsal and reported as a quality diagnostic. The corrected construction-allocation value was 0.138 rather than the confounded 0.037. Because no recovery marker was issued, no post-death inventory activity entered acquisition metrics.

This rehearsal remains apparatus evidence only. Exact item/block identities, block coordinates, attackers, and raw routes were never emitted.

Frozen pilot design🔗

The pre-collection protocol is frozen at digest 320e6bf8a8e2a3978e6125da3be5c2e0b4db52b1e8bea869b41b17225c65877c. The live 22-mod apparatus exactly matched manifest digest ebcb2db3723b6398b696d7b117b5a5dcee81242ed677d2a3cb97bb3b284acb59 at freeze time.

The battery contains 12 eight-minute sessions: exploration, acquisition, construction, and controlled recovery tasks across three seed blocks. Worlds within a block share a seed but remain separate saves. Task order is rotated across blocks to reduce simple order effects. Sessions under seven minutes, over ten minutes, lacking an explicit stop, or missing required markers are excluded under predeclared rules.

The compiled fingerprint has seven coordinates: exploration breadth, route closure, plan revision rate, resource selectivity, construction allocation, recovery efficiency, and deliberation latency. Each coordinate is aggregated at the seed-block level, receives a 95% block-bootstrap interval, and reports a leave-one-block-out range. Missing values remain null; there is no learned embedding, imputation, or single combined score.

The compiler refuses incomplete or duplicate session matrices and emits no session- or block-level vectors. Public projection is limited to aggregate coordinates, uncertainty, exclusions, protocol and apparatus digests, and the claim boundary.

The event sources, task battery, and analysis plan are frozen. Pilot collection may begin in protocol order. None of the rehearsal streams may enter evidence.

Pilot completion🔗

Collection finished on 2026-09-10 with 12 accepted sessions across all four tasks and three seed blocks. Interrupted and superseded attempts remain visible in the owner-safe collection ledger and were excluded before compilation. The deterministic aggregate contains all seven frozen coordinates, block-bootstrap intervals, and leave-one-block-out ranges without a single combined score.

This is a mixed result: the protocol produced an interpretable aggregate and earned exploratory follow-up, but three blocks from one participant cannot establish a stable fingerprint beyond this task battery. Labnotes 002 and 003 therefore test protocol-bound task recognition and feature-family contribution without making identity, personality, or population claims.