shot‑5.1.1 · episode harmonizer — auditable data harmonization

ROOM 2 · SORT THE MULTI-DATASET INBOX

Build one policy-input collection.

Combine compatible runs. Match what each action means. Fix only what the evidence supports. When every required label agrees, the collection is ready to try baking—not proven to work.

Pick compatible runs0 episodes
Match meaningwaiting
Clear flags0 open
Check collectionnot ready

ONE NEXT MOVE Start with the practice pair Delivery + repair are preloaded. They share one planar control family.

V2 · DEBUG THE DATA

Treat each mismatch like a bug.

Reproduce it. Check the evidence. Apply a rule. Replay the result. A saved rule can suggest the next fix, but never applies itself.

SUGGEST ONLY

MATCHING RUNS 2 starter runs · 2,507 frames

delivery relay1,422 frames · planar 4D
repair assembly1,085 frames · planar 4D

LABEL CHECK Owner must check 4 fields

action unitsreference framecommand modetiming

RULE STACK · MIXED ROBOTS Never force a match.

If the robot, mode, units, or frame disagree, stop and ask the owner.

send matched evidence to policy lab →

Optional control sandbox · not episode evidence

Planar control practice

seed 42

agent object goal

Goal Move the object into the target region.

Observed The agent is scanning the seeded field.

Goal-conditioned feedback Approach from behind the object.

Executed action [0.00, 0.00, 0.00, 0.00]

Learn the 4D control family This sandbox does not create, replay, or change a source episode. WASD / arrow keys · drag · R reset

Step 1 · the candidate group

Combine runs that can learn together

START HERE · PRELOADED PRACTICE Delivery + repair make one planar starter.

They use the same 4D controller, so you can practice aligning meaning, units, frame, and mode. They are not claimed policy-eligible yet—the final collection check decides that.

Episode Satchel · check compatible evidence to combine

Sources will appear here.

Checked cards form one candidate collection. Episode boundaries stay separate. G1 is a boundary test: its 29D humanoid controls must not be mixed into the planar starter.

Inspect one episode’s receipt Optional detail · does not change the checked group open

Episode detail preview

This dropdown changes the receipt shown below; it does not add or remove episodes from the Satchel.

Robot type: —
Frames: —
Source FPS: —
Action width: —

Preflight Awaiting source receipt

Select an episode to inspect its mappings and unresolved semantics.

Preview + output settings read-only by default

Auto-refresh safe preview ON debounces local receipt and read-only batch-analysis checks when the Satchel changes. It never compiles or writes. ON · read-only

Maximum frames

LEARNING ANALYTICS INBOX · EVERY CHANGE IS REVERSIBLE

Clear the robot team’s mismatch inbox.

receipt preview

same run mail, same decision ledger. The sorting floor is a playful view of these exact receipt flags—not a second source of truth.

MOONBASE MAILROOM · LOCAL REVIEW SHIFT

Sort the run mail before launch.

File an evidence-backed proposal, ask the owner for an explicit label, keep the raw meaning unknown, or hold the letter. These choices update the inbox receipt only. Source episodes never change.

shift clock 01:00

filed0

held0

open mail0

Start a shift when you are ready. The clock pauses if this page is hidden.

WAITING AT THE ARRIVAL TUBE

No letter loaded

Start the shift to load a real flag from this inbox.

vitroovius question batch load an envelope that needs an owner answer Edit this question, then add the pink envelope to one shared outbox. No label is invented.

0 pink questions in outbox

review log loss accepted matches only · lower is better

waiting for a reviewed match

01
Auto-map safe fields Exact fields retain declared names, shapes, timing, and units.
→
02
Review flagged mappings Auto-inferred, lossy, or unresolved semantics wait here.
→
03
Validate combined LeRobot view Awaiting an explicit local harmonize action.

exact auto-inferred needs-review lossy abstained

Search this inbox

Rules 0

Propose-only rules match a field, compatibility signature, and source family. Nothing runs without your review.

No local rules saved.

Local command help · no AI call

show blocking show unresolved show all select action.units abstain selected clear selection list rules apply rule <id> to selected

Selected stack

All flags

Source as sender · field as subject · evidence as preview

local only

0 selected Choose a stack, select it, then apply one supported action.

Choose an episode detail preview to inspect its review queue.

Selected multimodal time-series · descriptive overview

What is actually in this Satchel?

local metadata preview

These views summarize declared metadata and optional local analysis output. “Unknown” is not treated as “missing,” correlations are descriptive—not causal—and schema clusters are deterministic signatures, not learned embeddings.

Modality + missingness matrix

present / external / flagged / absent / unknown

Select episodes to compare declared modalities.

Deterministic schema clusters

robot · action width · frame rate

No selected schema signatures.

PCA episode constellation

global PCA ↔ focus garden · structural similarity only

Select three or more episodes to build a structural map.

Selected-batch inbox feed

merged into the Harmonization Inbox above

No selected-batch flags yet.

Auditable transformation

From heterogeneous source episode to auditable LeRobot view

receipt —

Preserved source

RAW EPISODE

waiting for source

action—
timing—
language—

Explicit transforms

MAPPING RECEIPT

0exact

0inferred

0lossy

0unresolved

0preserved

Compiler decision

AWAITING RECEIPT

target schema —

No compatibility claim is made until all required semantics are resolved.

Adapter ledger

Field-by-field decisions

exact auto-inferred needs-review lossy abstained

Select a source to load its auditable mapping receipt.

Open complete field ledger 0 fields

Source field	Canonical field	Status	Evidence
No mapping loaded.

CUTTING TABLE · TRUTH, CRITIQUE, ADVICE, ACTION

Four-lane episode editor

Goal: keep the strongest aligned evidence and raise its harmonization evidence score.

0.00 s · f0

FIRST CUT · 3 MOVES

1Scrub an exact frameTap a lane or use ±1f. Every mark snaps to a stored parquet timestamp.
2Mark the useful evidenceSet in + out for a keep, or park the playhead where a split should begin.
3Check the prediction, then saveAll four lanes cut together. Save unlocks only when every required evidence gate passes.

Loading the local frame contract…

— / 100

HARMONIZATION EVIDENCE SCORE waiting for exact lane evidence

Editing aid only · never a policy-eligibility, causality, or safety claim.

source read-only action lane immutable derived video status pending exact frames pending

REQUIRED SAVE GATESScore what is actually recorded.

current full episode

◇usable frame rangetarget pending

◇persistent labelscoverage pending

◇event / critic coveragecoverage pending

◇action alignmentexactness pending

no range marked

1×

PREDICTED EDIT

mark a range or park the split frame —

No score is inferred until the exact local timeline has loaded.

Annotationdescriptive truth

Criticgoal comparison

Coachadvice shown

Executed actionimmutable evidence

Phone controls: drag sideways to pan, tap a track to scrub, use ±1f for exact frames, then set in/out. Source episodes are never overwritten; derived video is omitted and receipted as lossy.

FINAL CHECK

Build one trusted run set.

Keep the raw runs. Check names, shapes, timing, units, frames, and action meaning. Stop if the runs cannot share one valid LeRobot view.