Implementation revision · September 13, 2026

A functioning self that wants to discover what it can do.

Wellness is a perception of the self: its body, usable senses, reserves, capabilities and learned relations. Intention should change that perceived future, including gaining reliable capabilities while the body is already comfortable.

This plan replaces the deficit-only interpretation. It specifies mechanisms, implementation boundaries, failure tests and the minimum environments for every ability. It does not equate working software with a developed being.

What is done and what remains

Training is running. Purposeful reaching and rolling have not yet passed our tests.

Done

Training now uses where the body is touched and how well it can sense and move.
Saved learning and body state survived the checked upgrades and restarts.
We repaired a separate model that predicts what a move will do and touch. Its code tests pass.
The dashboard shows experiments and results. This plan links the records needed to continue the work.

Still to do

Find out whether the repaired predictions are accurate, including in unfamiliar situations.
Show that they help the model choose movements that maintain its body or gain a new ability.
Prove purposeful actions, such as reaching or rolling, before giving the new system control.
Finish testing whether the caregiver helps learning. Test scheduled stillness once enough recording space is available.

The original controller still moves the body. The repaired predictor has not yet passed its movement-and-touch experiment.

Status checked 2026-09-14T22:59:05-07:00. Parked by your decision until the baby can act. All its finished work is kept.

Progress to date and handoff

Work remains in progress. Updated 2026-09-14T22:59:05-07:00. This is a dated record. Check the live supervisor and files before acting.

H83 is the production run: resumed 2026-09-14 from H82 final snapshot 4083925 with the scheduled rest holds dropped (operator decision, grounded in the corrected four-pair stillness reading). The corrected reading also RETRACTS an earlier chat claim of a first learned skill: against the same-probe untouched-brain control, no twin learned toy-tracking; the held twin lost nearly all of what the untouched brain had in the one informative world. Caregiver study parked by operator until the baby can act; its partials and 8 completed fits preserved. Stillness full study superseded by the operators decision; its preserved pairs are the record. Predictor v2 development negative stands; input-signal probe still queued. Purposeful behavior remains unproven; 0 of 8 registry abilities passed.

Continuation boundary: H82 was deliberately ended by operator instruction on 2026-09-14 with snapshot_and_stop at step 4083925 (complete snapshot verified: all six required files). H83 resumed from that exact snapshot with ONE change: --rest-control choice (the scheduled rest holds are dropped, operator decision 2026-09-14, grounds in docs/experiment_review_2026-09-13/stillness_preliminary_reading_20260914.md). Launch command receipt: Saved/mimo_bridge/h83_launch_command_20260914.json. NO crash-restart supervisor is attached to H83; the overnight watcher reworks restart mechanism is unconfirmed - attach or confirm before relying on automatic relaunch.

Complete machine-readable handoff · Short completion status

What is actually live

Last verified learning update: 4083925 at 2026-09-14T22:59:05-07:00
Body control: The original controller moves the body; the new controller only watches.
Sensing and self learning: Training uses local touch and spatial body information to learn about the body's condition and capabilities. Exact settings are in the linked handoff.

Read before continuing

Keep one production trainer. Preserve the newest complete learned and physical state. An earlier proof snapshot is not the source for restarting a newer live run.

The new chooser is an observer. A software test or auxiliary prediction score does not grant it control of the body.

What follows the predictor: confirm it on new worlds, check its predicted changes to the perceived self, prove better choices in copied brains, then require visible purposeful behavior before verified adoption and retention checks. Detailed sequence, missing connections and failure branches.

Project memory: the editable resume_handoff.json is the persistent handoff memory for this work. The workstream directory indexes the supporting notes. No separate assistant memory file has been created for this session; a replacement LLM should not rely on conversation history or unstored agent state.

Completed work and its limits

Reworked the implementation plan, all 25 ability requirements across four tiers, gain tree, crib explanation and feasibility review. Reviewed the UEMA patent, all 18 Omnibus DOCX extracted texts, prior external review and UE transfer interfaces.
Grounded spatial self learning and local skin modulation deployed. Latest production migration preserved 390 earlier tensors, 328 Adam histories, all eight continuing physical worlds and runtime/RNG evidence.
Independent acting/learning state implemented. New chooser is an unqualified observer. Added stable execution identity, reference-cache invalidation, per-body evidence identity and atomic complete checkpoint publication.
Dashboard, old dashboard, history and actual video playback verified. Twelve corrected native reports attach observations to15 tier entries; none is an acquisition pass. Added current-work card and local-skin deployment receipt.
Automatic checkpoint retention is active and protects recovery plus active experiment sources. Authorized alternating old snapshots/videos removed, including2.285GB of older copied proof snapshots on C and1.338GB of older intermediate stillness smoke snapshots on D.
Located invalid complete-learner interpretation of the old audit: frozen encoder plus continuing predictor; first 50 ms of 250 ms held orders. It cannot license action selection or establish forgetting.
Outer loaded-function binding, exact shadow-only migration and independent eight-body native save/reload proof completed. Production H81 now continues from H80 final step 4018725. This protects future execution identity; it does not certify the earlier unbound interval or grant action qualification.
Resource-pause continuation for copied caregiver/stillness fits is implemented. The preserved native pixel-equality failure was localized to a few renderer channels differing by 1/255 despite exact physical state. A separate identical-input native comparison passed all 13 learning/optimizer/RNG/command/physical preservation checks. The original caregiver study resumed; inspect its current status below rather than treating historical continuation checks as a current running status.
Separate raw causal readout completed80 matched situations, exact replay/encoding checks and all four declared fits. Original model and Adam state stayed exact. All20 untouched-world/group comparisons failed action resolution. Body position contrast0.230-0.331cm versus absolute prediction error0.432-0.564cm; pressure contrast0.196-0.517kPa versus error2.33-7.77kPa. This is a bounded new-readout failure, not proof of model incapability.
Corrected future checkpoint parameter inventory: native H81 has5728601 parameters plus33240 buffer values, with no missing weights. Nine writer/inventory/atomic-publication tests pass. Existing saved receipts remain untouched; running H81 picks up this metadata-only writer change at its next process start.

Work underway and the exact next steps

acting_learner

Completed checkpoint metadata audit and storage eligibility review.

Last recorded state: Complete: core actor/continuation, metadata and storage reviews; independent v2 review closed with8adversarialtests. New checkpoint inventory applies at next normal H81process start; no weights were missing.

Implementation: Source/Python/lem/training/checkpoint.py
Evidence: Saved/mimo_bridge/H81_checkpoint_parameter_audit_20260913.json; Saved/mimo_bridge/receipt_inventory_tests_20260913b.xml

tiers_research

Exact interrupted caregiver/stillness fit continuation, resource backoff, original study resume and truthful dashboard reports.

Last recorded state: Caregiver source refusals RESOLVED: pin handles the auxiliary edit; the residual mismatch was config_sha256 hashing the raw --checkpoint string (absolute at fit creation, relative in the readiness-note command; measured f8bd2c53 vs saved c6c395cb, reproduced both ways; in-vivo diff log latest_diff.jsonl). Resume relaunched with absolute path + pin (receipt caregiver_resume_pinned_20260913c_receipt.json), worker admitted: yoked_0 resumed at exactly 1,610 updates, verified learning at 1,651/9,000 with finite loss (segment_0002, worker pid 20052). Standing finding: canonicalize binding paths AFTER this study completes; the module is inside the pinned binding now. 8/24 copies complete; 1,610-update partial intact.

Implementation: Source/Python/mimo_bridge/tools/caregiver_fit_continuation.py; Source/Python/mimo_bridge/tools/caregiver_fit_workers.py; Source/Python/mimo_bridge/tools/prove_caregiver_fit_continuation.py; Source/Python/mimo_bridge/copied_execution.py; Source/Python/mimo_bridge/tools/caregiver_source_pin.py (audited; in use by the resumed study via sitecustomize startup pin)
Evidence: docs/experiment_review_2026-09-13/study_runner_and_dashboard_readiness.md

self_learning

Implement separate v2 auxiliary readout repairs and development-only factorial; preserve frozen v1.

Last recorded state: V2 development factorial completed 2026-09-13: development_only verdict; every condition action-blind in-sample (best single cell 0.0064 on the 0-to-1 action-error-removed scale); effect sizes a median 2950x the numeric floor; flat across budgets 40-240. Located class: model fitting/learning coverage - unused action input. Frozen encoder and v1 artifacts stayed exact. Next design choice is open; no v3 declared. See docs/experiment_review_2026-09-13/causal_readout_v2_development_result.md.

Implementation: Source/Python/lem/model/causal_readout_v2.py; Source/Python/mimo_bridge/tools/causal_readout_v2.py; Source/Python/mimo_bridge/tests/test_causal_readout_v2.py; Source/Python/mimo_bridge/tests/test_causal_readout_v2_review.py
Evidence: docs/experiment_review_2026-09-13/causal_readout_v2_development_plan.md; docs/experiment_review_2026-09-13/causal_readout_v2_verification.json; docs/Causal_Readout_V2_Independent_Review_2026-09-13.md; docs/experiment_review_2026-09-13/causal_readout_v2_development_result.md; Saved/mimo_bridge/causal_readout_v2_development_20260913a/report.json

Read current supervisor, live stats, caregiver study status and measurement queue first. Keep one production trainer. H81 migration is complete; never restore an older proof source over its newer learning.
Commits c39d3a1ee and e1e8b365d are pushed to origin/feature/mimo-bridge (user-approved 2026-09-13). Pre-existing tracked snapshot deletions and unrelated untracked files remain deliberately uncommitted; commit future work in batches with the same exclusions.
COMPLETED 2026-09-13: the declared v2 development factorial ran and failed - all four conditions action-blind in-sample; located class is model fitting (unused action input), not information (effects 2950x the numeric floor) or budget (flat 40-240 updates). Result note: docs/experiment_review_2026-09-13/causal_readout_v2_development_result.md. Choosing the next predictor design is OPEN operator-facing work; no v3 is declared and fresh confirmation worlds remain unconsumed.
After development, freeze the selected predictor and declare new confirmation worlds before capture. Then validate predicted changes to the perceived self, copied choice benefit, visible purposeful behavior and retained learning before adoption. Follow docs/experiment_review_2026-09-13/after_predictor_implementation_sequence.md; body/pressure accuracy alone is not a wellness or choice pass.
Caregiver study PARKED by operator 2026-09-14 until the baby demonstrates action; partials and 8 completed fits preserved on D:. Do not relaunch without operator instruction.
Stillness question ANSWERED for now by operator decision: holds dropped in H83 based on the corrected four-pair reading (docs/experiment_review_2026-09-13/stillness_preliminary_reading_20260914.md). The preserved pairs are the record; no further stillness fits without a new declaration. The 24-degree in-view metric weakness recorded there applies to any future gaze study.
The 32 experiment originals were retired 2026-09-13 with operator authorization after per-file SHA re-verification; archives and .retired.json receipts remain on D:. The operator still needs to upload the archives (2.34 GiB) to Google Drive; the manifest is beside them and in Drive. Recheck free space and retention before further study launches.
Update resume_handoff.json, workstream_documents.json and completion status after each result. Register every new worker note/review/report and any newly created memory file. Rebuild the main HTML plan so another LLM can resume from these files without conversation history.

Find each workstream's documents

Start with the first note in each group, then its latest evidence. Historical failures remain linked and must not be treated as current passes.

Editable document directory. When a worker adds or updates a project record, register it here and rebuild this plan.

Acting learner and checkpoint preservation (14 records)

The acting_learner workstream owns the acting/learning split, code identity, saved-state preservation and independent readout reviews. Start with the overview and current decorated-continuation note.

Document or record	What to use it for	Scope
Acting and learning overview `docs/Acting_While_Learning_2026-09-13.md`	Architecture, responsibilities and preservation requirements.	Design and implementation
Current continuation protection `docs/Decorated_Actor_Continuation_2026-09-13.md`	Outer loaded-function checks, completed H80-to-H81 migration, present limits and historical commands.	Current; migration already completed
Original execution identity `docs/Acting_Execution_Identity_2026-09-13.md`	Inner code-identity contract; read the decorated-continuation note for later protection.	Earlier foundation
Independent migration review `docs/decorated_actor_migration_review_2026-09-13.md`	Adversarial review of actor migration and exact saved-state handling.	Review
Checkpoint retention rules `docs/Checkpoint_Retention_2026-09-13.md`	Registered storage, active sources, recovery protection and alternating older snapshots.	Current implementation; check live policy/status
Eight-body native continuation proof `Saved/mimo_bridge/native_decorated_validation_20260913/report.json`	Two fresh processes, real updates, complete save/reload and source preservation.	Passed integration; no ability or choice license
Earlier local-skin continuation proof `Saved/mimo_bridge/native_local_shadow_validation_20260913b/report.json`	Preceding native startup/save/resume validation.	Historical successful proof
Earlier audit-state carry proof `Saved/mimo_bridge/native_local_shadow_audit_20260913/report.json`	Preserved pending and completed observing-audit evidence.	Historical integration result
Original fingerprint failure `Saved/mimo_bridge/native_local_shadow_validation_20260913/actor_fingerprint_diagnostic.json`	Retained diagnosis before corrected canonical identity.	Historical failure; not a current pass
Actual saved parameter inventory `Saved/mimo_bridge/H81_checkpoint_parameter_audit_20260913.json`	Shows complete weights and the stale earlier descriptive count.	Read-only audit
Metadata-writer verification `Saved/mimo_bridge/receipt_inventory_tests_20260913b.xml`	Nine inventory and checkpoint publication tests.	Passed software tests
Live retention status `Saved/mimo_bridge/checkpoint_retention_status.json`	Actual latest cleanup result, protected bytes and free space.	Live file; reread before acting
Completed duplicate-proof retirement `Saved/mimo_bridge/completed_duplicate_proof_retirement_20260913.json`	Exact retired old source and retained copies, hashes and limitations.	Storage operation receipt
Independent first-readout review `docs/Causal_Readout_Independent_Review_2026-09-13.md`	Packet, timing, leakage, source and gate checks before the first study.	Review of frozen v1

Self learning, sensing and action prediction (22 records)

The self_learning workstream owns grounded self learning, the corrected sensing/action assay and the separate predictor repairs. The first predictor failed; its examined worlds are development data for later work.

Document or record	What to use it for	Scope
Current first-readout explanation `docs/Matched_Causal_Readout_2026-09-13.md`	What was implemented and what the completed negative establishes.	Implemented and measured; no adoption
V2 repair plan and exact commands `docs/experiment_review_2026-09-13/causal_readout_v2_development_plan.md`	Achieved body coordinates, coupled movements, shared skin-site learning, development factorial and fresh-confirmation boundary.	Separate auxiliary; see current declaration
What follows the predictor `docs/experiment_review_2026-09-13/after_predictor_implementation_sequence.md`	Confirmation, predicted self consequences, choice trials, visible behavior, adoption and failure branches.	Next steps; not completed claims
V2 implementation verification `docs/experiment_review_2026-09-13/causal_readout_v2_verification.json`	Combined software tests and actual recorded-packet learning/save/reload with frozen source preservation.	Implemented; development experiment not yet run
Why the old audit could mislead `docs/Opportunity_Timing_and_Learning_Validity_2026-09-13.md`	First 50 ms versus held 250 ms action; old encoder/new predictor mismatch.	Located measurement limitation
Earlier self-learning repair notes `docs/self_learning_repair_2026-09-13.md`	Earlier grounding work and the reasoning behind its implementation.	Background; use current live receipts
Local skin and shadow activation `docs/Local_Skin_Shadow_Activation_2026-09-13.md`	Preservation checks for grounded sensing and the unapplied chooser.	Deployment note
Skin mechanics assumptions `docs/virtual_skin_mechanics_2026-09-13.md`	Local pressure, resolution, contact calibration and virtual-body limits.	H79 run references are historical
UE transfer requirements `docs/UE_Transfer_Contract_2026-09-13.md`	Coordinate, units, location and regional-sensitivity requirements for the richer avatar.	Current interface review
Frozen first-study review `docs/experiment_review_2026-09-13/causal_readout_review.md`	Errors in physical units, located input/objective/site-coverage limitations and next experiment.	Completed negative; no refitting
Exact saved-result reproduction `docs/experiment_review_2026-09-13/causal_readout_frozen_review.json`	Reproduces reserved-world scores and measures original training packets without fitting.	Read-only diagnostic
Learning-objective decomposition `docs/experiment_review_2026-09-13/causal_readout_objective_decomposition.json`	How much of the old objective described differences between commands.	Diagnostic of frozen v1
Contact-site training coverage `docs/experiment_review_2026-09-13/causal_readout_channel_coverage.json`	Sites lacking changing training examples.	Diagnostic of frozen v1
Dashboard result binding `docs/experiment_review_2026-09-13/causal_readout_dashboard_review.json`	Reviewed conclusion bound to the unchanged report SHA.	Published review
First-study declaration `Saved/mimo_bridge/causal_readout_study_20260913b/declaration.json`	Original worlds, conditions, budgets and gates.	Frozen before measurement
First-study launch receipt `Saved/mimo_bridge/causal_readout_study_20260913b/launch.json`	Exact invocation and exited worker identity.	Completed run
Raw sensory packet index `Saved/mimo_bridge/causal_readout_study_20260913b/packet_index.json`	All 80 immutable situations and the recorded inputs/outcomes.	Preserved source data
First-study full result `Saved/mimo_bridge/causal_readout_study_20260913b/report.json`	All held-out comparisons, source preservation and qualification exclusions.	Failed accuracy gate
Independent v2 review `docs/Causal_Readout_V2_Independent_Review_2026-09-13.md`	Cross-handle coupling, input/source integrity and optimizer validity.	Review before development fitting
V2 development declaration `Saved/mimo_bridge/causal_readout_v2_development_20260913a/declaration.json`	Fixed factorial, budgets and development-only status; fresh confirmation excluded.	Declared before the completed run
V2 development result and located failure `docs/experiment_review_2026-09-13/causal_readout_v2_development_result.md`	All four repairs stayed action-blind in-sample; information and budget ruled out by the predeclared controls.	Completed development negative; no v3 declared
V2 development full report `Saved/mimo_bridge/causal_readout_v2_development_20260913a/report.json`	All 4,800 comparisons, integrity fields and per-condition controls.	Completed run; development only

Tier research, environment and experiment continuation (20 records)

The tiers_research workstream owns developmental evidence review, caregiver/stillness comparisons, resource scheduling and copied-fit continuation. Start with the readiness note for actual running state and next steps.

Document or record	What to use it for	Scope
Current study readiness and next commands `docs/experiment_review_2026-09-13/study_runner_and_dashboard_readiness.md`	Current caregiver launch, copied-state proof scopes, rendering floor, resources and stillness prerequisites.	Primary current working note
Tier researcher working-document directory `docs/experiment_review_2026-09-13/tiers_and_studies_working_directory.json`	Worker-maintained list of all current studies, failures, continuation checks and external live files.	Canonical worker index
Caregiver monitoring repair and restart status `docs/experiment_review_2026-09-13/caregiver_status_io_repair.md`	Monitoring collision, preserved partial learning, restricted repair and source-version investigation.	Current continuation follow-up
All gain-of-function requirements `docs/gain_function_2026-09-13/requirements.json`	All 25 abilities, dependencies, minimal environments and evidence requirements.	Tier specification
UEMA and developmental research mapping `docs/gain_function_2026-09-13/research_and_spec_mapping.html`	Patent/specification connections, infant comparisons and stated uncertainty.	Research and specification
Source inventory and reading scope `docs/gain_function_2026-09-13/source_inventory.json`	Enumerated UEMA/Omnibus sources, hashes, coverage and relevance.	Source trail
Earlier experiment review `docs/experiment_review_2026-09-13/review.md`	Earlier experiment and instrument findings; newer readiness note takes precedence for operations.	Historical review
Completed behavior review `docs/experiment_review_2026-09-13/completed_behavior_review.html`	Readable linked assessment of corrected native behavior observations.	Measured observations; no acquisition pass
Individual behavior review directory `docs/experiment_review_2026-09-13/completed_developmental_reviews.json`	All individual report identities and their assessments.	Evidence index
Current continuation storage estimate `docs/experiment_review_2026-09-13/caregiver_continuation_storage_overhead.json`	Measured new companion overhead and original full-study reserve.	Measured minimum, not a long-run ceiling
Earlier remaining-fit estimate `docs/experiment_review_2026-09-13/caregiver_remaining_storage_after_benchmark.json`	Original-format fit sizes before complete continuation companions.	Historical estimate; use new overhead too
Unchanged stillness declaration and launch arguments `docs/experiment_review_2026-09-13/stillness_v3_ready.json`	Eight matched pairs, learning/retention budgets, source and resource prerequisite.	Prepared; full study not yet run
Original end-to-end negative `Saved/mimo_bridge/caregiver_fit_continuation_native_20260913_v2/report.json`	Independent fits differ; this report alone cannot attribute the cause to restart.	Retained failure
Traced end-to-end comparison `Saved/mimo_bridge/caregiver_fit_continuation_native_20260913_v3/report.json`	Per-update provenance locates the initial rendering difference.	Retained pixel-equality failure
First different image diagnosis `Saved/mimo_bridge/caregiver_fit_continuation_native_20260913_v3/raw_input_diagnosis.json`	Identical first learning update and later one-pixel observation difference.	Located observation discrepancy
Physical continuation and rendering floor `Saved/mimo_bridge/caregiver_render_cutpoint_20260913/report.json`	Exact simulator/camera/scene state; a few fresh-render channel differences of 1/255.	Physical pass; pixels not exact
Learning continuation with identical inputs `Saved/mimo_bridge/caregiver_recorded_continuation_20260913/report.json`	Exact learner, Adam, RNG, orders and physical outcomes across save/resume.	Component preservation; no benefit-trial input substitution
Earlier full caregiver-study launch `Saved/mimo_bridge/caregiver_exchange_2026-09-12/resume_launch_20260913_finish.json`	Original-budget restart before the later monitoring read collision. Read the current readiness note for the latest continuation.	Historical launch; learning saved on later stop
Continuation and scheduler tests `Saved/mimo_bridge/caregiver_continuation_review_20260913/tests.xml`	Focused copied continuation, resource and stillness verification.	Software tests
Monitoring read-collision repair tests `Saved/mimo_bridge/caregiver_status_io_review_20260913/tests_final.xml`	47 tests: temporary optional monitoring errors preserve the fit; stale observations block additional workers; required completion records remain strict.	Passed software tests; independently reviewed

Integration, dashboard, storage and final handoff (21 records)

Root maintains this handoff and the link directory. These files distinguish verified deployment, actual behavior, pending experiments and storage permission.

Document or record	What to use it for	Scope
Editable progress handoff `docs/gain_function_2026-09-13/resume_handoff.json`	Current run, completed work, pending work, exact next steps and constraints.	Update before rebuilding the plan
Short completion status `docs/gain_function_2026-09-13/completion_status.md`	Plain-language completion limits and current work.	Dated summary
Machine-readable completion status `docs/gain_function_2026-09-13/completion_status.json`	Test scope, deployment evidence and unresolved work.	Dated summary
H81 live migration verification `docs/gain_function_2026-09-13/decorated_shadow_live_verification.json`	Exact H80 final source and actual finite production updates.	Deployment receipt; no ability pass
Fresh H81 checkpoint verification `docs/gain_function_2026-09-13/H81_published_checkpoint_verification.json`	Published native files, physical companion SHA and post-save learning.	Verified recovery checkpoint
Earlier local-skin live verification `docs/gain_function_2026-09-13/local_skin_live_verification.json`	Original local-sensing integration and training preservation.	Earlier deployment receipt
Representation contract `docs/gain_function_2026-09-13/representation_contract.md`	Preserve information, physical meaning and sources through model components.	Implementation contract
Trace and evidence contract `docs/gain_function_2026-09-13/trace_contract.md`	Required provenance for tests and developmental claims.	Evidence contract
Dashboard work registry `docs/development_dashboard/work.json`	Current experiments, questions, decisions and linked run records.	Live application data
Dashboard evidence registry `docs/development_dashboard/evidence.json`	Receipts and scopes for model and separate-experiment claims.	Live application data
Automatic experiment declarations `docs/development_dashboard/queue.json`	Original and recurring measurement schedules and budgets.	Live queue specification
Live training supervisor `Saved/mimo_bridge/training_continuation.json`	Current process ownership, run and continuation command.	Live; read before any trainer operation
Live measurement queue `Saved/mimo_bridge/developmental_queue.json`	Active tests, completed records, resource pauses and queued work.	Live; do not duplicate an active test
Temporary measurement-slot reservation `Saved/mimo_bridge/queue_reservation_reload_20260913.json`	Preserved active forecast and expiring hold; release is recorded here.	Completed coordination receipt
Final core continuation regressions `Saved/mimo_bridge/final_continuation_regression_20260913.xml`	109 affected tests, prior to the separate metadata-only fix.	Passed software tests
Dashboard regressions `Saved/mimo_bridge/final_dashboard_regression_20260913c.xml`	42 dashboard/evidence/experiment-progress tests.	Passed software tests
Queue launch-hold regressions `Saved/mimo_bridge/queue_hold_regression_20260913.xml`	11 tests including preserved active job and automatic expiry.	Passed software tests
Dashboard and video browser checks `Saved/mimo_bridge/dashboard_finish_20260913/browser_results.json`	Desktop/mobile, legacy pages, history and actual video playback.	Browser verification
Disposable test-fixture cleanup `Saved/mimo_bridge/completed_test_scratch_cleanup_20260913.json`	Only completed test scratch; reports and actual experiment evidence retained.	Storage receipt
Older C proof-snapshot thinning `Saved/mimo_bridge/storage_thinning_receipt_c_decorated_20260913.json`	Retained source/final proof while retiring an intermediate snapshot.	Storage receipt
One older D smoke-snapshot thinning `Saved/mimo_bridge/storage_thinning_receipt_d_one_smoke_initial_20260913.json`	Other initial and all four final states retained; direct replay limitation recorded.	Storage receipt

Operating state, storage and permission boundary

Active run at this update: C:\Users\bhart\Documents\UEMA_devA_private\Saved\mimo_bridge\experiments\H83_no_scheduled_rest_20260914
Supervisor and trainer process IDs at this update: None; [37160]
Source of truth: Saved/mimo_bridge/training_continuation.json and active run replay_training_stats.jsonl
Newest verified complete recovery checkpoint: D:/UEMA_recordings/checkpoint_runs/H81_decorated_shadow_20260913/checkpoints/ckpt_step_4019025
Historical migration source: D:/UEMA_recordings/checkpoint_runs/H82_dephased_babble_20260914/checkpoints/ckpt_step_4083925
Free space when recorded: C: 9.23 GiB; D: 28.83 GiB.
Pending archival permission: Resolved 2026-09-13: the operator explicitly authorized retiring the 32 archived originals. All 32 were deleted only after their original AND archive SHA256 re-verified; each has a .retired.json receipt beside its retained .xz archive under developmental_tests_20260913/development-v3-action-forecast-step-3996400/attempt_1. UPLOAD_MANIFEST.json beside the archives (and in the operator's Google Drive folder UEMA_recordings_archive_2026-09-13) lists every expected hash. The operator will upload the 2.34 GiB of archives to Google Drive later; until then the archives exist only on D:.
Storage protection: H80 final4018725 and auxiliary source4016875 remain pinned, together with current H81 recovery. Completed old H80proof source4017125 was retired only after its6native files matched the retained CPU source exactly and no active/shared dependency remained; adapted full episode with original-value archive and final proof/report remain. C intermediate proof4017127 and one of two D smoke initials retired; source/final proof, other smoke initial and all4smoke finals retained. Current D headroom barely exceeds the measured10.73GiB minimum; longer runs and extra pauses may require more. Retention protects evidence even when its6GiB target cannot be met.

D:/UEMA_recordings/.checkpoint_retention.json
Saved/mimo_bridge/checkpoint_retention_status.json

User authorizes implementation that preserves training and wants ongoing training. Do not ask again for ordinary reversible fixes or the already authorized deployment. Preserve weights, optimizer, physical episode, RNG, held orders, current percepts and memory.
User explicitly deferred the MIMo avatar voice for discussion. Do not implement or activate it now. Caregiver speech is separate.
Do not lower scientific gates after seeing results or count software tests as acquired abilities. Keep original failures and source identities. Current no-roll/no-grasp findings are limited observed behavior, not proof of impossibility.
No bound actor module may change during its continuation without an explicit validated boundary. The frozen production manifest lists 29 modules. New auxiliary tools are deliberately separate.
The 32 archived originals were retired 2026-09-13 under the operator's explicit authorization, with per-file SHA re-verification receipts. The retained .xz archives on D: are now the only copy of those records until the operator uploads them to Google Drive; do not delete or rewrite the archives.

Evidence, commands and commit status

Git: feature/mimo-bridge. Core implementation, tests and supporting documents are committed as c39d3a1ee (126 files). The follow-up commit e1e8b365d adds the optional telemetry-read repair, the caregiver source pin and this handoff's documentation. Both were pushed to the remote branch on 2026-09-13 after the user gave explicit approval in chat; the push output confirmed b60bd6263..e1e8b365d.

Previously pushed commits: 9e79ab5d4, b60bd6263. Use git status to inspect this working tree; do not infer that later edits are pushed.

Set-Location -LiteralPath 'C:\Users\bhart\Documents\UEMA_devA_private'
$env:PYTHONPATH='Source/Python'
$env:PYTHONIOENCODING='utf-8'
$env:OMP_NUM_THREADS='1'
Get-Content Saved/mimo_bridge/training_continuation.json
Get-Content Saved/mimo_bridge/developmental_queue.json
Get-Content docs/experiment_review_2026-09-13/study_runner_and_dashboard_readiness.md
Get-Content docs/experiment_review_2026-09-13/causal_readout_v2_development_plan.md
# Edit resume_handoff.json and the workstream document directory before regenerating:
.venv_mimo/Scripts/python.exe docs/gain_function_2026-09-13/build_workstream_directory.py
.venv_mimo/Scripts/python.exe docs/gain_function_2026-09-13/build_plan.py
# Dashboard: http://127.0.0.1:8766/ ; live JSON: /api/progress

Update docs/gain_function_2026-09-13/resume_handoff.json and rerun build_plan.py whenever the live boundary, test state, next step or permission state changes.

1. What is wrong, and what direction is credible?

The earlier architecture could train predictions for days without learning a reason to roll. A low average prediction error can describe a mostly unchanged scene. A scalar comfort objective can prefer a stationary body. Adding a caregiver does not, by itself, connect controllable discoveries to the self's future. Increasing memory slots does not create those discoveries.

The credible route is a closed loop: experience actual bodily consequences; distinguish what action controls; predict the alternatives on the timescale the body responds; preserve functioning; and pursue demonstrated opportunities to learn useful distinctions. Here, useful means extending or maintaining the self's functioning and access, not pleasing an external task scorer.

Engineering judgment: this architecture can support the proposed early motor functions if its mechanics allow them and the new learning loop produces enough controlled experience. Current evidence does not justify assuming all four stages will emerge from unattended prediction training. We now test physical possibility, information availability, motivation, action selection and retained behavior separately so failure has an actionable location.

The initial MIMo brain is roughly 5.5 million parameters; the added grounded/spatial self model brings this configuration to roughly 5.73 million. This is a body-learning test bed, not the planned 280-million-parameter UE brain. Parameter count is neither an ability measurement nor a substitute for bodily experience. Measure simulated seconds, completed action consequences and retained control per unit of compute.

2. The complete perception–intention loop

Physical contacts first produce intrinsic signals: local pressure, vibration, sliding, force and joint state. The body's modulation operator changes available sensing or capacity. Encoders preserve local identity and spatial relations. The self contains both observed bodily functioning and learned beliefs about consequences; an observed function channel is not itself learned self-recognition.

The predictor receives recent motion and the actual held, smoothed command sequence plus elapsed time. A hypothetical command is not an achieved pose. Future self state is predicted as a change from current grounded state. A frozen acting copy supplies stable perception and choices while the learner updates from actual experience. A promoted replacement keeps its own continuous history and the latest learner's earned weights.

When functioning is threatened

Compare alternatives using local availability, capacity and uncertainty. Prefer a feasible future that restores functioning. A temporary effort cost can be acceptable; a fixed pose or zero movement is not the homeostatic target. Distinguish injury proxies, sensor availability and muscle fatigue instead of averaging them into one number.

When the body is already well

Use completed evidence of improving, action-dependent prediction to identify where another action could acquire reliable distinctions. Compare against persistence, held commands and shuffled commands. Keep a historical best so forgetting and relearning do not earn repeated credit. Mastered repetition satiates. Uncontrolled noise earns no opportunity.

A positive opportunity is a processing-state component of the perceived self. It is not a label saying “roll now.” Existing unlabelled motor exploration provides bootstrap experience when the new audit has no evidence. Its contribution and any scheduled overrides remain visible. The acquisition route must beat a matched comparison in actual learned control while maintaining viable function before replacing the checked controller.

3. Concrete mechanisms and their ramifications

Component	Implementation and reason	What could still fail; required check
Body function	`body_function.py` and `function_transition.py`: apply modulation to the next actual sensory frame; preserve the last experienced frame and existing fatigue. Do not retroactively call shadow impairment experienced pain.	Pressure may be averaged over sensor support instead of tissue footprint. Verify force, actual contacted site and area at the same instant. No arbitrary threshold reduction to manufacture a recovery example.
Current self and forecast	`self_state.py`, `self_state_integration.py`: grounded observed state plus learned local belief; predict action-caused changes with temporal history. Balanced local losses retain affected and healthy sites.	Observed passthrough must not be reported as learned recognition. Report present readout error and future action error separately, per domain and on fresh worlds.
Spatial vision and touch	`spatial_self.py`: local self queries attend spatial encoder regions. Initially zero output preserves old behavior. Existing global sensory targets remain unchanged; local touch sites retain stable identity.	Spatial tokens cannot invent an occluded object. Mask padding, test same-global-summary/different-layout scenes, and require actual camera-region data in the native training path.
Learnability as self state	`learning_opportunity.py`, `opportunity_integration.py`: strictly prospective prediction audits in frozen coordinates, controlled sensory effects, retained anchors and historical best.	Noisy TV, duplicate sensors, forgetting/relearning and repeated mastered actions must fail to create incentive. More fingertip sensitivity is a biological wiring choice; duplicating identical channels cannot multiply credit.
Stable action while learning	`acting_snapshot.py`, `self_intention.py`: freeze the full perceptual/controller copy; continue the learner; stage, test and promote candidate copies with independent history.	Bind receipts to weights, execution code, timing, body/sensory contracts and configuration. Failed or unqualified promotion preserves current control. Changing an encoder invalidates the corresponding behavioral license.
Memory and expanded self	Persistent history carries across physical episodes and checkpoints. Self associations track recurring states and their relation to functioning; the 6,400-slot table retains novelty and interference counters.	A slot is not a discovered object or proven knowledge. A caregiver/tool relation belongs to the functional self only when its controllable effect survives removal, substitution and delay tests.
Longer action and attention	Preserve continuous episodes, recent executed sequences and recurrent/persistent memory. The current transformer context of 128 frames at 50 ms is about 6.4 simulated seconds.	Long episodes do not enlarge attention. Delayed retrieval and multi-step plans require memory-specific tests. Do not lengthen imagination until one-step and sequence consequences are distinguishable at the actual physical scale.
Environment	Use an uninjurious support with room to move for posture; movable objects for manipulation; support edges for standing; responsive adult and grounded speech for social/language learning.	A welded mobile cannot establish grasping a free object. Adult following can hide lack of pursuit. Include absent/replayed adult and removed/relocated objects; no caregiver-directed rolling reward.
Training continuity	Publish a checkpoint only after weights, Adam, runtime and physical companion are complete. Keep canonical configuration beside externally stored checkpoints. Supervisor resumes a validated complete ancestor.	Interrupted disk writes, stale process IDs and a missing USB drive must not silently discard learned evidence. Alert on space, suspend copied experiments first and preserve production save capacity.
Capability evidence	`developmental_observation.py`, `developmental_evidence.py`, copied probe and durable queue: define measures before fitting, restore identical worlds, retain failures, run acquisition/retention/transfer.	Pure body telemetry cannot pass self-recognition. A finished queue job is a measurement, not an acquired skill. Physical reachability requires an actual feasible reference, not a boolean receipt invented by a caller.

The pressure and pleasure interpretation

Force is measured in newtons; pressure is force divided by a physically justified area. A sharp edge can concentrate stress at constant total force. In MIMo, dividing 7.49 N by 97.92 cm² gives about 765 Pa; that measured adapter result does not establish the true tissue footprint of a small sphere. We must not call it a biologically calibrated pain reading. Toddler pain thresholds cannot be derived from impact force alone: stopping distance, contact shape, tissue, location and duration matter.

The resolution variable q is an operator index, not “percent of information remaining.” At q=1 the current operator preserves the original signal. Lower values select a coarse quantization grid. For a 100 kPa range, nine bits give roughly 196 Pa spacing; twelve bits roughly 24 Pa. Whether that loses a relevant distinction depends on the pressure contrast being sensed. Therefore a generic q > 0.8 cannot certify intact perception. Calibrate actual distinguishability at each relevant site and keep muscle-capacity criteria separate.

Small positive modulation should have headroom or improve another real processing capacity; clipping every healthy sensor at its maximum makes that route inert. Healthy development must not require injuring the body first. The present controlled-learning opportunity supplies an additional processing-state route, while positive sensory modulation remains explicitly measurable rather than assumed.

4. Every gain of function and its minimum opportunity

The old “THE 24” page actually named 25 abilities in five tiers. All 25 are retained below in four stages. The stages organize tests; they do not force a universal infant sequence. Active adult instruction is not a prerequisite for rolling or crawling. Infant ages are comparisons, not mandatory simulator waiting periods.

Stage 1 · Body and sensory access

Self / world distinctionbody floor

Natural driver / interpretation: Own motion supplies correlated touch and proprioception; the outside world supplies partly independent changes.
Minimum environment: A movable body, gravity, a firm noninjurious support, room for limbs and sustained sensorimotor experience. Basic caregiving that sustains a real infant is assumed; active adult instruction is not part of the motor test.
Mechanisms required: executed_action_history, proprioception, local_touch
Resolve this gap: A predictable wall is not part of the motor body. Correlation alone cannot distinguish common causes.
Infant comparison: Birth and early infancy; operational contingency sensitivity is not adult self-recognition.
Acquisition test: Select which changes were self-caused when external motion is replayed with matched timing; reverse the action-effect link.
Retention and transfer: Repeat the same direct test after intervening learning and in withheld arrangements; preserve the earlier test definition.
Specification: UEMA §§11.1, 19.2–19.3

Locate the body without lookingbody floor

Natural driver / interpretation: Body motion and self-touch reveal where parts are and how they can move, even without seeing the hand.
Minimum environment: A movable body, gravity, a firm noninjurious support, room for limbs and sustained sensorimotor experience. Basic caregiving that sustains a real infant is assumed; active adult instruction is not part of the motor test.
Mechanisms required: proprioception, executed_action_history, body_frame_coordinates
Resolve this gap: A pooled prediction error can fall while left/right or crossed limbs remain confused.
Infant comparison: Early infancy; a complete mature body schema has no single onset age.
Acquisition test: Predict held-out limb position and self-contact location with vision masked; perturb one joint and retain correct identity.
Retention and transfer: Repeat the same direct test after intervening learning and in withheld arrangements; preserve the earlier test definition.
Specification: UEMA §§3.4, 5.4, 11

Recognize an action’s consequencebody floor

Natural driver / interpretation: Controllable sensory effects support discovering agency and later selecting repeatable effects.
Minimum environment: A movable body, gravity, a firm noninjurious support, room for limbs and sustained sensorimotor experience. Basic caregiving that sustains a real infant is assumed; active adult instruction is not part of the motor test.
Mechanisms required: executed_action_history, causal_prediction, proprioception
Resolve this gap: Candidate command difference must first produce a measurable physical difference. Recognition is not successful action selection.
Infant comparison: Two-month-olds differentiated analog from random sound consequences in a sucking task; this is task-specific.
Acquisition test: Identify the actually executed action among physically distinct alternatives in new worlds; reject delayed, swapped and externally replayed effects.
Retention and transfer: Repeat the same direct test after intervening learning and in withheld arrangements; preserve the earlier test definition.
Specification: UEMA §§6, 19.2, 23

Look at a target and remain on itone visible object

Natural driver / interpretation: Looking can resolve uncertain self-relevant observations and sustain access to changing objects; pain is not required.
Minimum environment: The body and support plus one visible feature or object whose bearing or accessibility can vary. Continuous entertainment, speech and injury are unnecessary for this test.
Mechanisms required: spatial_vision, eye_actuation, head_eye_geometry
Resolve this gap: A motion reflex or teacher can produce a look without the model maintaining an intention.
Infant comparison: Oculomotor control develops in early infancy; pursuit and voluntary target choice are separate measurements.
Acquisition test: Relocate, remove and interrupt the target, then test fixation after arrival; compare integrated gaze and model-authored gaze separately.
Retention and transfer: Repeat the same direct test after intervening learning and in withheld arrangements; preserve the earlier test definition.
Specification: UEMA §§5.1, 6.4, 12

Support and aim the headbody floor

Natural driver / interpretation: Gravity, support contact, effort and access to vision supply consequences of supporting the head.
Minimum environment: A movable body, gravity, a firm noninjurious support, room for limbs and sustained sensorimotor experience. Basic caregiving that sustains a real infant is assumed; active adult instruction is not part of the motor test.
Mechanisms required: neck_actuation, gravity, proprioception, vestibular
Resolve this gap: An actuator responding once does not establish sustainable head control. Check torque, damping and control delay.
Infant comparison: Early infancy; posture and reaching interact rather than following a rigid stage sequence.
Acquisition test: Hold several reachable head bearings against gravity, then recover after a small perturbation; compare limp and external-support controls.
Retention and transfer: Repeat the same direct test after intervening learning and in withheld arrangements; preserve the earlier test definition.
Specification: UEMA §§3.4, 9, 12

Stage 2 · Directed sensing and changing posture

Recognize self-caused visible motionbody floor

Natural driver / interpretation: Visuo-proprioceptive contingency links an observed moving part to an executed action.
Minimum environment: A movable body, gravity, a firm noninjurious support, room for limbs and sustained sensorimotor experience. Basic caregiving that sustains a real infant is assumed; active adult instruction is not part of the motor test.
Mechanisms required: spatial_vision, executed_action_history, causal_prediction
Resolve this gap: Hand visibility is exposure, not recognition. This ability is helpful, not mandatory for tactile reaching or rolling.
Infant comparison: Early infancy contingency tasks; this is not mirror self-recognition or proof of consciousness.
Acquisition test: Predict and distinguish own moving limb from matched externally moved or delayed visual motion; retain performance with a new viewpoint.
Retention and transfer: Repeat the same direct test after intervening learning and in withheld arrangements; preserve the earlier test definition.
Specification: UEMA §§11, 19

Control weight across supportsbody floor

Natural driver / interpretation: Support stability, limb loading and access to a desired sensory state make shifts consequential.
Minimum environment: A movable body, gravity, a firm noninjurious support, room for limbs and sustained sensorimotor experience. Basic caregiving that sustains a real infant is assumed; active adult instruction is not part of the motor test.
Mechanisms required: trunk_actuation, limb_actuation, support_contact, vestibular, body_dynamics
Resolve this gap: Center-of-mass displacement caused by collapse or external forces is not controlled weight shift.
Infant comparison: Develops within each posture; skill does not automatically transfer from sitting to crawling or standing.
Acquisition test: Move support loading toward a chosen reachable side and recover when the support shifts; test both directions with mirrored initial conditions.
Retention and transfer: Repeat the same direct test after intervening learning and in withheld arrangements; preserve the earlier test definition.
Specification: UEMA §§3.4, 9, 12

Act before an expected consequenceone visible object

Natural driver / interpretation: A learned cue can forecast accessible contact, loss of function or another useful bodily consequence before it happens.
Minimum environment: The body and support plus one visible feature or object whose bearing or accessibility can vary. Continuous entertainment, speech and injury are unnecessary for this test.
Mechanisms required: time_memory, causal_prediction, persistent_intention
Resolve this gap: Pre-event activity alone can be a timer or motion reflex. A violated prediction must change future behavior.
Infant comparison: Six-month-olds showed predictive reaching to moving objects in a tested interception task.
Acquisition test: Change a predictive cue while current body condition is matched; measure appropriate preparation before the event and its disappearance when the contingency is broken.
Retention and transfer: Repeat the same direct test after intervening learning and in withheld arrangements; preserve the earlier test definition.
Specification: UEMA §§10, 12

Reach a selected contact locationone reachable object

Natural driver / interpretation: Contact, exploration and access to a sought object or bodily site provide the purpose; seeing the hand is not universally required.
Minimum environment: One physically reachable movable object with observable contact consequences; alternate body sites can supply self-contact experience without an object.
Mechanisms required: arm_actuation, contact_geometry, target_bearing, causal_prediction
Resolve this gap: Following a commanded joint angle proves the plant, not selection of a seen target.
Infant comparison: Reaching begins around four months in Thelen’s longitudinal sample; six-month-olds also reached to remembered sounding objects in darkness.
Acquisition test: Relocate/remove the target and shift it mid-reach; score actual hand contact, correction and stopping across new positions.
Retention and transfer: Repeat the same direct test after intervening learning and in withheld arrangements; preserve the earlier test definition.
Specification: UEMA §§10, 11

Move an object by striking itone reachable object

Natural driver / interpretation: The strike produces a repeatable sight, sound or contact effect that the model has learned to seek.
Minimum environment: One physically reachable movable object with observable contact consequences; alternate body sites can supply self-contact experience without an object.
Mechanisms required: arm_actuation, movable_object, causal_prediction
Resolve this gap: A higher contact count from indiscriminate movement is not a deliberate strike.
Infant comparison: Early infant contingency learning motivates testing effect-specific action; batting has no universal mastery date.
Acquisition test: Move or remove the strikeable object; measure targeted hand contacts and the desired object effect, including trials where contact no longer causes the effect.
Retention and transfer: Repeat the same direct test after intervening learning and in withheld arrangements; preserve the earlier test definition.
Specification: UEMA §§10, 19

Turn toward a heard eventone located sound

Natural driver / interpretation: Sound can resolve where an unseen event with learned relevance occurred.
Minimum environment: One spatially located sound source whose position can vary independently of visible events. Speech is optional for orienting.
Mechanisms required: spatial_audio, head_eye_geometry, neck_actuation, eye_actuation
Resolve this gap: Stereo audio and a moving head do not prove sound-directed choice; exclude simultaneous visible motion.
Infant comparison: Auditory-directed reaches were tested at five and seven months; sound localization and reaching are distinct.
Acquisition test: Play an unseen source at different bearings; reverse its bearing while preserving waveform and volume, then test silent and nonspatial controls.
Retention and transfer: Repeat the same direct test after intervening learning and in withheld arrangements; preserve the earlier test definition.
Specification: UEMA §§3.3, 5.3, 12

Lift the head to inspect somethingone visible object

Natural driver / interpretation: A head lift reveals previously inaccessible information while consuming effort and changing support.
Minimum environment: The body and support plus one visible feature or object whose bearing or accessibility can vary. Continuous entertainment, speech and injury are unnecessary for this test.
Mechanisms required: neck_actuation, spatial_vision, vestibular, causal_prediction
Resolve this gap: A neck trajectory or a repeated lift may have no relation to the target. A rewarding toy is not the only possible reason to lift.
Infant comparison: Early infancy postural and visual control interact; no single milestone age is claimed.
Acquisition test: Relocate/remove a below-axis target, move it during the lift, and test stopping once visible; compare current needs and alternative targets.
Retention and transfer: Repeat the same direct test after intervening learning and in withheld arrangements; preserve the earlier test definition.
Specification: UEMA §§9–10, 12

Roll and use the new posturebody floor

Natural driver / interpretation: Exploratory movement, support redistribution and access to new views can lead to rolling without imposed pain or a directing adult.
Minimum environment: A movable body, gravity, a firm noninjurious support, room for limbs and sustained sensorimotor experience. Basic caregiving that sustains a real infant is assumed; active adult instruction is not part of the motor test.
Mechanisms required: trunk_actuation, limb_actuation, neck_actuation, gravity, friction, body_dynamics
Resolve this gap: An empty comfortable world need not make rolling preferable on every trial. A limp-body rotation is not a learned roll.
Infant comparison: Infant postural development varies; rolling is not one of the six WHO milestone windows.
Acquisition test: Measure whole-body supine-to-prone transitions and stability; change the sought consequence to the opposite side and interrupt midway; compare passive drift.
Retention and transfer: Repeat the same direct test after intervening learning and in withheld arrangements; preserve the earlier test definition.
Specification: UEMA §§3.4, 9, 19

Support the chest while proneprone floor

Natural driver / interpretation: Support and head elevation expand vision and free movement while distributing load.
Minimum environment: Safe prone support and free arms/head, with ordinary visual and tactile variation. A completed roll is not required to supply a prone starting state.
Mechanisms required: arm_actuation, trunk_actuation, neck_actuation, support_contact
Resolve this gap: Do not withhold all prone experience until the learner has rolled; test transition and posture as separate abilities.
Infant comparison: An infant can practice prone support when placed prone; a self-produced roll is not a prerequisite.
Acquisition test: From a prone start, support the upper body with the arms, retain head access and recover after support disturbance; compare externally supported and limp trials.
Retention and transfer: Repeat the same direct test after intervening learning and in withheld arrangements; preserve the earlier test definition.
Specification: UEMA §§3.4, 9

Stage 3 · Sustained manipulation and floor mobility

Acquire an object with the handgraspable object

Natural driver / interpretation: Retaining an object permits further tactile exploration or future actions; grasping need not be driven by deficit relief.
Minimum environment: One unattached object with size, mass and friction within the hand’s measured grasp envelope.
Mechanisms required: finger_actuation, palm_geometry, contact_geometry, local_touch, object_dynamics
Resolve this gap: Finger closure without an acquired object is not grasp. Self-membership alone is not behavioral incorporation.
Infant comparison: Manual contact and grasp develop through the first year; object size and support change the task.
Acquisition test: Acquire then lift an unglued object, carry it under changed orientation and release it; reject welded, frictionless and externally held controls.
Retention and transfer: Repeat the same direct test after intervening learning and in withheld arrangements; preserve the earlier test definition.
Specification: UEMA §§3.1, 11.2

Keep a held object from slippinggraspable object

Natural driver / interpretation: Loss of a currently useful relation predicts loss of available action; local slip and pressure guide correction.
Minimum environment: One unattached object with size, mass and friction within the hand’s measured grasp envelope.
Mechanisms required: finger_actuation, relative_slip, contact_geometry, object_dynamics
Resolve this gap: Blind tightening may keep an object but fail context-sensitive control or crush a compliant object.
Infant comparison: Infant grip adaptation depends on object and experience; a human age threshold is not supplied for this exact test.
Acquisition test: Reduce friction or perturb load after an established grasp; measure correction, retained object and excessive force, against no-slip and useless-object controls.
Retention and transfer: Repeat the same direct test after intervening learning and in withheld arrangements; preserve the earlier test definition.
Specification: UEMA §11.3

Put down and release an objectobject and support

Natural driver / interpretation: Releasing can free a hand, lower effort or enable the next interaction; larger self need not mean permanent possession.
Minimum environment: A graspable object and a support on which placing it changes load or frees the hand.
Mechanisms required: finger_actuation, arm_actuation, contact_geometry, persistent_intention
Resolve this gap: Dropping anywhere is not placement; preventing every self-boundary contraction would make useful release impossible.
Infant comparison: Manual means-end control develops across later infancy and toddlerhood; no universal age is claimed.
Acquisition test: Place on a changed support location, release without external attachment and use the freed hand; move the support mid-action and remove it.
Retention and transfer: Repeat the same direct test after intervening learning and in withheld arrangements; preserve the earlier test definition.
Specification: UEMA §§10–11

Coordinate both hands on one objecttwo hand object

Natural driver / interpretation: Two hands can support, rotate or explore what one hand cannot; simultaneous need is physical, not a named reward.
Minimum environment: An object whose size or loading makes coordinated two-hand interaction physically relevant.
Mechanisms required: bilateral_arm_actuation, finger_actuation, cross_hand_touch, object_dynamics
Resolve this gap: Two hands touching at once is not coordinated control; do not require two separately mastered grasps before any bimanual experience.
Infant comparison: Four infants followed through their first year alternated uni- and bimanual reaching tendencies.
Acquisition test: Carry or rotate an object requiring both hands; perturb one contact and score coordinated compensation, against a one-hand-sufficient object.
Retention and transfer: Repeat the same direct test after intervening learning and in withheld arrangements; preserve the earlier test definition.
Specification: UEMA §§5.2, 10–11

Maintain sitting balancesupported to free sitting

Natural driver / interpretation: A stable upright base frees the hands and changes visible/accessed space; no aversive training is necessary.
Minimum environment: A floor and physically attainable seated posture; external placement or temporary support supplies experience but never counts as independent sitting.
Mechanisms required: trunk_actuation, hip_actuation, vestibular, support_contact, body_dynamics
Resolve this gap: A fixed base, locked joint or caregiver holding the trunk can manufacture a sitting score.
Infant comparison: WHO sitting without support window: 3.8–9.2 months, spanning the first to ninety-ninth percentiles.
Acquisition test: From a reachable seated state, maintain balance and correct pushes; separately test getting into sitting. Remove external supports during scoring.
Retention and transfer: Repeat the same direct test after intervening learning and in withheld arrangements; preserve the earlier test definition.
Specification: UEMA §§3.4, 9

Travel on the floor toward a consequenceopen floor

Natural driver / interpretation: Exploration, access to objects and people, and moving between surfaces can motivate travel without pain.
Minimum environment: Enough floor area and friction for body-length travel with ordinary reachable surfaces or objects. A caregiver can offer destinations but need not continually intervene.
Mechanisms required: trunk_actuation, limb_actuation, friction, support_contact, body_dynamics, persistent_intention
Resolve this gap: Some infants use other locomotor forms. Score purposeful travel separately from a specific hands-and-knees gait.
Infant comparison: WHO hands-and-knees crawling window: 5.2–13.5 months; some children skip it.
Acquisition test: Travel to a selected changing destination, correct a detour, stop/use the destination, and generalize floor friction; compare passive drift and destination absence.
Retention and transfer: Repeat the same direct test after intervening learning and in withheld arrangements; preserve the earlier test definition.
Specification: UEMA §§9–10, 19

Use a kneeling support postureopen floor

Natural driver / interpretation: Kneeling can change access and provide a support arrangement for manipulation or rising.
Minimum environment: Enough floor area and friction for body-length travel with ordinary reachable surfaces or objects. A caregiver can offer destinations but need not continually intervene.
Mechanisms required: hip_actuation, knee_actuation, trunk_actuation, support_contact
Resolve this gap: Sitting is a possible route, not a required previously mastered state; confirm knees can contact the floor physically.
Infant comparison: A transition pattern, not a universal mandatory milestone or a WHO age window.
Acquisition test: Enter or sustain a knee-supported posture when it enables access, correct perturbations and leave it when the consequence changes.
Retention and transfer: Repeat the same direct test after intervening learning and in withheld arrangements; preserve the earlier test definition.
Specification: UEMA §§3.4, 9–10

Stage 4 · Composed action and expanded access

Use a held object to affect anothertool and out of reach object

Natural driver / interpretation: A held object extends reachable consequences, so its relation matters when it enables a valued outcome.
Minimum environment: A graspable effective tool, an initially inaccessible consequential object, and real object-object contact; compare ineffective and substitute tools.
Mechanisms required: grasp_geometry, object_object_contact, entity_relations, persistent_intention, causal_prediction
Resolve this gap: Holding a stick or changing a slot membership score is not tool use; causal gap and substitute controls are essential.
Infant comparison: In one longitudinal rake task, spatially separated retrieval became reliable after about eighteen months; easier contact cases occurred earlier.
Acquisition test: Retrieve an out-of-reach object with an effective tool, reject a similar ineffective tool and transfer to a different-looking effective one.
Retention and transfer: Repeat the same direct test after intervening learning and in withheld arrangements; preserve the earlier test definition.
Specification: UEMA §§11.2–11.4

Rise using a supportstable graspable support

Natural driver / interpretation: A reachable stable support permits a new viewpoint and access with different limb loading.
Minimum environment: A support reachable from the floor, capable of accepting grip force and body load without hidden attachment.
Mechanisms required: support_grasp, hip_actuation, knee_actuation, trunk_actuation, support_contact
Resolve this gap: Standing after being placed upright is different from executing the transition. Support force must pass through actual contacts.
Infant comparison: WHO standing with assistance window: 4.8–11.4 months; this is not identical to independently pulling up.
Acquisition test: Rise from a lower posture by contacting a support and extending the body; change support position and compliance; reject invisible lifting assistance.
Retention and transfer: Repeat the same direct test after intervening learning and in withheld arrangements; preserve the earlier test definition.
Specification: UEMA §§3.4, 9–11

Balance while standingflat floor

Natural driver / interpretation: Upright balance permits other actions and access; stability costs and opportunities change with posture.
Minimum environment: A physically attainable upright stance on a supportive floor; a teacher may place the body for practice, but no aid remains during independent balance scoring.
Mechanisms required: ankle_actuation, hip_actuation, trunk_actuation, vestibular, support_contact, fast_feedback
Resolve this gap: Pull-to-stand is one route to standing, not a logical prerequisite for learning balance from a placed stance. Check fast-loop stability first.
Infant comparison: WHO standing alone window: 6.9–16.9 months.
Acquisition test: Maintain unsupported standing and recover from calibrated perturbations; vary support surface and remove all external aids.
Retention and transfer: Repeat the same direct test after intervening learning and in withheld arrangements; preserve the earlier test definition.
Specification: UEMA §§3.4, 9–10

Move while uprightsupport and open floor

Natural driver / interpretation: Changing access and self-generated exploration drive repeated locomotor practice, including travel without an apparent destination.
Minimum environment: A rail for cruising and adjacent free floor for distinct independent-step tests; space for balance corrections.
Mechanisms required: leg_actuation, ankle_actuation, support_contact, fast_feedback, body_dynamics
Resolve this gap: Cruising and independent stepping are two subtests. Neither establishes general walking or a one-year developmental equivalence.
Infant comparison: WHO walking with assistance window: about 6.0–13.7 months; walking alone: 8.2–17.6 months.
Acquisition test: Score supported cruising separately from unsupported steps, then destination correction, stopping and changed ground; compare external support and passive propulsion.
Retention and transfer: Repeat the same direct test after intervening learning and in withheld arrangements; preserve the earlier test definition.
Specification: UEMA §§9–10

5. Attack the plan before calling it progress

Failure	Why it can fool us	Concrete resolution
Comfortable immobility	A deficit-only score has no reason to explore while healthy.	Prospective controlled-learning opportunity; historical-best anchors; matched learning comparison with viable functioning. Count newly retained control, not motion volume.
Self-confirming prediction	A top-down replacement becomes its own target.	Keep actual sensory targets independent of inferred reconstruction. Test that altered top-down reports cannot change the physical training target.
Predictable damage	Prediction error falls while sensing worsens.	Observed local availability remains part of self. Predictability alone cannot satisfy the viable-function criterion.
Passive recovery	An object leaves before the decision; holding still appears purposeful.	Show impairment source remains at decision time and a feasible intervention changes the actual sensed distinction; compare hold and source removal. Keep the failed passive fixtures.
Motor movement mistaken for rolling	A trunk joint responds while whole-body dynamics cannot turn the body.	Physics-only reference search under actual ROM, force, smoothing and timing; score front-down relative to gravity with support, not yaw. Search failure is inconclusive unless the physical constraint is proved.
World luck mistaken for progress	A toy happens to start in the fovea, or the adult delivers it.	Paired identical initial states; independently arranged worlds; relocation/removal/interruption; acquisition plus retained performance after intervening learning.
Reward from duplicate pixels or pain	Summing more sensors or using error rewards can create an exploitable incentive.	Group redundant evidence; require action-dependent improvement; filter nonviable futures and impaired learning episodes; audit regional sensitivity separately from sensor duplication.
Contaminated action trial	Smoother, momentum, recurrent state or RNG carries over from another candidate.	Restore full physics and learner state for each branch. Replay recorded commands exactly. Compare true writer/reader units and achieved pose conventions.
Moving goalposts	A poor predictor passes a smaller gate invented after measurement.	Freeze meaningful effect and error bounds before fitting; preserve all negative worlds. Route failure to mechanics, representation, evidence quantity or motivation without relabeling it.
All software passes, no function	Unit tests prove wiring; they do not prove intention.	Three separate receipts: integration, independent physical behavior, and live retention. Dashboard uses these scopes explicitly.

What the September 8 review still contributes

The entire external review brief was reread, including its caveats and prior mistakes. Its demands for real physics, action provenance, direct behavior and retention of failed results remain binding engineering requirements. Its old numerical readings are historical. The claim that every later function must wait for visual self-credit is not a valid present prerequisite: proprioception, touch and mechanics can support posture learning. The claim that nothing else needed building for posture also requires a current whole-body reachability test and a motive to acquire control.

6. Execution order, tests and deployment

Current deployment: grounded spatial self learning and local skin modulation are live. A separate frozen chooser observes the same world while the existing learner continues to control the body. The migration preserved every earlier tensor, optimizer history and physical episode. This is a working integration, not a passed developmental test. Read the current verification and remaining work; machine-readable receipt.

Learning-audit limitation found in the live path: the existing observer grades updated predictor weights using its older frozen sensory coordinates. That measures compatibility with the old encoder, not the continuing learner's complete ability. Its 50-millisecond outcome also samples only the beginning of each 250-millisecond held command. A zero reading cannot establish absence of learning, and a positive reading cannot qualify the healthy-exploration controller. Preserve these records as diagnostics. A replacement must train and evaluate compatible representations, retain fixed targets, and test outcomes over actual executed command sequences. The observer remains unqualified while this is resolved.

Preserve continuity and complete saves. Validate weights, optimizer moments, current perception, physical world, action smoother, RNG, history and curriculum. Put large checkpoint files on monitored storage. Recover from the last complete save after a hardware/storage fault; record the exact unsaved interval.
Activate real function experience and continued grounded learning. Migrate every old parameter and optimizer state unchanged. Import only the added grounded module's earned evidence from a copied study. Newly added spatial weights start neutral. Do not replace the latest old brain with an earlier copied-study checkpoint.
Measure physical opportunity and actual behavior immediately. Run target relocation, removal, interruption and settle tests; run body-only floor movement and rolling reachability; record action authorship, contacts, orientation, simulated exposure and full replay receipts.
Qualify action-sensitive self forecasts and healthy acquisition. Separate skin and motor domains; the useful motor forecast is not evidence of a working touch forecast. Check prediction error against the physical separation needed to choose. Acquisition may qualify on newly controlled distinctions or physical task success, with retained functioning and retention.
Compare interventions by learned function. Keep the original caregiver-versus-replay study; add the scheduled-stillness comparison with equal physical experience and equal exploration schedule. No result based only on lower prediction error or greater movement can decide these studies.
Run subsequent stages when physical prerequisites exist. Free-object manipulation, floor mobility, support-assisted transitions and tool retrieval use the same evidence contract. An unavailable measurement is reported as unassessable, never zero. Network self-recognition uses representation tests, not a renamed body coordinate.
Promote only a qualified acting candidate. Stage it on the actual stream; preserve candidate history through promotion; continue learner updates. A failed test leaves the old controller operating and produces a named repair task. No endless waiting for another permission question is built into routine testing.

Every queue job records its checkpoint, declaration, exact command, output directory, process identity and result hash. Stopped measurements retry in a new directory with bounded retries. Experiment timing is estimated from completed work when available; wall-clock promises are not fabricated. Experiment failure must not stop unrelated production learning.

Acceptance is concrete: all old trained state survives migration; actual next-frame modulation is observed; spatial information reaches self forecasts; actor and learner continue independently; noise/forgetting controls do not create false opportunity; the recorder measures real rolling rather than yaw; missing/empty/NaN evidence cannot pass; completed behavior results appear in the dashboard and history; the old dashboard still updates.

Actor/learner verification · Declared automatic experiments · Behavior evidence contract · Acting and learning implementation. Current deployment and measurement results belong in the linked dashboard, not in permanently optimistic prose.

7. Transfer to UE Avatar and remaining scientific limits

Transfer interfaces should carry units, stable intrinsic site identity, local position/normal, actual elapsed time, available resolution, regional response and executed actuator history. UE's denser fingertips should preserve their additional distinctions and regional sensitivity. Do not collapse them to MIMo's pooled summary or interpret sensor count as proportional pleasure. Recalibrate contact footprint, force limits and action cadence on UE; MIMo thresholds and handle indices are not automatically transferable.

Keep old motor/perception weights and add neutral adapters where layouts differ. Test identical physical meaning across both bodies, then measure retention and transfer of learned relations. More parameters and richer sensing may increase capacity, but success depends on adequate information, controllability and experience. The avatar's own vocal tract remains deferred by the user's instruction; caregiver speech is already a separate environmental channel.

Expanded self is a functional hypothesis: an object, person or place becomes represented in relation to continued capability and regulation. Causal dependence is not ownership; ownership is not consciousness. Tool use must survive changed tools and goals. Loss should provoke context-appropriate regulation, not a hard-coded replacement purchase. No complete empirical equation of consciousness exists here; we can implement and falsify the observable developmental mechanisms without claiming that software integration has created subjective experience.

8. Specification and research trail

The source inventory enumerates all eighteen Omnibus Word documents, their hashes, reading scope and relevance, plus the older external brief. It explicitly distinguishes documents read in full from those enumerated for related work. The full specification mapping and primary research links each infant comparison and its limits.

US11537850B2: simulated senses, perceptions, enhancement/impairment and processing changes. The disclosure supports embodiment mechanisms; it is not a validated biological pain scale.
Perception Space: raw, modulated, encoded and functional perception. Keep these levels separate in code and measurement.
UEMA core dual-register wellness/intention specification: preserve a perceived self and anticipatory regulation. Do not identify wellness with a weighted scalar comfort report.
Developmental maturation, tactile calibration and fast-gain specifications: intrinsic sensory structure, physical calibration and optional demonstrations. Demonstration can supply experience; it does not count as independently acquired behavior.
Adolph et al. (2012): natural locomotor practice motivates measuring physical experience and varied action, not simulated infant-age delays.
WHO Motor Development Study (2006): overlapping achievement windows; no universal required sequence.
Kuhl, Tsao and Liu (2003): responsive interaction supported a tested speech contrast; recordings are not an assumed substitute for social contingency, and this is not a prerequisite claim for rolling.