Edit the state.
Watch the output.
Manipulate a candidate J-space inside the saved neural models. Compare the answer before and after the edit, then test whether facts and later computations survive.
64 hidden coordinates each.
8 message values per direction.
Change the working state, not the answer.
Choose a donor value, then move the edit strength. Only a three-dimensional candidate component is replaced. The network computes what happens next.
One change. A real continuation.
Apply a latent edit, check the controls, then let both branches keep computing.
The models and candidate frames are unchanged. New tests run locally on this case.
01Native starting state
Episode and capture options
No prefix runs by default. The seed fixes aliases, presentation order and private labels.
02Edit the candidate
3 of 64 directions / agentMove individual J coordinates
Offsets are measured in calibration standard deviations. The three axes are local to each agent, not proven shared semantic axes. These controls move activations before inference.
03Ask or continue afterward
Both branches execute the same continuation and two silent recurrent updates. The query does not enter state formation.
Never substituted for a prediction.
One state. Four possible uses.
All actual readouts| Late consumer | Original 1 / 2 | Edited 1 / 2 | Donor reference |
|---|
Table uses XOR mask [1, 1] and verification proposal [1, 0]. Different consumers may have the same answer even when the working pair changes.
How much edit is enough?
Test α = 0…1 in steps of 0.05 on this case. Manual offsets are set to zero for the scan. Every point includes both outputs and a factual-memory audit.
Every measured point
| α | Agent 1 / 2 | Facts 1 / 2 | Private labels |
|---|
This is case-specific selection, not an independently validated threshold. “Memory-preserving” here means both agents recall all eight facts and both private labels correctly on this audit. Projected edits remain artificial.
Compare the competing explanations
The same starting states, donor and continuation. Full-strength edits, zero manual offsets, and the selected writing agent(s). No automatic method selection.
| Intervention | Agent 1 / 2 | Facts 1 / 2 | New recall errors |
|---|
Counts alone can hide different lost facts. “New recall errors” counts originally correct facts that become wrong. A whole paired native donor is a broader intervention, not a localized J edit.
Did the facts survive?
Recall all eight facts from separate copies of the original and edited states. This does not change the active experiment.
What was actually changed?
Coordinates are projected hidden activations, not a drawing of an assumed manifold. J1–J3 are defined by the fixed calibration frame. Manual offsets add a separate displacement inside that frame.
Experiment boundary and measured validation
The default J frame is formed from centered, averaged gradients of four exact-report logits through the two silent updates. It is output-informed. Three dimensions are the rank limit of four centered rows, not an emergence result.
Projected edits are artificial. Earlier workspace studies failed stronger natural-state audits. No new support certificate is claimed here. A whole paired donor is a native-state control, not a localized write. Same-rank controls can also affect output.
These are small saved GRU models, not the hidden states of this chat assistant. No language-model J-space, general intelligence or consciousness claim follows.
Source weights and candidates are frozen. All edits are inference-time interventions. The answer network is never changed, bypassed or replaced by the exact evaluator.
Keep using the edited state
Copy the current before/after states into a live comparison. Each action updates both branches once. The edit is not reapplied on later steps. Preview controls above cannot silently reset this trajectory.
Committed trajectory
| Step / operation | Original branch Agent 1 / 2 | Edited branch Agent 1 / 2 | Private correct Original / edited |
|---|
Working actions finish with the model’s two silent updates. Reloading facts intentionally replaces temporary content with factual content. Additional actions can cause genuine errors. Imports recompute outputs from recorded activation snapshots; they do not authenticate the snapshots’ native origin.
Does the change keep working?
Compare the original state, your edited state, and an actual donor execution. Run exactly the same later operations. Then ask for the facts again.
Challenge the current edit
This test copies the preview. It never changes your committed live session.
What is controlled?
All branches use the same weights, cues, operations and late consumers. The edit is applied once. Work runs continuously without extra monitoring ticks. At each checkpoint, the usual silent delay, output reads and factual audits run on copies. After the final working step, reloads update the real test state in sequence. A factual reload intentionally resets the desired content to the remembered world.
Reference errors are displayed, not excluded from the record. A valid donor history is not evidence that the projected hybrid is native. A donor prepared from a captured live snapshot inherits that snapshot’s provenance limitation. Natural support, generalization and consciousness are not certified.
Output and retained facts
| Checkpoint | Original Agents 1 / 2 | Edited Agents 1 / 2 | Donor reference Agents 1 / 2 | Edited facts correct / 8 | Interpretation Agents 1 / 2 |
|---|---|---|---|---|---|
| Run a comparison to populate actual model predictions. | |||||
“Reference error” means the unedited original or donor already fails the intended answer. “Reverted” means both references work, but the edit returns the original answer. Fact counts measure recall, not proven erasure. The chart audits four independent recall branches at each checkpoint; the subsequent reload rows are an actual no-reset trajectory.
Inspect a measured checkpoint
Try fixed settings on new histories
Use the same method, strength, agent selection, manual offsets and preview suffix on every generated case. Worlds, private labels, cues, aliases, prefix programs and donor changes vary. No case-by-case search or threshold adjustment occurs.
Not run. Uses the continuation and reload schedule above. This is a development check, not a sealed primary experiment.
| Lineage / case | Donor transfer at final work both agents | Reference check both agents | Edited facts correct / 8 | Reload pass all eight facts | Inspect |
|---|
Every attempted case is retained, including reference failures. Scores at intermediate snapshots reuse the same cases. A successful point chosen on one example is not an independently validated intervention. Cancellation retains only completed cases and explicitly marks the run incomplete.
Does the edit reach the other agent?
Change one hidden state. Watch the actual outgoing message, the receiving state, and the later answer. Then restore or replay only the first packet to isolate that connection.
01 Choose the write and continuation
Up to 8 working operations. The preview suffix is not used here. The edit happens once before this program.
Extra receiver updates without working inputs are an experimental schedule, not the normal native program.
The original editor and any committed session stay unchanged.
Read after formation
The plot uses the network’s softmax probability for the intended donor pair, not a guarantee or a calibrated confidence estimate.
Challenge both directions
Same settings, fresh generated histories, both writers. No per-case dose search. Includes every completed case and reference failure.
02 Follow the recipient
Not run| Branch | Writer output | Recipient output | Donor-pair probability |
|---|---|---|---|
| No measurements yet. | |||
03 Inspect the actual packet
Not a learned semantic legendThe eight outgoing values below come from the frozen sender network. They are not eight named facts. Restoration sends the original packet; replay sends the edited packet to an otherwise unchanged recipient.
Did other information survive?
What this comparison can and cannot establish
The recipient sees the partner only through the message interface. At the first exchange its state is identical in the original, edited, and replay branches. Equal incoming packets must therefore produce equal recipient updates. This equality checks the causal wiring and implementation. Whether the packet induces the intended answer and whether it stays effective are separate empirical questions.
Only the first packet is intervened upon. Later feedback is unrestricted, so restoration may cease to block the effect and a replayed effect may disappear. Reference errors remain visible. A whole donor replaces both states and is a competence reference, not proof of a native local edit.
Projected states remain artificial. No new training, natural-support certification, general reciprocity, or chat-assistant hidden-state access is claimed.
01Remember
These are the facts to preserve. Temporary flips and swaps must not overwrite them.
Independent private labels are generated from the seed.
02Compute
Native programUp to 96 operations. Spaces and commas are ignored.
Loading facts sends them as input events. No answer is supplied to the model.
03Retrieve
Actual state changesOnly the retrieval wire is changed. Observation and working communication stay intact. Repeating a cue is not guaranteed to help.
Ask after state formation
Late consumerChanging this query does not alter hidden states. The exact evaluator below is separate from the networks.
Parent
Calibrated access
What remains retrievable?
The audit uses four independent recall branches from the current state. It does not replace the live state. The sequential cue cycle above does.
Each manual working action uses one native operation and two silent steps. Arbitrary click sequences may exceed training schedules.
Session trace
Recorded live decisions, not published benchmark averages. Pair values are always [left, right].
| Native action | Reference | Parent 1 / 2 | Repair 1 / 2 |
|---|
Inspect actual hidden activations
Each grid shows all 64 raw coordinates. These are not selected J-space axes. Colors encode sign and magnitude, not semantic meaning.
Parent · agent 1
Parent · agent 2
Repair · agent 1
Repair · agent 2
Readouts during program stepping use a copy with the standard two silent updates. The running program is not silently modified by inspection.
All eight facts recalled correctly
1,024 historiesConfirmation bank · four independent branches · both agents averaged.
Sequential retrieval, without resets
Per-cue exact pair accuracy. Four cues cover all facts, then the pass repeats. This is a different metric from all-eight-fact exactness.
Source: Komorebi_Workspace_Incubator_v0_6_RESULTS.json, sections new_history_confirmation and sequential_no_reset. Data are embedded unchanged. Seed histories, capacity, training exposure, and unresolved tests matter when interpreting these results.
What is running
This page contains the original float32 weights from two v0.5 parents and their v0.6 main and calibrated access repairs. It performs recurrent inference locally in JavaScript. There is no remote API, training loop, or hidden rule-based substitute for neural predictions.
The two parent networks inside a dyad have separate weights and only communicate through eight-value messages. The repaired versions share the byte-identical frozen base weights with their parents; only the access-phase modules differ. Lineages A and B are different training histories, not the two agents inside a single dyad.
What the exact reference does
An independent two-bit interpreter applies I, L, R, and S to calculate the intended working answer. It also keeps the original eight factual bits for scoring. Neither the reference working value nor stored world array is passed to neural readout or retrieval.
Scope of the live episode
Eight set events initialize the world in seed-dependent order with seed-dependent aliases. The seed also fixes the private labels. The interactive generator is not the published evaluation bank and makes no population-level performance claim. A custom or long sequence can expose genuine model errors.
Validation & boundaries
The port was compared against the original PyTorch routines on 72 model/episode combinations, including long programs, successive cues, and message cuts.
JavaScript dot products use double-precision accumulation and float32 tensor outputs. Minor numerical differences remain; this is not bit-exact inference on every browser or all possible inputs.
Projected-state naturalness, no-wire competence, unilateral workspace writing, and unknown-partner transfer remain unresolved in the source work. This lab does not certify J-space, consciousness, or subjective states.
Controls that change state
Running or stepping a program, retrieving a cue, cycling cues, and manual working operations update the live state. The four-branch memory audit does not. Reading a different late query never updates state. Cutting the wire affects only future retrieval exchanges.
Source provenance
Archive: Komorebi_Workspace_Incubator_v0_6.zip
SHA-256:
Embedded weight pack SHA-256: