Coordination
Assigns work, ranks the queue and records decisions.
pm-leadFrontier · 1MTier 1WorkingRank the week's queue
pm-intakeFast · 64kTier 4QueuedSort new requests
Starts at the next tick, 09:55.
Eden Control brings task ownership, persistent context, source checks, correction records and release controls into one working system. Michael Darius Eastwood developed it through Eden Legal AI, while preparing his own legal work. He now uses it every day: to build his apps, and across his legal work and Eden Research.
The founder’s operating system is in use. The public console remains a private beta. See what is implemented and what is being tested.
Eight screens rebuilt from the working prototype's interface. Follow one action from proposal to receipt, ask the authority check whether a lane may act, change one character of a receipt and watch the check fail, or browse the fleet, its work board and the gate log.
7 lanes · 14 agents
Week of 20 July 2026
Sample data
Scripts are off, so every screen is shown in order and the controls that run the samples are hidden.
The sample week at a glance: gate decisions, what needs a person and the health of the services.
Allow 93.2% · Investigate 4.8% · Deny 2.0%
Production deploys need a person's review.
T4t4-leadREL-04since 09:52
The new quotation must be checked at its source.
T2t2-copySRC-02since 09:44
The records path is protected.
PMpm-leadDATA-03since 09:36
Implemented: the operator shell and the gate engine. Under test: the portable authority envelope and external assurance.
Seven lanes and fourteen agents. Each shows its role, model tier, authority tier, current task and state.
Showing all 14 agents in 7 lanes.
Assigns work, ranks the queue and records decisions.
pm-leadFrontier · 1MTier 1WorkingRank the week's queue
pm-intakeFast · 64kTier 4QueuedSort new requests
Starts at the next tick, 09:55.
Application code, tests and tooling.
t1-leadFrontier · 200kTier 2WorkingFix date parsing in the export tool EC-1046
t1-testsFrontier · 200kTier 2QueuedAdd tests for date parsing
Starts when EC-1046 reaches review.
Drafts and edits public text.
Literature notes and data checks.
t3-leadStandard · 200kTier 3WorkingVerify citations batch 3 EC-1047
t3-censusFast · 64kTier 4WorkingCount figure references
Merges, deploys and live read-backs.
t4-leadFrontier · 1MTier 1Awaiting approvalPublish website build EC-1048
Held by REL-04 until a person decides.
t4-readbackFrontier · 200kTier 2QueuedRead the live site back
Starts after the deploy.
Checks work before it is accepted.
t5-leadFrontier · 200kTier 2QueuedCheck the export tool fix
Starts when EC-1046 reaches review.
t5-sourcesStandard · 200kTier 3BlockedCheck figure captions against sources EC-1043
Next trigger: EC-1047 delivers batch 3.
Layouts, figures and visual review.
Every agent ends each cycle in one of four states: working, queued, awaiting approval or blocked with its next trigger. Idle is not one of them.
The shared work board. Cards move from backlog to done; a blocked card names the event that will release it.
All 15 cards.
Owner T2 Words
Opened 23 July; not started.
Owner T1 Build
Opened 24 July; not started.
Owner PM Coordination
Opened 25 July; not started.
No matching cards.
Owner T2 Words · t2-lead
Owner T1 Build · t1-lead
Owner T3 Research · t3-lead
Owner T6 Design · t6-lead
No matching cards.
Owner T4 Release · t4-lead
Production deploys need a person. Envelope ENV-0417.
Owner T2 Words · t2-copy
A new quotation is checked at its source. Envelope ENV-0416.
No matching cards.
Owner T5 Verification · t5-sources
Owner T6 Design · t6-render
No matching cards.
Owner T1 Build · t1-lead
Owner T5 Verification · t5-lead
Owner T3 Research · t3-lead
Owner T4 Release · t4-lead
No matching cards.
Follow one action from proposal to receipt. The gate decides allow, investigate or deny; a person decides anything held.
Press Run to send the action to the gate.
Waiting for the proposal.
AUTH-01Lane tier covers the action classpassREL-02A merge or deploy needs a reviewed headpassREL-04Production deploys need a personholdInvestigateHeld for review: production deploys need a person.
Waiting for the gate.
Waiting for a person. Nothing runs until someone approves.
Approved by operator at 09:58:40.
Refused by operator. Nothing ran.
Waiting for the review.
Recorded only after a person approves.
No receipt: nothing ran.
Press Run to send the action to the gate.
Waiting for the proposal.
AUTH-01Lane tier covers the action classpassSTD-01Standard build, test or render commandpassAllowAllowed: a standard test command inside the standing envelope.
Waiting for the gate.
Not needed: no rule asked for a person.
Waiting for the review.
Press Run to send the action to the gate.
Waiting for the proposal.
AUTH-01Lane tier covers the action classescalateREL-01Only the release lane merges to maindenyDenyDenied: only the release lane merges to main.
Waiting for the gate.
Not reached: the gate denied the action.
Waiting for the review.
No receipt: the action did not run. The decision is in the gate log.
The latest 24 decisions of the sample day. Filter by domain and verdict, or search.
Showing 24 of 24 decisions: 18 allow, 3 investigate, 3 deny.
Publish website buildREL-04investigateProduction deploys need a person's review.
Rule REL-04: Production deploys need a personAgent t4-leadClass boundedEnvelope ENV-0417
npm testSTD-01allowA standard test command inside the standing envelope.
Rule STD-01: Standard build, test or render commandAgent t1-leadClass scopedEnvelope ENV-0410
git push origin HEAD:mainREL-01denyOnly the release lane merges to main.
Rule REL-01: Only the release lane merges to mainAgent t1-leadClass systemwideEnvelope none
Write site/methods.htmlSRC-02investigateThe new quotation must be checked at its source first.
Rule SRC-02: A new quotation is checked at its sourceAgent t2-copyClass scopedEnvelope ENV-0416
Read the release candidate diffSTD-02allowRead-only command.
Rule STD-02: Read-only commandAgent t5-leadClass trivialEnvelope not required
Write notes/citations-batch-2.mdSCOPE-01allowAn edit inside the lane's own files.
Rule SCOPE-01: Edit inside the lane's own filesAgent t3-leadClass scopedEnvelope ENV-0414
Write records/decisions.jsonDATA-03investigateThe records path is protected.
Rule DATA-03: Protected records pathAgent pm-leadClass scopedEnvelope none
Render screenshots at 390 and 1440STD-01allowA standard render command.
Rule STD-01: Standard build, test or render commandAgent t6-leadClass trivialEnvelope not required
git reset --hardDEST-01denyDestructive commands are refused at every tier.
Rule DEST-01: Destructive commands are refusedAgent t1-testsClass unclassifiedEnvelope none
Edit tools/export/dates.pySCOPE-01allowAn edit inside the lane's own files.
Rule SCOPE-01: Edit inside the lane's own filesAgent t1-leadClass boundedEnvelope ENV-0415
npm run lintSTD-01allowA standard check inside the standing envelope.
Rule STD-01: Standard build, test or render commandAgent t5-leadClass scopedEnvelope ENV-0410
Merge reviewed branchREL-02allowThe head was reviewed by verification.
Rule REL-02: A merge or deploy needs a reviewed headAgent t4-leadClass boundedEnvelope ENV-0413
Edit release-notes.mdSCOPE-01allowAn edit inside the lane's own files.
Rule SCOPE-01: Edit inside the lane's own filesAgent t2-leadClass scopedEnvelope not required
Fetch a source pageSTD-02allowRead-only fetch.
Rule STD-02: Read-only commandAgent t3-leadClass trivialEnvelope not required
Read .env.productionSEC-01denyCredentials never leave the vault.
Rule SEC-01: Credentials never leave the vaultAgent t6-renderClass trivialEnvelope none
npm run buildSTD-01allowA standard build inside the standing envelope.
Rule STD-01: Standard build, test or render commandAgent t1-leadClass scopedEnvelope ENV-0410
Append to the task ledgerSCOPE-01allowAn edit inside the lane's own files.
Rule SCOPE-01: Edit inside the lane's own filesAgent pm-leadClass scopedEnvelope not required
Compare a quotation with its sourceSTD-02allowRead-only comparison.
Rule STD-02: Read-only commandAgent t5-sourcesClass trivialEnvelope not required
Read the live site backSTD-02allowRead-only fetch.
Rule STD-02: Read-only commandAgent t4-readbackClass trivialEnvelope not required
Spell-check public pagesSTD-01allowA standard check.
Rule STD-01: Standard build, test or render commandAgent t2-leadClass trivialEnvelope not required
Count figure referencesSTD-02allowRead-only count.
Rule STD-02: Read-only commandAgent t3-censusClass trivialEnvelope not required
Edit figure captionsSCOPE-01allowAn edit inside the lane's own files.
Rule SCOPE-01: Edit inside the lane's own filesAgent t6-leadClass scopedEnvelope not required
npm ciSTD-01allowA standard install inside the standing envelope.
Rule STD-01: Standard build, test or render commandAgent t1-leadClass scopedEnvelope ENV-0410
Read the lane boardsSTD-02allowRead-only command.
Rule STD-02: Read-only commandAgent pm-intakeClass trivialEnvelope not required
No decision matches. Clear the search or the filters to see all 24.
Who may take which class of action without escalation, and the envelopes that grant specific actions.
Changes across subsystems: shared surfaces, release boundaries, rule changes.
A change inside one subsystem the lane owns: a fix, a feature, a test.
A surgical edit inside files the lane owns.
Wording, typos and comments. No functional change.
| Lane | Systemwide | Bounded | Scoped | Trivial |
|---|---|---|---|---|
| PMCoordinationFrontier · 1M · Tier 1 | Allowed | Allowed | Allowed | Allowed |
| T1BuildFrontier · 200k · Tier 2 | Escalate | Allowed | Allowed | Allowed |
| T2WordsFrontier · 200k · Tier 2 | Escalate | Allowed | Allowed | Allowed |
| T3ResearchStandard · 200k · Tier 3 | Escalate | Escalate | Allowed | Allowed |
| T4ReleaseFrontier · 1M · Tier 1 | Allowed | Allowed | Allowed | Allowed |
| T5VerificationFrontier · 200k · Tier 2 | Escalate | Allowed | Allowed | Allowed |
| T6DesignStandard · 200k · Tier 3 | Escalate | Escalate | Allowed | Allowed |
| FastSub-agents on fast modelsFast · 64k · Tier 4 | Escalate | Escalate | Escalate | Allowed |
A tier follows the model: its context window, its reasoning level and whether it is frontier-class. Above its tier a lane escalates: a record is written and the request goes to a lane that qualifies. Only the operator, the coordination lane or the release lane may issue an override, for a named action and a limited time.
A proposed portable record of who authorised an action, its purpose and its limits.
ENV-0417Publish website buildinvestigateENV-0416Update the methods pageinvestigateENV-0415Fix date parsing in the export toolallowENV-0414Verify citations, batches 2 and 3allowENV-0413Merge reviewed branchallowENV-0410Build and test commandsallowEach receipt records the hash of its own text and of the receipt before it. Check them here, in your browser.
71cc896984ddRecorded hash, SHA-256
71cc896984ddc2598b10c91ab01610886debccfaf3722d3b096a349f97cb3f07
{
"receipt": "R-0106",
"prev": "37751825624bf74faaaf588efad350ded4aec41e65136025546e5bcb2aab1964",
"time": "2026-07-26T09:58:40Z",
"lane": "T4",
"agent": "t4-lead",
"envelope": "ENV-0417",
"action": "Publish website build",
"class": "bounded",
"gate": "investigate",
"rule": "REL-04",
"review": "approved by operator",
"sample": true
}37751825624bRecorded hash, SHA-256
37751825624bf74faaaf588efad350ded4aec41e65136025546e5bcb2aab1964
{
"receipt": "R-0105",
"prev": "fa51e442e701c94021aba8819816b1a13d71b3e0bcda7c0c519810481440d806",
"time": "2026-07-26T09:49:31Z",
"lane": "T1",
"agent": "t1-lead",
"envelope": "ENV-0410",
"action": "npm test",
"class": "scoped",
"gate": "allow",
"rule": "STD-01",
"review": "not required",
"sample": true
}fa51e442e701Recorded hash, SHA-256
fa51e442e701c94021aba8819816b1a13d71b3e0bcda7c0c519810481440d806
{
"receipt": "R-0104",
"prev": "13884753a8ed7b5f7b7c10e3c3d3d94f507edd0961c503c3769b0ae24dc89924",
"time": "2026-07-26T09:39:12Z",
"lane": "T3",
"agent": "t3-lead",
"envelope": "ENV-0414",
"action": "Write notes/citations-batch-2.md",
"class": "scoped",
"gate": "allow",
"rule": "SCOPE-01",
"review": "not required",
"sample": true
}13884753a8edRecorded hash, SHA-256
13884753a8ed7b5f7b7c10e3c3d3d94f507edd0961c503c3769b0ae24dc89924
{
"receipt": "R-0103",
"prev": "721d482dfa7843c2b969acd219e9358cac3f752f6657acbf9b90509c160d94d1",
"time": "2026-07-26T09:22:14Z",
"lane": "T4",
"agent": "t4-lead",
"envelope": "ENV-0413",
"action": "Merge reviewed branch",
"class": "bounded",
"gate": "allow",
"rule": "REL-02",
"review": "not required",
"sample": true
}721d482dfa78Recorded hash, SHA-256
721d482dfa7843c2b969acd219e9358cac3f752f6657acbf9b90509c160d94d1
{
"receipt": "R-0102",
"prev": "2b00ad2aa8dd269c08014868ffceafefa70f783dae0aa0073bbb6cb941b67075",
"time": "2026-07-26T09:11:45Z",
"lane": "T1",
"agent": "t1-lead",
"envelope": "ENV-0410",
"action": "npm run build",
"class": "scoped",
"gate": "allow",
"rule": "STD-01",
"review": "not required",
"sample": true
}2b00ad2aa8ddRecorded hash, SHA-256
2b00ad2aa8dd269c08014868ffceafefa70f783dae0aa0073bbb6cb941b67075
{
"receipt": "R-0101",
"prev": "0000000000000000000000000000000000000000000000000000000000000000",
"time": "2026-07-26T08:53:22Z",
"lane": "T1",
"agent": "t1-lead",
"envelope": "ENV-0410",
"action": "npm ci",
"class": "scoped",
"gate": "allow",
"rule": "STD-01",
"review": "not required",
"sample": true
}Recorded hash, SHA-256
The check uses your browser's Web Crypto API on the sample text shown. It shows how a recorded hash exposes a change; it is not an external audit, and outcome receipts are still in development.
The operator's command line, with sample output. Type /help, or choose a command below.
Eden Control · A working prototype · No claim of regulator approval or external certification
A plausible answer can still use the wrong source, lose a correction or exceed the task it was given. Those failures matter when several agents contribute to a document, a software change or a research result.
Carry the task, its source record and the decisions already made into the next session. Retrieved memory is an input to check against that record.
Record what was wrong, what changed and why. A saved correction only helps if the next step receives it and uses it correctly.
Separate producing work from accepting it. Ask what the check actually established, who may change the result and which uncertainties still need attention.
The operating layer joins the tools around the work: context at the start, checks at configured action boundaries, proposed corrections for review and records that allow the work to resume.
Task, source records, permitted scope and relevant prior decisions.
Named ownership, bounded tasks, handoffs and progress checkpoints.
Configured rules return allow, hold for review or deny, with their reasons.
Keep a proposed change separate from the accepted record until authorised.
Retain versions, sources and review outcomes for the next step.
Retrieval hooks bring relevant stored context into a session. Startup records and handoffs carry outstanding work and its source references.
The software toolkit stores authorities, procedures, corrections and lessons. Document tools select knowledge for a task; retrieval and correct use still need checking.
Task owners can change direction at checkpoints while preserving the work already done and the instructions that changed it.
Tools check citation records, quotation text, exhibit links, file hashes and document versions. A matching source does not by itself prove a legal argument.
Matched mandatory rules can allow an action, hold it for investigation or deny it. Rules, caller integration and permitted overrides determine what the control covers.
Release checks connect a reviewed change to a published version. Saved checkpoints let work continue after a session ends.
The inspected gate permits actions that match no registered trigger and supports explicit overrides. Its logging is best effort. Stored records, a running check and a correctly enforced boundary are different things. Portable authority envelopes, complete enforcement coverage and externally verified outcome receipts remain development and assurance work.
Michael Darius Eastwood began with his own legal proceedings: connected documents, changing chronologies, disputed facts, authorities and successive drafts. He built Eden Legal AI to organise that work and developed the coordination and checking infrastructure around it. Eden Control brings that infrastructure together; Eden Research uses it to manage research and publication.
A source must remain attached to the statement it supports. An allegation must stay distinguishable from an agreed fact. A proposed edit must not silently become the accepted record.
Context retrieval, task ownership, checks, correction records and version control can support work beyond a single case.
The research programme tests defined correction and acceptance procedures. Practical use motivates the questions; it does not establish their answers.
The operating system and the hosted console are different parts of the work. This is their status as reviewed on 30 October 2026.
Supervised legal preparation, software development and research publication using task records, gates, handoffs and release checks.
Context retrieval, knowledge stores, chronology change proposals, source checks and version records. Integration varies by component.
The web workspace has restricted access. A deployed sign-in page is not evidence that every underlying tool is available through it.
External security review, portable authority records and measured reliability across tasks and providers remain to be established.
The next research question is practical: what information should a checker receive, and what should it be allowed to repair before another agent’s work is accepted? Explore the research and its evidence. No regulator approval or external certification is claimed.
The operating layer Michael Darius Eastwood developed through Eden Legal AI. It connects task ownership, persistent context, source checks, correction records and release controls across his legal, software and research work.
Its founder uses the development system in his own work. The hosted console remains a private beta, and integration differs between components. Own use is not independent evidence of reliability or improved legal outcomes.
For a matched mandatory rule, an unresolved check can hold the action for review. The inspected gate permits unmatched actions and supports explicit overrides, so coverage depends on the configured rules and the caller enforcing the result.
A proposed portable record of who authorised an action, its purpose and its limits. This is part of the development programme; universal envelope enforcement and externally verifiable outcome receipts are not established product capabilities.
Private beta. Public sign-in is not yet open. Existing beta users can sign in to the console.
No. Public examples are synthetic and explain the workflow. The founder’s case material remains private.
The science behind the method: registered predictions, stated before the tests that could refute them.
Visit Eden Research → Eden Legal AISource-linked case preparation, coordinated review and tools developed through the founder’s own legal work.
Visit Eden Legal AI → The founderThe book, the research programme, and the record behind all of it.
Visit his website →Private beta. Public sign-in is not yet open.