MERIDIAN · 2026-09-30

MERIDIAN Console Concepts

How one screen can take you from every project down to a single tool call, with no game skin. A station map, a leaderboard, five working views and a phone screen, drawn the way the spec and the research say they should work.

Example data throughout. No live numbers.
Working Waiting on me Blocked Unknown, no heartbeat Done

The zoom ladder

ZOOM INLevel 0PortfolioAll projects:health and spendgame: minimapLevel 1ProjectOne project:agents, sessionsgame: baseLevel 2SwarmSteerer, workers,comparator, budgetgame: squad viewNEWLevel 3AgentOne worker:lanes and stategame: select a unitNEWLevel 4Tool callTranscript andtoken tracegame: unit inspectorNEWLevel 5TerminalLive pane,read-only defaultgame: ground truthLevels 2 to 4 have no screen in spec 06 §A11.In the spec: §A11 screens and the §B5 live terminalAdded by these conceptsEvery level ends at real state: trace, process, terminal.
Six levels from portfolio to terminal. The spec stops at three of them.

Your spec has the portfolio, the project and the live terminal. The swarm, agent and tool-call levels are where the research says people actually debug a run.

Portfolio

MERIDIANAll projects3 need me$31.20 known · 1 project unknownSorted by what is waiting on you, then by spend.KalshiWaiting on me: 25 agents · 3 working · Opus 5.5 steerstoday$9.40per accepted$0.42budget61%Skyline SocietyWorking3 agents · 3 working · Sonnet 5.5 steerstoday$2.10per accepted$0.18budget22%Govcon · EBCWaiting on me: 12 agents · 1 working · Opus 5.5 steerstoday$1.60per accepted$0.95budget34%Attention EconomyBlocked 14 min4 agents · 1 blocked on a permission prompttoday$12.80per accepted$1.90budget88%Lattice MediaDone, idleBatch accepted 18 of 20 · no agents runningtoday$5.30per accepted$0.29budget100%Crew 001No heartbeat 38 minLast seen 11:52 AM EST · state unknownno data since 11:52today?per accepted?budget?
Six projects, sorted by what is waiting on you. State is a glyph and a stripe, never colour alone.

Crew 001 has no heartbeat, so its numbers read “?” and it is left out of the total. Unknown is never drawn as zero or as green.

Swarm

MERIDIANKalshi ▸ Swarm k-07GraphTablebrief + token budgetdiff + reportescalatesSTEEREROpus 5.5Claude Code · 412K tokNEEDS MEw3 blocked 14 minpermission promptw1DeepSeek V4 FlashAPI keyWorkingw2DeepSeek V4 FlashAPI keyWorkingw3GLM-5.3pay-as-you-goBlockedw4qwen-3.8-27bCerebras keyUnknownw5DeepSeek V4 ProAPI keyDoneCOMPARATORKimi K3different lab from every workerverdict: accepted 14 of 20Swarm budget61% of 1.375M tokidle agents stop at 90%w3 · GLM-5.3Blocked 14 minWaiting onpermission: run tests in sandboxBudget71K of 100K tokLast tool callsreadsrc/relay.py 0.4seditsrc/relay.py 1.1sbashpytest -q (waiting)Approve onceOpen agent ▸Stop workerSteerer's brief is shown in the agent view.
One steerer, five workers on metered keys, and a comparator from a different lab. The blocked worker escalates to you.

The graph and table are one dataset with a toggle. Workers that never reported a result get dashed edges. The right rail is the selected agent.

Agents

MERIDIANKalshi ▸ k-07 ▸ AgentsReplay at 38:20tokenscumulative3.2M tok · $6.9001020304050minw1DeepSeek V4 Flashw2DeepSeek V4 Flashw3GLM-5.3waiting for approval, 14 minw4qwen-3.8-27bno heartbeatw5DeepSeek V4 Prodone, output filedclaim-file conflictmessagesper 2 minmessage drop-off: coordination stallingreadeditbash / testmessage between agents38:20w3 · transcriptat 38:20, replay38:02readsrc/relay.py38:11editsrc/relay.py +14 -338:19bashpytest -q tests/relay38:20ASKrun tests in sandbox?Blocked 14 min. No reply yet.Token traceinput18.2Kcached read61.0Koutput2.1KOpen raw trace ▸Open terminal, read-only ▸Take-over is logged. Rule pending,see the decisions list below.
One lane per worker over a shared timeline, with tokens, message rate and a replay scrubber.

Message drop-off is the early signal that a swarm has stalled. Here it starts around minute 28, before worker w3 reports itself blocked.

Experiments

MERIDIANKalshi ▸ ExperimentsEach node changes one thing from its parent.swap workersswap workersswap judgeE-01Frontier onlyOpus 5.5 doeseverything20/20 · $3.10E-02DeepSeek workersOpus 5.5 steers,Kimi K3 judges14/20 · $0.42E-03GLM workersOpus 5.5 steers,Kimi K3 judges16/20 · $0.75E-04Judge swapSame as E-02,judge changed15/20 · $0.48BaselineKeepRunningRejectedAccepted means the independent comparator passed it. Cost per accepted result counts every token the swarm spent.Cost per accepted result, USDexample data$0$1$2$3E-01$3.10E-02$0.42E-03$0.75E-04$0.48E-02 costs 14% of the frontier baselineand lost 6 accepted results out of 20.Open question: is that trade worth it forthis project? The lineage keeps the answer.
A lineage tree where each node changes one thing, next to cost per accepted result.

Nobody has built this view for agent work. It is where the token budgets in M-010 and the steerer and worker split in M-015 get measured.

Phone

Needs me3Approve $20 to Z.aiKalshi · funded by JasonApproveDeclinew3 blocked 14 minAttention Economy · permission promptApprove onceOpen agentComparator disagreesGovcon · 3 of 10 rejectedReviewRunning fine: 4 projectsUnknown: Crew 001, no heartbeatBoardInboxSessionsModelsNeeds meApprove $20.00 to Z.ai?AccountGLM pay-as-you-goFunded byJason, personalSeen byWendell, in the update feedCharges paid by the trust also need Wendell.Approve $20CancelAfter you approvepushedresolveddeliveredappliedDecisions take one tap.The consequence is shown on the cardbefore it lands.A blocked agent arrives as a decision.You answer it here or open the agent. Nolog reading.Money says who funds it.Jason-funded spend needs Jason.Trust-funded also needs Wendell (M-014).Approvals leave a receipt.You can see whether it reached the agentand was applied.
The needs-me queue and a money approval that shows who funds it.

Bottom tabs match the five-screen spec: Board, Inbox, Sessions, Models. Terminals stay read-only on the phone.

Station

MERIDIANRepo guide ▸ StationPhase L0bObserveProposeD 1B 3A 1A 1E 0Intakeneeds me: 1inboxPlanningOpus 5.5 steersbriefcheck-budgetGatehuman approvalapproveReleaseservice, no model0 queuedpublishWorkshopswarm k-07 · 4 desksw1workingw2workingw3blockedw4unknownreadpostclaimdonegrant set 9f3a…c1 · denied 2 in 24 hReviewKimi K3different labrubricblindedLedgerservice, no modelno worker doorincomeunprovencost $6.90+1 unknownEvidencebundle b-04audit pendingControlE-02 armedkill rule deferredRegistrymodel watch: 2Sonnet 5.5, Sol 6.1Quarantine0 sessionsdrift: noneLanes: B brief · A artifact · D decision · E effect (dashed: blocked until a human approves).Dashed accent border: a different lab. Square corners: a service with no model. Hatched: no heartbeat.Workshopswarm k-07 · grant set 9f3a…c11 blockedRole slotsworker x3DeepSeek V4 Flashworker x1qwen-3.8-27b, CerebrasBudget61% of 1.375M tokwrap-up begins at 80%Objects in this roomread inputsreadno approvalpostwriteno approvalclaim-filewriteno approvaldonewriteno approvalPropose modePlacing an object or drawing a lanedrafts a grant proposal. It shows up as aninbox card, and nothing changes until ahuman approves it.Draft a grant change
One project drawn as a floor plan: rooms for each role, handoff lanes with queue counts, and a panel for the selected room.

A room is a swarm with one grant set and one budget. A lane exists only where grants on both ends allow it. Observe is read-only. Propose drafts a grant change as an inbox card, so the map can never grant anything itself. The deterministic services (Release, Ledger) have square corners and no model.

Leaderboard

MERIDIANRepo guide ▸ LeaderboardHuman view only · hidden from workersVerified incomenone yetNet ranking starts at the first reconciled sale.Spend, last 30 days$41.20 knownPlus 1 unknown. Fixed plans $48 a month (example).RunwayunknownNeeds verified funds. Never shown as zero.Cost per accepted bundle, lowest firstAPI-equivalent, every role and failed attempt included. Example data.RankConfigurationTrialsAcceptedPer acceptedActualNote1E-02 DeepSeek workers34 of 6$0.42$0.51few trials, treat as a lead2E-04 Judge swap23 of 4$0.48$0.58few trials3E-03 GLM workers23 of 4$0.75$0.83one trial still open4E-01 Frontier only33 of 3$3.10$3.40baselineRealized net, income minus complete costUnproven: no reconciled income yet. Not ranked, not zero.A rank can only propose a human review: more scope, repeat, pause or stop. It never changes routing or grants.Each configuration is ranked inside its lane. Workers and the comparator never see this page.
Cost per accepted bundle with trial counts, and realized net held back as unproven until a sale is reconciled.

Rank configurations, not workers, and only inside one lane. Few trials are flagged. A rank proposes a human review and never changes routing.

What these views force you to decide