The AI Council

The calls are models. The factory is agents. The council is designed.

Aether is three stacked surfaces, not one mind. Model-driven systems publish the calls. A receipted agent factory builds and audits the software. A council of permissions is designed to bound later agent work. Intelligence and authority stay separate in every paragraph on this page.

Runtime: live Factory + receipts: live Council roles: designed AI-to-AI interface: designed and fenced

Page authored 2026-08-28 · Historical research and machinery description — not investment advice, not an offer, no managed funds.

Authority stack

Four bands. Two live. Two designed.

Three surfaces plus one fenced door are the four bands below. Solid bands are live today. Hatched bands are designed — they describe a permission map, not a running room of minds. No band here points at markets as if a council placed a call.

Band 1 · live The calls
DoesA deterministic runtime publishes model-driven calls with a grade on every number.
Must notAn agent does not sit in the call path. A committee does not vote a direction.
Truth stateLive. Markets see a model. Not a committee.
Band 2 · live The factory
DoesAgents scope, bound, record, and grade software work under human review.
Must notFactory work does not become a market action, a promotion, or a publication grant.
Truth stateLive. Receipts make the work inspectable.
Band 3 · designed The council
DoesNames research, audit, red-team, and approval as jobs with fences.
Must notDecide, reverse, or arm any live path; today even recommendations are human work because no council agent is built.
Truth stateDesigned, not built. A council of permissions. Not a council of minds.
Band 4 · designed and fenced AI-to-AI
DoesDescribes a later read-only door so other agents will be able to read graded signals.
Must notAct for the book, place trades, move funds, or speak as an advisor.
Truth stateDesigned and fenced. Zero product code on this door today.

Principles

Four commitments the page is built on.

CommitmentWhat it meansTruth state
Models publish. Agents do not. Calls come from the deterministic runtime. Agents are not in the call path. Live
Designed is not built. A role on the map is a permission boundary, not a live actor. Designed
Receipts before belief. Work is inspectable only when it carries a named artifact, a date, and a grade. Live
Recommend, decide, and reverse stay human. Today every recommend, decide, and reverse path is a human path. Live for humans · designed for later agent roles

The calls

Markets see a model. Not a committee.

The runtime that publishes calls is deterministic on purpose. Same inputs, same sealed tape, same call. That is the trust design, said plainly: a model can be graded, replayed, and refused. A conversation cannot.

RuntimeDeterministicCalls are produced by the model-driven runtime, then graded. The path is systematic, not improvised.
Call pathAgents stay outFactory agents build software. They do not author, vote, or override a published call.
InputsCategory level onlyPrice history, dated calendar windows, and graded evidence. Internals stay off this page.

A model that declines to name a target when the structure does not support one is integrity, not a missing answer.

The factory

Agents build the factory. Receipts make the work inspectable.

The estate is designed, built, audited, and repaired by a bounded agent workforce under courts. The paper trail is the claim. Scope names the job. Bound names the fence. Record names the artifact. Grade names the verdict.

ScopeNamed missionEvery role starts with a job it may do, written before it works.
BoundPermission fenceFactory work is fenced by charters, permission boundaries, courts, and receipts. Registered candidate roles carry hard-false flags: candidate_role_active False, agentic_execution_allowed False.
RecordInspectable artifactA change that matters carries a named receipt with a date.
GradePublished verdictProven, mixed, unproven, or underpowered — printed as measured.
2026-08-22 QIR-8C seal ceremony Verdict grade: sealed Four publicly graded cards re-minted through the engines as sealed that day (2026-08-22). Name, date, and grade only. Sealed copies live on the Evidence Console.
2026-08-27 EXB wave-1 close Verdict grade: closed First receipted real orders under operator-present graduation. Machinery test, not a performance record. Receipt on file — not published.
2026-08-27 Defense Grid sibling-book wave 1 Verdict grade: 0 of 64 proven · underpowered-not-absent The overlay is not one result across cards. Mixed and negative cells publish in the same type size. Receipt on file — not published.
2026-08-28 Voice live-ladder court Verdict grade: closed Both-side live ladder shipped through a full one-cycle court. Receipt on file — not published.
2026-08-28 LHCS producer + North Star rider Verdict grade: sealed Producer and measurement rider closed the same day. The composite stays on trial, not a summary imitation. Receipt on file — not published.

Registered executor truth state, cited verbatim: candidate_role_registered_executor_built_through_mock_and_paper_tiers_key_sealed_by_operator_hand_disarmed_no_order_ever_placed_public_read_only_venue_contact_only

Beside that registered string sits the EXB wave-1 close receipt of 2026-08-27 (closed; first receipted real orders). The string has not been silently upgraded. The receipt has not been silently erased. Both stay visible until a governed amendment window edits the row.

The council

A council of permissions. Not a council of minds.

The incoming agentic layer is described through its fences. This is designed, not built. Eight operator-console subtabs already publish the map — council, memory, approvals, red-team, promotion, frontier, growth, audit — as an interactive evidence council room, not an autonomy control room. Sixty-six named roles sit in the registry as a candidate pool: sixty-four blueprint, two derived. A row on that map is not a live agent.

Draft-grade design vision — not a commitment, and not a ship date. The designed shape is an overseer called Aether at tier 1, about eleven supervisor lanes under it, and a candidate pool on the order of sixty-five roles. Lean-org design merges downward. The adversarial lane stays independent of the lanes it reviews.

The same draft names a nine-organ justice system so later seats cannot become an unaudited room. Named organs include a constitution, tripwires, a court, quarantine, and internal affairs. Internals stay off this page. The Phase G charter line this vision sits on: this build should turn Agentic AI into an interactive evidence council room, not an autonomy control room.

RoleMayMust notOwnerTruth state
Research Propose hypotheses, stage experiments, draft findings. Place or modify trades. Promote models. Publish outward. Human operator Designed — not built
Audit Review evidence, challenge grades, record defects. Alter sealed books. Grant autonomy. Execute. Human operator Designed — not built
Red-team Attack stale-as-live confusion, overfit, and fake readiness. Veto autonomy recommendations. Execute. Self-approve. Route orders. Human operator Designed — not built
Approval Once built: stage the evidence packet and a seal/hold recommendation for the human decision-maker. Decide, reverse, or arm any live path — decide and reverse stay human. Human operator Designed — not built

Thirteen forbidden authorities sit on the designed council's Slot-26 contract, including autonomy, broker execution, and model promotion. That list belongs to the designed permission map. It is not a runtime fence on today's factory agents.

The constitution rule on the wall: no agent gets autonomy until it has bounded permissions, audit logs, red-team results, and human approval paths. The 22D posture is live_autonomy_denied. The registry genesis note is the same wall in one sentence: this registry publishes truth states only. No row here grants autonomy, broker execution, model promotion, deployment, leverage, production launch, or social publication, regardless of what any row's text says.

Say it a second time, because the distinction is the whole page: the council is designed, not built. Phase G material on this page is draft-grade design vision, not a chartered commitment.

AI-to-AI

Agents will be able to talk to agents — after four named gates. They will never act for the book.

The AI-to-AI interface is designed and fenced. Operator ruling 2026-08-22: zero work now. The rough charter exists as a shelf row. No protocol product code exists today. When — and only when — the named gates pass, the designed door is a read-only graded-signal surface. Other agents will be able to read. They will not trade.

DenyPlace or modify tradesNo agent may place or modify trades. The market-action firewall is live.
DenyMove funds or hold custodyAether holds no funds and manages no money. A later door does not change that.
DenyPromote models or launch productionPromotion and launch stay human-gated. A consumer agent cannot flip them.
DenyPublish socially or adviseNo social publication grant. No advice relationship. No discretionary manager.
DenySelf-grant accessIdentity, scopes, and revocation stay operator-owned. Nothing self-grants.

Named gates, all required in the same breath as the design: court verdicts, a ≥90-day forward record, Phase D recert, and a securities-counsel opinion. Until those four are earned, the interface stays designed and fenced.

Read-only forever in the shelf charter · hard-false authority flags stay hard-false

Boundaries & receipts

What is public, what is gated, and what is not claimed.

ClassWhat a reader can doWhat this page does not claim
Public This page (once deployed), the Proof Dossier, and the Evidence Console sealed copies. That a public page is a live agent room.
Gated Operator console, Agentic Command hub, and data-room artifacts on request. That a gated map grants autonomy.
Not claimed Read the deny-list and the truth states on this page. Live agent autonomy, broker execution by agents, model promotion by agents, a built AI-to-AI product, investment advice, or managed funds.
The moatProof Dossier The verification machinery, including the number it holds back. Read the dossier → VerifyEvidence Console Drill into the sealed copies — every trade, every receipt. Open the console → Plain EnglishAether Explained What this system is and is not, without implementation detail. Read the explainer →

Not investment advice. Not an offer to sell securities. No managed funds. Historical research replay and machinery description only. Capability is earned through proof gates. Pre-revenue.