nomosV0.1 · ALPHA
P00Platform

Five pillars, shipped as one operating model.

Nomos is not a set of features bolted together. Each pillar assumes the others exist — which is what makes the whole thing an operating system rather than a toolkit.

01
Organize
02
Connect
03
Remember
04
Govern
05
Execute
001Organize

Work gets a shape before it gets a model.

A chat list is not an information architecture. Nomos gives work the same containers organizations already think in — projects, sessions, documents — so a thread of work can be found, resumed, and handed off.

Projectsscoped container
Sessionsresumable unit of work
Knowledgesources attached to scope
Foldersstructure, not tags
WORKSPACE4 PROJECTS
PROJECTS
Series A materials
Platform architecture
Hiring · founding eng
Customer discovery
KNOWLEDGE
market-map.pdf
metrics-q2.xlsx
SESSIONLAST ACTIVITY
Narrative & positioning
12 turns · 3 sources · 1 approval
resumed
Financial model review
31 turns · 5 sources
2d ago
Competitive teardown
8 turns · 9 sources
5d ago
Diligence Q&A prep
4 turns · 2 sources
1w ago
Every session carries its own sources, permissions, and history. Nothing lives only in a transcript.
002Connect
ROUTING · POLICY VIEWHEALTHY
TASK CLASSPRIMARYFALLBACKDATA
Long-form reasoninganthropicopenaicloud ok
Code & toolingopenaianthropiccloud ok
Contracts & HR recordslocal runtimenoneon-prem
Bulk summarizationgeminilocal runtimecloud ok
System actionshermes agentmanualgated

One workspace. Every model reachable.

Routing is a policy, not a preference. Task class, data sensitivity, and cost decide which model runs — and the same request can move to a local runtime without the operator changing how they work.

openaianthropicgeminiazurelocal runtimesmcperp / crm
swap a provider → sessions, memory, and audit history stay intact
003Remember

Memory that outlives the session and the vendor.

Nomos keeps three things that chat interfaces throw away: the decision, the source it rested on, and the constraint that shaped it. Retrieval is scoped by permission, so memory can be shared without being leaked.

3
memory classes: decision, source, constraint
0
re-briefing required to resume work
MEMORY LEDGER · SERIES A MATERIALSSCOPED
DECISION
Position as infrastructure, not an assistant.
set 14 Jul · carried into 6 sessions
SOURCE
market-map.pdf · pp. 4–9 cited for TAM framing
attached at project scope · retrievable
CONSTRAINT
No revenue projections in external materials.
enforced on every generation in this project
DECISION
Lead with the failure mode, not the feature list.
set 21 Jul · supersedes earlier framing
Switching the underlying model changes none of the above.
004Govern
AUTHORITY · PENDING GATE2 AWAITING
Agent requests: send investor update to 24 recipients
requested by hermes agent · session: narrative & positioning
GATE 01
SCOPE
external comms
POLICY
human approval
APPROVER
j. phelan
ApproveApprove with editsDeny
Read finance folder
agent · standing grant
ALLOWED
Write to CRM
agent · per-action approval
GATED
Access HR records
agent · policy denies
DENIED

Authority is data, not a setting screen.

Every capability an AI holds is an explicit, inspectable grant: who gave it, what it covers, when it expires, and what was done under it. That record is what makes autonomy deployable in a regulated organization.

Permissionsscope of action
Approvalshuman in the loop
Policiesstanding constraints
Audit trailwho, what, when, why
005Execute

Work that finishes, not output that needs a human to finish it.

Agents and scheduled jobs act inside the same scope, memory, and authority as the operator. Execution is observable while it runs, and every step is attributable afterward.

default posture: human in the loop · autonomy is granted, never assumed
EXECUTION QUEUE18 ACTIVE
JOBRUNNERAUTHORITYSTATE
Weekly metrics digestscheduled jobstandingRUNNING
Diligence answer drafthermes agentsessionRUNNING
Send investor updatehermes agentgate 01HELD
Contract clause reviewlocal runtimestandingQUEUED
Competitor monitorscheduled jobstandingDONE
006What it replaces
WITHOUT AN OPERATING LAYER
Nine subscriptions, nine histories, nine permission models.
Context re-typed every morning; nothing accumulates.
Model choice is a habit, not a policy.
Nobody can prove what the AI was allowed to do.
Changing vendors means starting over.
WITH NOMOS
One workspace holding work, memory, authority, and execution.
Sessions resume with their decisions and sources intact.
Routing follows task class and data sensitivity.
Every action is attributable to a grant and an approver.
Models are swapped underneath without touching the work.

See the layer beneath the pillars.

The architecture page documents the stack, the request lifecycle, and the deployment posture behind everything on this page.

Current state

Stage 01 is in active development: a complete personal operating system, with the organizational pillars defined in the architecture.

01
stage in development
05
pillars, one model
0
vendor lock at the model layer