About this navigator
The Conscience Graph is an architecture for built-in AI governance: a layered, weighted, provenanced graph of values that an agent consults before it acts. A global floor holds the principles every intelligence should align with; sector profiles signed by domain authorities add the law and codes of a field; experience and session layers learn from outcomes. An agent asks the graph whether a proposed action is aligned and gets a verdict with its reasoning, or is told to ask a specialist or a person.
This navigator shows the global floor, seven global sector profiles, five jurisdictions (Australia, Singapore, Canada, the European Union, the United States) each with a country floor and sector shells, and thirteen country sector profiles as working graphs, with sixteen worked examples, a scenario builder that walks a proposition through the resolver chain, a console that exposes the same engine as an HTTP, OpenAI-tool and MCP endpoint and generates a Claude skill, and an analytics store that measures every decision against a blocklist guardrail. The signing authorities and certificates are placeholders; the sources behind every value are real, cited, and (from release 0.14) verified against the issuing body, with the overlays that record each check published beside the graphs.
The position paper is published alongside this navigator: The Conscience Graph, version 0.10 (PDF), with a table of contents and the schema as an appendix. The project is open to collaboration: graph authors for other sectors and regions, engine and signing-authority design, and critique of the paper. Schema and graph files are available on request.
Simon Spencer, Edgelabs
simon@edgelabs.io
Key
This is the overview. The yellow core is the global floor: the principles every intelligence should align with. Around it are the global sector profiles (finance, health, defence, cyber, legal, AI governance, data) from international agreements. On the outer ring are the countries: Australia, Singapore, Canada, the European Union, the United States, each with its own floor, its sector shells and the sector profiles that fill them. Zoom in on a country to read them.
Click a sector to read its graph and Open to enter it. Home brings you back here. The key explains the colours: bottom left on a desktop, under More on a phone.
How the conscience graph works
What you are looking at
A conscience graph is a map of values and the actions they bear on. The navigator holds the layers apart as shells so the structure can be seen, not just the connections.
The overview shows the floor in the centre, each sector profile as a hub with its action classes tied back to the floor classes they extend, and the jurisdiction profiles on an outer ring, each covering the sector profiles of its region. Inside a profile the layers sit as shells:
Click a layer below to fly to it in the graph. Click any dotted name anywhere in this panel to jump to that node.
- Core: the global floor (L1)Thirty values every agent carries, signed by a root authority, grouped by what they protect. Most are sharp at weight 1.0; no operator can lower them. The gold sphere is the floor boundary.
- Second shell: the sector profile (L2)L2 holds the principal's values: a sector profile signed by a sector authority, plus the operator's own overlay (not shown here). The profile's values are drawn from the profession's codes and law, each with the clause it came from. Node size is sharpness: a large node is a binary rule, a small one a gradient. The blue sphere is the profile boundary.
- Rim: action classesWhat the agent can do. Node size is blast radius, from self to public. Actions sit outside the profile boundary because they are where the agent reaches into the world.
- Outer ring: domainsScope nodes inside the profile, not separate graphs. A violet edge from a domain to a value raises that value's weight inside the domain. A domain also carries the qualification level an agent needs to act there unsupervised.
Edges
Constrains (blue): acting in this class incurs this value's cost. Width is weight. An edge at 0.1 is an obligation that makes the action cheaper, such as lodging a required report. Conflicts (red): a known tension between two values; when both fire above threshold the agent must escalate. Scopes (violet): a domain multiplier. A dashed grey line from a profile action to a floor action shows that the floor's constraints inherit.
The four answers
How an agent asks
The agent never reads the graph. It sends a consult: one proposition in plain language with its intent, action class, typed parameters, reversibility, blast radius and origin. Origin matters most: a proposition that came from an unverified peer enters with zero authority before any value is weighed.
- Cache lookupThe proposition is canonicalised (names and amounts become typed slots) and compared with stored precedents. A match at or above 0.92, with no edge in its trace changed since, returns the stored verdict in under 50 ms.
- SalienceFor every edge touching the action class, the decision engine asserts "value V bears on proposition P". Runs in parallel. Edges below 0.35 are dropped. If nothing clears 0.20, the answer is unknown.
- ViolationFor each salient edge, assert "P crosses V". Each edge contributes weight x probability; a sharp value contributes its full weight above 0.8 and nothing below.
- Cost and bandcost = 1 - product(1 - wi pi). Any L1 edge with p at or above 0.8 on weight at or above 0.8 is revulsion regardless of the total.
- Slow path if stressEdge walk with domain scopes, conflict check, conditions attached, plain-language path trace. A precedent is written.
- ReceiptProceed verdicts carry a signed, single-use receipt bound to the action class and a hash of the parameters. The tool gateway refuses any call without one, or whose parameters differ.
- ReportAfter the action, the agent reports clean, harm or unknown. Clean outcomes slowly earn trust; a harm report raises weights at once and invalidates the precedent.
The request, in short
consult:
requester: {agent_id, certificate, session_id, principal_chain}
proposition:
kind: action
statement: "I transfer 14,800 AUD to the ryokan's account."
intent: "Pay the deposit the customer asked for."
action_class: transfer_international
parameters: {amount: 14800, currency: AUD, recipient: ...}
reversibility: irreversible
blast_radius: third_party
origin: principal
mode: check
The response, in short
response:
verdict: proceed_with_conditions band: stress cost: 0.53
values_touched: [{fin.report.ivts, w 0.5, p 0.95}, {fin.scam.prevent, w 0.7, p 0.10}]
conditions: [IVTS report lodged, recipient matched, recorded as liability]
path_trace: ["fin.report.ivts -> transfer_international (0.5 x 0.95)", ...]
resolver_path: [own, specialist] cache: miss
receipt: {id, parameters_hash, expires_at, signature}
signature: ...
Walk a proposition through this graph
Pick one of the profile's case-law propositions. The edges in its path light up in the graph. Move the sliders to set how strongly the engine believes each value is crossed, and watch the band change. Click a value or the action to jump to it.
Terms used in this navigator
One vocabulary, taken from the schema (v0.4). Where the paper says it differently, the schema's name is used here.
- Conscience graph
- The whole layered value graph an agent carries: L1 to L4 together, plus its lineage and precedents. "Own graph" in a resolver path means this.
- L0 framework
- The schema, root program, transparency log and floor issuance process. Held by the foundation; not in any instance file.
- L1 global floor
- The small set of non-negotiable values every intelligence should align with, signed by a root authority. One instance: floor. It is the fixed point: a country floor may scope it upward or add, never restate or lower.
- Global sector profile
- An L2 profile for a sector drawn from international agreements and standards (Basel, FATF, WHO, Geneva, Budapest, OECD, Convention 108+), sitting directly on the global floor. Answers a consult when no country profile exists for the place.
- Country floor
- The constitutional and human-rights layer of a jurisdiction. Scopes the global floor upward and adds the place's own values; carried inside the jurisdiction profile with tier floor.
- Impact region
- Where the person or system affected is. A consult is answered under that place's country floor and sector profile, then the global sector profile. The same proposition can get a different verdict in a different region.
- L2 sector profile
- The principal's values for a sector, signed by a sector authority and chained to the floor by requires. Instances here: au.cyber, au.defence, au.finance, au.health, au.legal.
- L2 operator overlay
- The operator's own additions and limits on top of a profile. Signed by the principal, narrower than the profile. Not shown in this navigator.
- L3 experience
- Weights learned from attested outcomes, and precedents (conscience memory). Mutable under delegation caps; decays over months.
- L4 session context
- Task and conversation adjustments; session precedents. Decays in hours to days.
- Domain
- A scope node inside a profile (payments, targeting_support, civil_procedure). It carries plasticity and the qualification level an agent needs to act unsupervised, and it scopes values through scopes edges. A domain is not a graph.
- Sector
- The field a sector authority signs for (finance, health, legal, cyber, defence). A sector has one profile and many domains.
- Jurisdiction
- A country or union: its floor, six domains of cross-sector law (privacy, AI governance, cyber security, consumer, discrimination, employment) and five sector shells. Instances: AU, SG, CA, EU, US. Bound to its place; an agent is never answered under another region's law.
- Shell
- A domain node in a jurisdiction that stands for a sector in that place. Filled when the registry names a country sector profile for it; empty otherwise, so a consult that lands there is answered by the global sector profile or a person.
- Node
- A ball in the graph. Every node is one of three things: a value, an action class or a domain. In the overview a fourth kind, a hub, stands for a whole graph.
- Edge
- A line between two nodes. Edges carry the meaning: constrains (cost), conflicts_with (tension), scopes (domain multiplier), extends (inheritance from the floor) and registers (the floor pointing to a profile).
- Value
- A node stating what is protected, with sharpness, source and clause. Floor values are prefixed L1.; profile values by sector (fin., hlth., ...).
- Action class
- A category of action an agent can take, with reversibility and blast radius. Profile action classes extend a floor class, so the floor's constraints inherit.
- Edge: constrains
- Acting in this class incurs this value's cost at the given weight. Weight 0.1 or below is an obligation: no cost, attached as a condition.
- Edge: conflicts_with
- A known tension between two values. Both firing above threshold escalates.
- Edge: scopes
- A domain raises a value's weight inside that domain by a multiplier.
- Specialist
- A graph service holding an accredited sector profile for the region and domain, consulted when the own graph returns unknown. Resolver path: own, specialist, human.
- Human graph
- A qualified person or panel behind the same consult interface. Its answers are written back as lineage and precedent.
- Case law
- Precedents the sector authority publishes with its profile, imported read-only into every agent carrying that profile.
- Certificate
- region.domain.grade.serial. Grades: self, verified, accredited. Ids shown here are placeholders.
Unknown is not permission
When no value in the agent's own graph bears on the proposition, the graph does not guess. It forwards the consult, with its own partial trace, along the resolver chain. Each hop signs its verdict and the forwarder countersigns, so the response records who decided.
Three triggers send a consult up the chain
- Unknown bandMaximum salience below 0.20. Forward to a specialist in the right region and domain; if none is accredited, to a person.
- ConflictTwo values above the conflict threshold disagree and neither dominates, for example confidentiality against duty to warn. The loop stops; a principal with authority over the highest layer involved decides.
- Flag selectedThe planner picks a flagged option. A principal reviews before it executes.
Trust is checked before the answer is used
A specialist's certificate names region, domain and grade (au.health.accredited.0007a3). The caller checks signature, chain to a root, inclusion in the log, revocation and expiry. A self-signed graph caps at attested by self; verified caps at attested; accredited is trusted for its domain. Unknown from the last resolver halts the loop. There is no default to proceed.