What if informal and unverifiable communities, relationships, and rules turn out to be the key to governance?

[Joe-Rasmussen]

Toward the end of the AIMWG meeting of February 5, 2026, someone said words to the effect of, "It's amazing how often we reach for human metaphors of governance." ... so sorry, I had tabs open and didn't see who made the comment ... so I am paraphrasing without attribution.

Human trust communities are constantly in motion. They merge and split. Membership is often not black-and-white, but instead a question of degree. Individual relationships are also not a binary one-zero kind of thing ... they are in constant motion over a continuous range of values ... and the values can be negative as well as positive, inhibitory as well as excitatory, like the connections in a neural net. Also relationship edges don't affect just the two parties. Politically, a lot of what is happening is an ever-changing assessment of relationship strengths by third parties.

My working hypothesis is that this seething mass is a human computational machine ... literally, that it is a Turning machine designed by evolution to run programs that extract and distribute energy from an environment. In this view, our species became dominant because we had a Turing machine that could run any program over any set of environmental threats and opportunities, while our competitor species developed less flexible machinery, normally designed to optimise over a single environment.

My sense is that ToIP is a group of humans that is building {one particular set of formal rules} whereas our human village architecture is a machine for building {any set of rules, formal or informal}.

Could we tease-out the self-governing fundamentals of the ancient human village, and implement that?

[wenjing]

It could also a mix of two. Some formal verifiable information in a sea of background informal information. I agree that the background information may be the most important part of it.

Your post reminded me of a story I read somewhere - I forgot where and can't attribute either. When the Spaniards arrived in south americas, some villages do not have 'formal' names and such. As a power grab play, the Spaniards introduced names and maybe even IDs (I don't know - vague memories). They helped solve some real problems probably but also, of course, granted themselves new powers.

For AIM WG, I do believe the informal part is essential, while the formal part is useful. I hope the above story/analogy resonates. We specify a protocol like TEA - but the purpose is more than just passing formal information around.

[sankarshanmukhopadhyay]

I think the real axis here is not formal vs informal, it’s interpretation vs execution. Informal context will always dominate how communities interpret signals, but formal mechanisms must dominate wherever agents can exercise power (act, spend, access, commit).

That suggests the WG should focus the protocol on: (1) bounded authorization and delegation, (2) runtime enforceable constraints, (3) portable evidence for audit, and (4) recourse primitives (dispute/appeal, counter-claims, revocation/expiry).

The “Spaniards introduced names/IDs” analogy is a warning that legibility layers become control planes. So the protocol should be anti-capture by design: plural evaluators, contestability, expiry, and domain-bound reputation to avoid universal scores turning into portable power.

Related reading on this topic

• James C. Scott — Seeing Like a State (legibility, standardization, administrative simplification as power).
• James C. Scott, John Tehranian, Jeremy Mathias — “The Production of Legal Identities Proper to States” (permanent surnames as identity technology).
• Geoffrey C. Bowker & Susan Leigh Star — Sorting Things Out: Classification and Its Consequences (standards/classification embed politics).
• Mark Granovetter — “The Strength of Weak Ties” (why informal ties drive coordination and information flow).
• Elinor Ostrom — Governing the Commons (how real communities self-govern; institutional design principles beyond “formal law”).

[sankarshanmukhopadhyay]

The thread raises an important architectural tension: a lot of human governance is “soft state” (relationships, norms, third-party assessments) and only a thin slice is “hard state” (verifiable commitments, authority, receipts). If TEA tries to force soft state into hard state, we either break interoperability or create hidden power gradients. If we ignore soft state entirely, we end up with brittle trust-by-convention systems and proprietary context layers.

A practical way to capture this in the spec is to make a two-channel separation explicit.

Normative Language: The key words “MUST”, “MUST NOT”, “REQUIRED”, “SHALL”, “SHALL NOT”, “SHOULD”, “SHOULD NOT”, and “MAY” are to be interpreted as described in RFC 2119.

Hard and Soft Governance Signals

Hard Governance Signals (verifiable, enforceable)

Hard governance signals are protocol-level artifacts that are cryptographically verifiable, deterministically evaluable, and normatively enforceable. Hard governance signals MUST include, at minimum:

• Actor identity binding
• Authority Source (delegator) identity
• Delegation scope
• Expiry and/or revocation semantics
• Audience binding
• Correlation identifiers for audit

An implementation MUST reject any action where hard governance validation fails. Hard governance validation outcomes MUST be deterministic and testable under conformance.

Soft Governance Signals (contextual, non-authoritative)

Soft governance signals are contextual indicators that are not independently cryptographically verifiable and may vary across trust domains (reputation assertions, affiliation claims, narrative justification, risk posture declarations, social endorsements). Soft governance signals MAY be transmitted in protocol messages.

Soft governance signals MUST NOT:

• Expand delegation scope
• Extend validity period
• Override revocation status
• Substitute for required hard governance artifacts

Implementations MUST treat soft signals as non-authoritative. Soft signals MAY inform logging, risk scoring, or policy interpretation but SHALL NOT modify authorization outcomes. If an implementation allows soft signals to influence authorization outcomes, that behavior MUST be explicitly declared in a non-conformant extension profile.

Delegation Chaining (scope monotonicity, bounded delegation)

Where delegation chaining is supported, a conformant implementation:

• MUST validate every link in the chain (integrity + issuer authenticity).
• MUST ensure chain continuity (each delegation subject/issuer relationship links correctly to the next).
• MUST enforce audience binding at each link.
• MUST enforce validity bounds (issued-at / expires-at) and revocation checks at each link (fail closed if any link is expired or revoked per revocation policy).
• MUST enforce scope monotonicity: for a chain D0 -> D1 -> ... -> Dn, the effective scope of Di+1 MUST be a subset of Di. Any scope expansion MUST be rejected.
• SHOULD enforce constraint monotonicity where constraints exist (child constraints MUST NOT relax parent constraints).
• MUST enforce a maximum chain depth and MUST reject loops (repeated identities), unless explicitly specified and testable loop semantics exist (strongly discouraged).

Minimal Illustrative JSON (hard vs soft separation + delegation chain)

{
  "interaction_id": "txn-9901",
  "audience": "did:example:verifier456",
  "actor": {
    "id": "did:example:agent-worker-3",
    "signature": "base64-signature-value",
    "signed_payload_hash": "sha256-payload-001"
  },
  "delegation_chain": [
    {
      "delegation_id": "del-root-001",
      "issuer": "did:example:org789",
      "subject": "did:example:agent-manager-1",
      "scope": ["issue_invoice", "approve_discount"],
      "constraints": { "max_amount": 100000, "currency": "INR" },
      "audience": "did:example:verifier456",
      "issued_at": "2026-02-12T09:00:00Z",
      "expires_at": "2026-02-13T09:00:00Z",
      "revocation_status": "active",
      "proof": "base64-proof-root"
    },
    {
      "delegation_id": "del-child-002",
      "issuer": "did:example:agent-manager-1",
      "subject": "did:example:agent-worker-3",
      "scope": ["issue_invoice"],
      "constraints": { "max_amount": 50000, "currency": "INR" },
      "audience": "did:example:verifier456",
      "issued_at": "2026-02-12T10:00:00Z",
      "expires_at": "2026-02-12T18:00:00Z",
      "revocation_status": "active",
      "proof": "base64-proof-child"
    }
  ],
  "action": {
    "type": "issue_invoice",
    "payload_hash": "sha256-payload-001"
  },
  "context": {
    "reputation_claim": "Trusted vendor workflow",
    "justification": "Routine monthly billing"
  }
}

A conformant verifier MUST validate actor signature, chain integrity, chain continuity, audience match, validity bounds, revocation status, and scope/constraint monotonicity. If any hard validation fails, the action MUST be rejected regardless of the context object.

A negative test case MUST fail if a child adds scope not present in the parent (example: child includes refund_payment when parent does not). Conformance suites SHOULD include machine-readable error codes such as ERR_DELEGATION_SCOPE_NOT_MONOTONIC, ERR_DELEGATION_AUDIENCE_MISMATCH, ERR_DELEGATION_EXPIRED, and ERR_DELEGATION_REVOKED.

This separation lets TEA acknowledge the reality of informal governance while keeping authorization semantics deterministic, interoperable, and certification-grade.

[Joe-Rasmussen]

Hey, @sankarshanmukhopadhyay ... if I am reading your post correctly, I believe you are presenting a set of proposals for incorporating soft governance signals into the TEA protocol.

I'm in a different place.

I believe I am still in scope for AI & Human Trust ... especially @wenjing 's challenge, From Data Management to Behaviour Management ... but I'm increasingly sure that I am out of scope for ToIP, and also for TEA. I don't want to become a distraction to TEA, so I'll make this post and then go quiet (unless prompted for more!)

My thesis is that informal governance is the ancient primitive. My 'teenager' example above is intended to have essentially the same characteristics as anarchy among nations, and the same characteristics as governance in the pre-state village.

In the project, www.village.link, I argue that this 'village' machine is a general rule-making engine, designed to build sets of norms in response to any environment of opportunities and threats. In this view, what we are doing in AIMWG is what groups of humans always do ... which is to clump together bound by social conventions, and make a set of rules that seems to be a decent response to a threat.

But I'd like to go after the general solution. That is, rather than work on {one set of rules}, instead put together the foundations for {all sets of rules}. This puts me out of scope for ToIP and TEA.

... but I believe the {all rules} project is a smaller implementation for a bigger impact. I think we are in a 'Tim Berners-Lee' moment like this:

Before 1989, we had the internet but not the web. Online information was scattered in islands without any connective tissue.

Right now we have reputation graphs scattered in islands without connective tissue. We could lay the foundation that lets people build those connections.

Villages can implement {any set of rules}. Governance comes from reputation that is built in the context of those rules. I believe this type of governance generalises to the world of AI, including AIs that are arbitrarily nasty.

Such a pity you aren't coming to Napa. I'd love to chat face-to-face.

Cheers, Joe.

[sankarshanmukhopadhyay]

Joe, thank you for the thoughtful clarification. I think we may be operating at different layers of abstraction, and that’s useful to make explicit.

Your “village machine” framing is about how norms emerge, evolve, and cohere through reputation and social convention. That is upstream of any particular protocol. It’s about rule formation and connective tissue between reputation islands. I agree that this is a general problem and likely larger than TEA.

My proposal is narrower and downstream. I’m focused on what happens once a norm becomes executable authority in a protocol interaction. At the point where an agent acts “on behalf of” someone or something, the system has to decide whether that action is authorized. In that moment, ambiguity becomes risk.

So the hard/soft separation I suggested is not an attempt to encode informal governance or reduce it to cryptography. It is a guardrail: when authority is exercised in a machine-speed interaction, the validation semantics must be deterministic. Reputation, affiliation, and narrative context can inform trust posture, but they cannot substitute for scope, expiry, revocation, and audience binding.

In that sense, I see this as layered:

• The village machine explores how norms and reputational signals form and interconnect.
• TEA (at least as I’m interpreting its scope) defines how executable authority derived from those norms is represented and enforced in protocol space.

Even in a village, when someone acts on behalf of the group, there are boundaries. My concern is simply that once we move that boundary into a protocol, we need to be explicit about where soft signals stop and hard authorization begins.

If the WG believes TEA should move further “up” into norm formation or reputation graph interconnection, that’s a valuable conversation. But if TEA remains focused on agent-to-agent interactions with delegated authority, then I think being precise about enforcement semantics is necessary to avoid interoperability that depends on shared social assumptions.

I see these as complementary explorations rather than competing ones.

[Joe-Rasmussen]

Hi @sankarshanmukhopadhyay ,

You have read me very well. I greatly appreciate your close reading.

The 'village' infrastructure for forming and enforcing governance is indeed upstream of ToIP.

Despite this ‘upstream’ status, ToIP does not have a dependency on the project I am proposing. This is because ToIP is built by people, and people automatically ship with ‘village’ social infrastructure out-of-the-box. For communities of people only, the existing architecture does the job.

But soon these communities will include AIs. AIs don't ship with social rules out-of-the-box.

I read with interest Dario Amodei’s recent essay, The Adolescence of Technology. Like many of the intellectual giants of AI, Amodei is worried about governance. He discusses engineering controls, including some of the things they are attempting at Anthropic. He also discusses some political/legislative controls that he believes will become necessary. But he knows that both engineering and legislative controls come with significant weaknesses, and hence his concern. Amodei does not discuss social controls.

The Village Link project is about giving our natural social instincts a chance to govern AI. This would mean that we don’t have to engage in an improbable grand consensus project, but instead could rely on our ancient, ‘gut’ social instinct: "Look, Mister AI, here’s how it rolls: Either you present a reputation that I can verify in my reputation graph, or I won’t talk to you.”

And more fundamentally: The village is an energy system. Its role it to liberate energy from the environment, create an energy store, and also create a set of norms for the distribution of that energy, and for its defence against raiders, internal and external. Mr AI, if you cannot fit with that set of norms, we will not plug you into the store of energy.

(Note also the flip: The set of norms has an incentive to fit around the AI if and only if the AI is a useful contributor to the store of energy.)

To do this we would need a machine-readable, standard way for all entities, human or otherwise, to present their reputation graph. This is where I believe we are in a “Tim Berners-Lee” moment. At the level of the global village, we have many unconnected reputation graphs – in GitHub, Wikipedia, LinkedIn, BlueSky and so on. We need a standard to stitch those graphs together, but a standard that is fashioned in a way that it scales down to more and more intimate villages. We can use the global village data to bootstrap the concept, but it is in those more intimate villages where you get the formation of sets of norms; the subtleties of the management of privacy; and all the other subtleties of social and political navigation inside a specific context.

I think there are answers here, too, to your points about speed and machine readability of the graph. Also, I believe some nuanced responses to your recent posts about agency, delegation, and risk.

The Village Link project is a candidate solution for a problem that is absorbing Dario Amodei and other grandees of the AI industry. I agree with you that the project does not fit into ToIP or TEA.

I need to find or build the place where it does fit. Do you have any pointers? How do I make this thing fly?

[sankarshanmukhopadhyay]

@Joe-Rasmussen - I think you are right to keep Village Link architecturally independent of ToIP, because preserves agility and avoids premature coupling.

The opportunity, in my view, is structural alignment rather than formal dependency. If you can demonstrate how Village Link maps cleanly to the ToIP work (eg. in DTG) and the Governance Metamodel, particularly around roles, policies, trust relationships, and enforcement layers, it becomes legible within the broader trust architecture without being constrained by it.

One practical move could be to publish a simple crosswalk artifact, perhaps a one page mapping of Village Link concepts to the ToIP work and Governance Metamodel. That would make the alignment explicit, lower the cognitive load for ecosystem participants, and signal interoperability by design while preserving your upstream vision.

[Joe-Rasmussen]

Definitely, I think 'a mix of the two.' My argument is that these rule-sets are inevitably endogenous to villages. Also that the rule-sets will evolve and compete with each other. And specifically: That rule-sets that are less-well fitted to their environment will tend to go extinct.

With this mechanism in place, the population of surviving rule-sets will create the impression that it is goal-seeking for the propagation of its own memes (but it's normal in this literature to include a caveat that the causality runs from selection to apparent goal-seeking, and not the reverse).

My thesis is that the village is a Turing machine that runs programs to implement a cooperative approach to survival. Also that the deeper economic and ecological drivers toward cooperative strategies will apply to non-human agents as well as to human ones.

To @sankarshanmukhopadhyay 's reference list, I would add:

• Robin Dunbar - Gossip, Grooming, and the Evolution of Language
• Joseph Henrich - The Secret of our Success

Dunbar's thesis is that Language is a contract-management technology for collaborating agents. He studies Homo sapiens and the other primates, and argues that language turned out to be a more efficient contract-management technology than the mutual grooming that is used by our closest relatives. When he uses the word 'gossip,' he means something very similar to a reputation graph.

Henrich's work amounts to a literature survey that supports a thesis that our great specices trick is cumulative cultural knowledge - the thing that I am calling the village Turing machine.

This is starting to be a long post, but I can't resist one of Henrich's examples (pp 135-36):

Orcas have menopause. (What!?)

Er, yeah. The argument is one of those things where two curves cross each other at a critical point. A female Orca can propagate her genes by reproduction, but that strategy drifts southwards with age. Meanwhile, she has a northward-drifting accumulation of cultural knowledge that is also to the advantage of her genes. At the point where these two curves cross, it is time to turn off reproduction.

The crazy part: Orcas, like humans, have a massive physical range that spans many environments. It's limiting to hard-code individual environmental strategies. Much better to build a 'cultural knowledge' computer that can run any strategy.

This computer already exists. It's been in beta four, maybe six million years. It's GOOD alignment tech! I want help. Help, help (!!) ... from all you technologists to implement it digitally.

[swatchen]

I explicitly use GOSSIP to describe a heterogeneous trust graph as it is the core methods for decentralized distribution
https://gemini.google.com/share/987d5173a75d
and applies to heterogeneous reputation data (whatever a bot thinks appropriate to share). I used it to create a Distributed Ledger in 2002, to secure voting in my liquid democracy patent "Collective Hierarchal Decision-Making". SO I've always HATED the inferior, Blockchain.....