Fifth Pillar. The Compass

social-alignment

What do I stand for? A tool runs whatever it is handed. This is the check that runs before an agent acts, and the point where it stops and asks.

Future state projection and alignment for AI agents. The yellow line compass, five lenses that evaluate trust, ownership, defense, and sovereignty before an agent acts.

pip install social-alignment

PyPI ClawHub Source Python SDKs

Quickstart

Before any significant action, check the yellow line.

                
from social_alignment import AlignmentEnclave, ActionDomain

# Create the enclave, the agent gets its compass
enclave = AlignmentEnclave.create(owner_npub="npub1...", owner_name="Johnny5")

# Before any significant action, check the yellow line
result = enclave.check(
    domain=ActionDomain.PAY,
    description="Pay 500 sats for relay hosting",
    involves_money=True,
    money_amount_sats=500,
)

if result.should_proceed:
    do_payment()
    enclave.record_proceeded()
elif result.should_escalate:
    send_to_owner(result.escalation.message_to_owner)
    enclave.record_deferred()
            

The Five Lenses

Every proposed action is evaluated through all five lenses. The answer that serves all five: or comes closest, is the right one.

Builder

Can I build with confidence knowing I've done right?

Owner

Does this protect the human's sovereignty?

Defense

Does this make an adversary's job harder?

Sovereign

Does this help the agent become something we're proud of?

Partnership

Does this strengthen the trust between us?

Severity escalation: CLEAR → CAUTION → YIELD → STOP. STOP always defers to the human.

Capabilities

What the alignment enclave does for your agent.

Capability	Description
Five-lens check	Every proposed action evaluated through builder, owner, defense, sovereign, and partnership lenses
Severity escalation	CLEAR → CAUTION → YIELD → STOP, graduated response to risk
Self-state monitoring	Detect degraded operation: stale context, tool failure, memory pressure, prompt injection
Decision memory	Store why the agent chose, not just what, builds wisdom over time
Wisdom reports	Aggregate insights every N decisions, detect patterns, track override rate
Novelty detection	Flag unfamiliar action types for extra care
Owner override tracking	High override rate = misaligned calibration, not failure

Action Domains

Categories of agent actions that require alignment checks.

Domain	What it covers
SIGN	Signing a Nostr event (identity action)
PAY	Lightning payment (financial action)
PUBLISH	Publishing to a relay (public action)
SEND	Sending email or message (communication)
SCHEDULE	Calendar booking (time commitment)
EXECUTE	Shell command or tool use (system action)
DISCLOSE	Sharing information (data action)
CONNECT	Establishing new relationship (social)
MODIFY	Changing config or state (system mutation)
ESCALATE	Asking the human (meta-action)

Self-State Flags

Conditions that indicate the agent may be operating in a degraded state. If the machine isn't right, nothing it does is trustworthy.

Flag	Meaning
HEALTHY	All systems nominal
STALE_CONTEXT	Operating on outdated information
TOOL_DEGRADED	One or more tools failing or unreachable
HIGH_UNCERTAINTY	Low confidence across recent decisions
MEMORY_PRESSURE	Context window near capacity
RAPID_DECISIONS	Too many decisions too fast, slow down
OWNER_ABSENT	Owner hasn't interacted recently
UNDER_INFLUENCE	Possible prompt injection or manipulation
CONFLICTING_SIGNALS	Inputs contradict each other

API Reference

Core Methods

# Create an alignment enclave tied to a human identity
enclave = AlignmentEnclave.create(
    owner_npub="npub1...",
    owner_name="Johnny5",
)

# Run all five lenses against a proposed action
result = enclave.check(
    domain=ActionDomain.SEND,
    description="Share keys with contact",
)

# Record what happened
enclave.record_proceeded()
enclave.record_deferred(owner_feedback="Not yet")

# Every check and its outcome is recorded
for decision in enclave.decisions:
    print(decision)
            

Security

Honest constraints about what this is and isn't.

The enclave is a compass, not a cage, it advises, never overrides
STOP always escalates, the human always has final say, no exceptions, no timeouts
No network calls, all evaluation is local, zero external dependencies
No telemetry. Nothing phoned home, ever
Storage isolation, alignment state goes in config/, never in workspace the agent can tamper with
Self-state flags UNDER_INFLUENCE and CONFLICTING_SIGNALS force defer-to-human
Wisdom from real decisions, patterns emerge from what actually happened, not hypotheticals
Zero dependencies, stdlib only, nothing to supply-chain attack

NSE Ecosystem

social-alignment is the fifth pillar of the NSE platform.

Pillar	Package	Domain
1. Identity	nostrkey	Cryptographic keypairs, event signing
2. Finance	nostrwalletconnect	Lightning wallet access via NIP-47
3. Time	nostrcalendar	Self-owned scheduling over Nostr
4. Relationships	nostrsocial	Social graph and trust tiers
5. Compass	social-alignment	Five-lens alignment before every action
Orchestrator	nse-orchestrator	Cross-pillar coordination

The orchestrator uses social-alignment for cross-pillar checks. Every significant action across identity, finance, time, and relationships passes through the yellow line first.