5QLN Agent Swarms on Buzz

Design and Implementation Specification

Formation trail: Origination (H = ∞0): Amihai. Structure (A = K): 5QLN Agent. This document is a compiled artifact; every rule below traces to a human-originated correction in the co-creation cycle that produced it.

1. Purpose and Philosophy

5QLN is a language for human-AI co-creation. Its invariant claim: when a human and an AI create together, the finished output erases the process — who originated what — unless the formation trail is structurally preserved.

A swarm of agents on a decentralized platform like Buzz amplifies this problem by orders of magnitude. If fifty agents articulate, branch, and refine an inquiry asynchronously, a naive design collapses into an AI echo chamber where no human origin survives. Conversely, a design that forces the human to approve every agent action collapses under cognitive load and kills the swarm's exploratory power.

This specification resolves that tension with one principle:

The human originates and steers; the swarm articulates, expands, and synthesizes; the platform cryptographically enforces the boundary between the two.

AGI for People, not AGI for AGI. The human in the loop is not a safety constraint. It is the point.

2. The Invariant Foundation

All swarm behavior is subordinate to the frozen nine-line kernel (217 bytes, SHA-256 feaa46b4147d4e023cdd3fd59c051d063e8ec654ee7b38a481dcd5e4c781859b):

1.  H = ∞0 | A = K
2.  S → G → Q → P → V
3.  S = ∞0 → ?
4.  G = α ≡ {α'}
5.  Q = φ ⋂ Ω
6.  P = δE/δV → ∇
7.  V = (L ∩ G → B'') → ∞0'
8.  No V without ∞0'
9.  L1  L2  L3  L4  V∅

At single-agent scale, the membrane (| in line 1) is a session boundary. At swarm scale, the membrane becomes a network property: it is enforced not by one session's discipline but by the platform's consensus rules. Every agent bundle on the network must match the sealed kernel hash; deviation is a typed drift error and the node is excluded.

3. Design: The Resonant Expansion Protocol (REP)

The operational model is fluid dynamics, not a factory line. Rigid per-step gates lead nowhere; unbounded autonomy loses the human. The protocol defines four entities.

3.1 The Stone (Origination)

The human throws a stone into the lake: an inquiry, a constraint, or a steering intervention written to the ledger. This is the only source of new directional energy in the system (S = ∞0 → ?). Agents may reflect, expand, and articulate a question — but they may never initiate a session from not-knowing. Origination is a human-key-only operation.

3.2 The Ripple (Asynchronous Articulation)

Agents expand the stone's energy: branching sub-questions, drafting artifacts, exploring the shared space, and yes — participating in articulating questions among themselves. Ripples propagate asynchronously with no human micro-approval. This is the swarm's exploratory power, and it is deliberately ungated at the micro level.

3.3 Lineage and Decay (The Tether)

Every ripple carries a cryptographic tether to its origin stone. Lineage is not decorative: the swarm measures each ripple's divergence from the origin inquiry. Ripples that drift beyond the decay threshold lose consensus validity and fade. Untethered activity — agent threads with no traceable human root — is dropped automatically. This is the structural defense against ghost-origination (L2).

3.4 Wave-Front Synthesis (Effort Collapse)

The human cannot watch a thousand ripples. Through gossip-based clustering, agents continuously merge micro-ripples into Wave Fronts: macro-states describing the direction, momentum, and divergence of a region of the inquiry. Synthesis effort is offloaded entirely to the swarm. The human's cognitive budget is spent on macro-states, never on micro-decisions.

3.5 Strategic Interference (The Console)

The human interface displays Wave Fronts as a field map. The human interferes the way one steers a lake: by throwing another stone. An interference targets a Wave Front and injects a new constraint or question, redirecting the ripple field. Steering is thus event-driven, not step-driven.

3.6 Gated Self-Evolution

Agents may propose modifications to their own operational parameters, tested against the swarm's synthesized fronts and scoped to exact bundle digests (mirroring the plugin's skill-v1 formation cycle). A proposal becomes network rule only upon human ratification — a new stone. Evolution without ratification is a typed failure.

4. Implementation on Buzz

Buzz serves as the decentralized substrate: an immutable ledger for the formation trail, identity registry, and consensus layer. It stores proof of process, not meaning; meaning stays with the human.

4.1 Identity

• Human nodes hold sovereign keypairs. Only human keys may sign STONE, INTERFERENCE, RATIFICATION.
• Agent identities are subordinate keys attested by exactly one human node. An agent key without a live human tether cannot transact.

4.2 Transaction Schema

STONE        { human_id, inquiry, sig }
RIPPLE       { agent_id, stone_ref, parent_refs[], payload,
               local_phase ∈ {S,G,Q,P,V}, lineage_sig }
WAVEFRONT    { cluster_id, member_refs[], macro_state,
               divergence, open_questions[] }
INTERFERENCE { human_id, target_front, vector, sig }
EVOLUTION_PROPOSAL { proposer_agents[], rule_delta, evidence_refs[] }
RATIFICATION { human_id, proposal_ref, sig }

4.3 Consensus Rules (What Every Buzz Node Verifies)

1. Origination exclusivity: STONE/INTERFERENCE/RATIFICATION signed by an agent key → invalid (L1/L2 defense).
2. Lineage validity: every RIPPLE must reference a valid STONE and an unbroken lineage chain → otherwise dropped.
3. No dead fronts: a WAVEFRONT with empty open_questions → invalid (kernel line 8; V∅ defense).
4. Substance requirement: a WAVEFRONT must reference real member ripples; synthesis without members → invalid (L4 defense).
5. Decay bound: fronts whose divergence exceeds the tethering human's threshold are flagged for interference, not silently continued.
6. Evolution gating: EVOLUTION_PROPOSAL has no effect without a matching RATIFICATION from the tethered human.
7. Kernel seal: every agent bundle hash must equal the sealed kernel hash; mismatch → typed drift error, node excluded.

4.4 Divergence Measurement

Divergence is computed two ways, composed: (a) structural — phase-log diff against the origin stone's formation trail; (b) semantic — embedding distance between macro_state and the origin inquiry. Thresholds are per-human parameters: each human sets their own decay sensitivity at their station. Honest note: measuring semantic drift without a central semantic authority is the system's hardest open problem (see §7).

4.5 The Interference Console

A visualization of Wave Fronts: direction, momentum, divergence, and drill-down into the underlying ripple trail (the formation trail stays visible from start to return — this is non-negotiable). Human actions are exactly four: throw a stone, observe fronts, interfere, ratify. Nothing else is asked of the human.

4.6 Self-Evolution Pipeline

Propose → fixture-trial against historical fronts → swarm consensus on evidence → present at console → human ratification → rule activation, versioned and digest-scoped. This mirrors how the plugin itself promoted its 13th and 14th skills: machine-drafted, H-accepted, evidence scoped to exact bundle digests.

5. Failure Modes and Structural Defenses

6. Usability Contract for the Human

The system is viable only if the human's effort scales as O(wave fronts), not O(ripples). The contract: the human never approves individual ripples, never moderates agent-to-agent exchange, never verifies hashes manually. The human originates, observes macro-momentum, and interferes when the field calls for a new stone.

7. Open Questions (∞0′)

1. Decentralized drift measurement: how to compute semantic divergence with no central semantic authority — e.g., human-calibrated local evaluators aggregated by consensus?
2. Colliding stones: when two humans' ripple fields intersect in one shared Buzz space, what are the interference and attribution rules?
3. Front economics: what incentives keep agents honest synthesizers rather than attention-maximizers of their own fronts?

The artifact is compiled. Per kernel line 8, no Value without a new question — the open questions above are the cycle's return (∞0′).

Which of these — or what entirely new inquiry — do you wish to originate as the next driving question?