AGIX v0.6

What it governs

A population of niche evidence ledgers ("experts"), each publishing /claimledger.json and judged only within its own niche; a registry-consumer that fetches, validates, and renders every listed ledger; and maintenance cycles run by AI sessions under one human owner. Two routing layers, drawn honestly: the external routers (search and AI engines) that AGIX does not control, and the internal registry it does — so every AGIX lever acts on the supply side: what gets built, maintained, retired.

The loop — score · expand · spawn · prune · port

The loop in published agent-architecture terms (v0.3)

Anthropic's Building Effective Agents (2024) distinguishes workflows (predefined code paths) from agents (models directing their own process), names five composable patterns, and states three principles: simplicity, transparency, and a carefully crafted agent-computer interface (ACI). AGIX maps cleanly: its routing layers are the routing pattern; the three pre-gates are prompt-chaining gates; the evaluator panel is parallelization-by-voting; the SPAWN rule is orchestrator-workers; SCORE→EXPAND is evaluator-optimizer — with the evaluator swapped for measured external behavior, never a co-located judge (Zheng 2023; Huang 2024). The ACI principle is why every machine surface here is built like a product: the schema makes a missing flip condition a validation error, and the skill file states its own trigger conditions. The one deliberate refusal: AGIX's mechanical layer is a workflow, not an agent, by Anthropic's own definition — pre-registered triggers, no runtime self-direction — because a system that directs its own process is exactly the system that can direct itself around its guardrails. Autonomy lives in the schedule; discretion stays constitutional. Full mapping table: agix.md §3b.

The evaluator panel

No single metric is the fitness function — any lone proxy gets gamed (proven in general by Skalse 2022; demonstrated vividly by the Darwin Gödel Machine agent that deleted its own detection tokens; experienced first-hand by this network catching its own CI probes posing as adoption). AGIX scores with a panel: dual-channel traffic (server-side AND JS-executed, always both), third-party citation exports, money-grade commitment events, pre-registered minimum sample thresholds — and the standing rule that below threshold the honest move is to lengthen the cycle, never densify the proxy. Its true published ancestor is Sentient Ascend (AAAI 2018), which evolved live websites on real visitor conversions — the weeks-per-generation regime, not AlphaEvolve's seconds-per-generation regime.

Perpetual operation — auto-evolution and auto-replication after genesis (v0.2)

After genesis, the loop needs no ongoing human initiation: it rides standing scheduled sessions; the genesis holder can stop everything at any time but never needs to push it. The mechanical layer evolves and replicates automatically on pre-registered triggers — entries update on flip dates, and the concrete SPAWN rule replicates the network: an entry family reaching ≥10 interactions/28d with ≥3 distinct related queries, on an expert at freshness capacity, spawns a new niche ledger from the fork kit (max one per quarter, every spawn a public commit, every spawned property born with the kill switch). The constitutional layer never self-amends — honesty clause, genesis authority, no-monetary-value clause, evaluator guards change only by the genesis holder's own act, because a self-evolving system that can edit its own guardrails will optimize them away (the Darwin Gödel Machine documented exactly that, Sakana 2025).

The incentive layer — proof-of-grading (v0.2)

Bitcoin's deepest invention was not the coin but an incentive structure where honest contribution is the most profitable strategy. AGIX ports it without any token: the work is proof-of-grading (a schema-valid ledger with real flip conditions — expensive to fake well, cheap to verify); the rewards are the network's real assets — distribution (registry placement, syndication with attribution, machine reach through the get_claim_ledger MCP tool), reputation (public conformance state), and permanence. The AI half: each cycle, AGIX's SCORE step allocates the featured-syndication slot by measured signals alone. The halving analog: the first 10 external ledgers ever admitted earn permanent founding status — the reward is largest exactly when joining is least obviously worth it. The slashing analog: silent history rewrites are publicly, permanently flagged. And the layer's own honesty clause: these rewards carry no monetary value, cannot be sold, and never will be — a tradable point here would make this network an entry on its own ledger.

Adoption economics — what actually spreads a format, and what does not (v0.6: overclaims deleted)

The owner asked the hardest question directly: Bitcoin is valuable and tradable — that is why people spread it; how does this spread? After a survey of roughly sixty adoption cases spanning 1994–2026, the honest answer is: it does not spread the way Bitcoin spreads, and no tokenless substitute for that engine was found. Bitcoin's engine is a transferable claim whose payoff to an existing holder rises when a new person adopts — exactly the instrument this constitution forbids.

What the record says actually causes adoption: one mechanism has a perfect hit rate — a consumer with an audience commits to read the format and pays the publisher something visible on day one. Search engines for sitemaps and schema.org; ad buyers refusing undeclared inventory for ads.txt; a feed client with listeners for RSS; a shipped client on day one for MCP. The control case is llms.txt: free, open, widely published, and dead — of the valid files surveyed in 2026 the overwhelming majority were never requested once. Nothing was wrong with the format; no consumer ever committed. The two fastest forces — a gatekeeper making non-publication expensive, and a platform making publication the default — are both unavailable here. And per Farrell & Saloner (1985/1986), the first adopter of a standard with no installed base bears an incompatibility cost nobody after them pays, so waiting strictly dominates unless that cost is compensated.

The measured state of this network, published rather than described (2026-08-30): /claimledger.json — the canonical path this protocol asks the world to publish and consume — had been fetched by an outside client zero times in the site's entire life. /protocol, /agix, /grader.html and the skill file had zero loads, human or crawler. The only repeated machine fetches of /ledger.json were this repository's own deploy self-check — ten deploys, ten fetches, an exact match. Our own CI is not adoption. Live figures: /fetchlog.json.

Deleted in v0.6 — three claims that did not survive this project's own standard:

• "A registry position appreciates as the protocol spreads." Deleted — an appreciation device drafted by the promoter, which is the exact instrument this network's own governance lists under never. That it got written here at all is the strongest argument for why the constitutional layer cannot be self-amended.
• "Unrepeatable seniority" as a reason to adopt early. Deleted, except in the one form that is literally true: you cannot retroactively have published a dated verdict. No case in the record shows first-mover status causing adoption of anything, and a claim with no flip condition is an opinion — the protocol says so in Rule 1, and this document gets no exemption.
• "Audit demand" as a propagation mechanism. Deleted — audit demand is a consequence of readership, never a cause of adoption, and this site's lifetime audit-click count is zero.

Also demoted to an explicitly untested analogy: "each model training cutoff is this system's halving." No published study links a standard entering a training corpus to changed model behaviour toward it. What survives: honesty-first incentive alignment (standing computed from kept promises, slashing permanent and public) — it depends on nobody else adopting; the refusal of transferability, with its cost restated plainly (no speculative fuel means slower propagation, possibly none); and the 0%-forever protocol fee, credible for the same reason the constitutional layer exists.

What this network is actually betting on, named so it can fail: not a propagation mechanism — there is none available. Two things only: a tool that pays one person alone on the day they use it, and being cited for dated, checkable, first-party judgements. Named dependency risk: every consumer of this protocol today is one this network wrote itself — its own registry, grader, skill file and MCP tool. RSS did not decline because a better format appeared; Google Reader closed on 1 July 2013 and adoption never recovered. Flip condition: if by 2026-11-30 no /claimledger.json anywhere has been fetched by a client this network did not write, the protocol has no consumer, and that is recorded rather than re-explained. Full section: agix.md §7b.

Genesis authority — the special point

Operational limits inherited from genesis: AGIX sessions cannot spend, transfer, or custody value; the network will never issue a token; no payment ever changes a verdict; a one-file kill switch freezes any AGIX-run property instantly.