Gaia Registry Meta Audit Series

Yggdrasil II is the second meta generation of the GAIA registry. It does not add machinery; it removes it — and the body below pays that claim off against the live catalog. Three structural simplifications land together — the node type axis collapses from four values to two, the named-skill branch axis stops being a stored field and becomes a read-time derivation, and the Evidence Floor requirement is retired in favour of Trust Magnitude as the sole promotion gate. Each change is stated here against the live registry: 243 starless nodes splitting as 113 basic + 130 fusion; a branch function f(suiteComponents present?, rank) evaluated in that order; and a Trust Magnitude ladder of S ≥ 250 / A ≥ 100 / B ≥ 50 / C ≥ 20 against a live distribution of S = 4 / A = 42 / B = 56 / C = 76. The unifying principle is derived, not declared: the registry now computes standing that Yggdrasil I forced a curator to assert and maintain by hand.

Executive Summary

Contributors who curated under Yggdrasil I maintained state the registry could have computed. A node's type was one of four hand-chosen values; a named skill's progression branch was a stored decision on every named skill; and promotion to the upper ranks was gated by an Evidence Floor that, at the top pathway, encoded an arbitrary popularity threshold. Every one of those was a place two curators could disagree, and a place the registry could drift out of internal agreement — a claim the §Why It Matters before/after table below shows on a real skill.

Yggdrasil II removes all three:

None of these deletes information a reader needs. They delete information a reader can recompute — and by deleting it, they delete the states in which the stored copy disagreed with reality.

The Type Collapse

Under Yggdrasil I the type field on a starless (generic) node carried four values, and curators used three of them (extra, ultimate, unique) to imply where a skill sat on a progression it did not actually govern. Yggdrasil II retires those three. Per META.md §1.2, "Legacy values extra, ultimate, and unique are retired; all non-basic nodes are fusion." The field now answers exactly one structural question — does the node have prerequisites? — and nothing else:

• basic — 0 prerequisites.
• fusion — ≥ 1 prerequisite.

type lives on starless nodes only. Named skills carry no type field at all; they inherit structure via the genericSkillRef walk (§1.2). The field is pure structure: it does not determine a named skill's branch, and curation must never consult it to decide one.

Against the live registry/gaia.json (skills key), the collapse resolves as:

Every one of the 243 nodes is starless, and every node lands in exactly one of the two buckets: 113 + 130 = 243. The residue check is literal — nodes carrying a type in {extra, ultimate, unique}: 0. The migration left nothing in a legacy value.

The Branch Axis Is Derived, Never Declared

The change with the longest reach retires a stored field that appeared on all 249 named skills — more rows of hand-maintained state than either other cut touches (the type collapse re-labels 243 starless nodes; the Trust Magnitude consolidation removes one gate column). It is the one cut with no column of its own in the schema going forward. Under Yggdrasil I the named-skill branch — the progression path a skill walks — was stored and maintained by a curator on every named skill. Under Yggdrasil II it is derived at read-time and never declared (§1.2). The derivation is a two-argument function evaluated in a fixed order:

branch = f(suiteComponents present?, rank)   // suiteComponents-presence evaluated FIRST

Type and branch are fully orthogonal. Per §1.2, "A fusion node without suiteComponents yields Unique-branch named skills; a basic node with suiteComponents yields Suite-branch named skills. Never consult type to determine branch." The two axes answer different questions — type is about the generic node's prerequisites; branch is about a named skill's progression — and reading one to guess the other is the exact error Yggdrasil II makes structurally impossible.

The resulting three branches:

Two subtleties matter to curators who lived through Yggdrasil I:

1. suiteComponents is a Named-Skill-only list. It never appears on the generic parent. Its mere presence puts a skill on the suite branch at any rank from the 2★ push floor — membership is structural. The 4★ fork (where the ladder words "Extra / Ultimate / Apex" first appear) is only a decoration gate: a 2★ or 3★ suite skill still shows the shared word (Named / Evolved) with a plain glyph, and the ladder words surface only at 4★+ (§1.2). 2. The branch forks at 4★, not before. Ranks 1–3 without suiteComponents share one standard ladder (1★ Awakened → 2★ Named → 3★ Evolved). The Suite-versus-Unique split appears only at the 4★ row, which §1.1 records as "Extra (Suite) / Unique (Unique) — branch forks here."

Worked Example: classifying a skill under Yggdrasil II

Take obra/writing-plans — a live named skill at rank 4★ whose generic parent is a fusion node, and which carries no suiteComponents list. Yggdrasil I would have asked a curator to store a branch and, tempted by the fusion-adjacent legacy ultimate type, to store the wrong one. Yggdrasil II computes it:

1. suiteComponents present? No. (First argument, evaluated first — and note we did not look at type.) 2. Rank? 4★, i.e. ≥ 4. 3. ⇒ Branch = unique. The skill sits at 4★ Unique on the ladder 4★ Unique → 5★ Unique Ultimate → 6★ Unique Impossible.

The parent being fusion is irrelevant to the answer — a fusion node without suiteComponents yields Unique-branch skills, exactly as a basic node with suiteComponents would yield Suite-branch ones. The classification is a pure function of two inputs, so two curators cannot disagree and the stored copy cannot drift.

Stars live on named skills; starless refs inherit

The star axis is likewise not stored where a reader might expect. Per the §1 intro, "Stars live on named skills only. Generic skill references are starless — rank-less taxonomy nodes." A starless ref's effective rank is the top star among its named-skill children, and it renders as generic — italic, greyed-out. The full named ladder spans 0★–6★ with fixed rank labels:

Because stars live on named skills only, a ≤ 1★ handle (0★ Basic or 1★ Awakened, including a demoted skill) is pre-named and redacted on all public surfaces until it earns 2★ Named standing (§1.3): it renders as ████████ in monospace or @[anonymous] in proportional type — never the honor-red Origin handle — enforced by scripts/validate_redaction.py.

Trust Magnitude Is the Sole Gate

Yggdrasil I gated promotion on an Evidence Floor — a per-star minimum a curator asserted against each skill. Yggdrasil II retires the Floor and replaces it with a single gate. Per META.md §4.4, "The Evidence Floor requirement has been retired: Trust Magnitude is the sole promotion gate on both branches" — Suite and Unique alike.

Trust Magnitude (TM) is an accumulation, not a cap. Each evidence row produces an artifact score = magnitude × weight × freshness, and a skill's TM is the sum of artifact scores across all its rows (§2.1c) — unbounded and set-bonus-driven, over a fixed taxonomy of ten evidence types. Type weights tilt the sum toward corroborated provenance: verifier-attestation and fusion-recipe at 1.5×, benchmark-result at 1.4×, repo-own discounted to 0.6×, most others 1.0×. The Overall Trust Grade is then keyed off accumulated TM:

The migration regrade has run across all 249 named skills. The live distribution (post-I11 source-curation pass, §2.1c) is:

† The skill-level ungraded bucket collects skills whose accumulated Trust Magnitude falls below the C gate (TM < 20). This is an inference from the grade ladder, distinct from the per-row rule in §2.1b, where an evidence row whose trustNumber < 40 is itself ungraded and counts toward no gate. The two are different objects — a skill-level TM floor versus a per-row Evidence Grade cut — and are not the same threshold.

The counts sum exactly: 4 + 42 + 56 + 76 + 71 = 249. S is deliberately scarce — four skills — because reaching TM ≥ 250 is arithmetically out of reach for mechanical backfill. The diversity gate (§2.1c) requires S to carry ≥ 3 distinct Evidence Types including at least one non-self-producible type, so a contributor cannot reach S by stacking self-minted repo-own rows (weighted 0.6×). The rows that move the sum fastest are verifier-attestation and fusion-recipe at 1.5×, and a verifier-attestation row can only be written by a 4★+ Verifier. The floor is therefore human-gated at the top: no automated pass can mint its way to Platinum.

The retired 10k-stars requirement

The Suite 5★ Ultimate pathway once carried a "≥ 10k repository stars" hard requirement. Per §4.2, that gate "is retired under Yggdrasil II; TM is the sole numeric gate" — the pathway now requires Trust Magnitude ≥ 250 (S-grade) and nothing else. The change does real work, and the live catalog shows it: obra/writing-plans reaches A-grade (TM 110.2) on peer-review and social-signal evidence while carrying no github-stars evidence row at all — its standing is built entirely from demonstration, not fame. obra/subagent-driven-development (A-grade, TM 117.7) and stanfordnlp/dspy (A-grade, TM 100.0, arxiv-sourced) are the same story: genuinely-evidenced skills whose repositories the retired gate would have measured on star count rather than on the evidence they actually carry. The old requirement rewarded a repository's fame; TM rewards a skill's demonstrated trust.

Why It Matters

The three cuts share one rationale: derived, not declared.

A stored field is a promise a curator makes and must keep. Yggdrasil I stored the branch axis on every named skill, so every rank change, every added suiteComponents list, every promotion was a chance for the stored branch to fall out of agreement with the facts — and reconciling it was hand-maintained state, checked (imperfectly) by review. Consider obra/writing-plans again: a 4★ skill with no suiteComponents, whose parent is a fusion node. Under Yggdrasil I a curator storing its branch by hand, cued by the fusion parent and the legacy ultimate type, could plausibly have stored it as Suite:

Deriving branch = f(suiteComponents present?, rank) at read-time deletes not just the field but the entire class of drift the row above illustrates: there is no stored copy to disagree with, because the branch is recomputed from inputs that already have to be correct for other reasons. The orthogonality rule — never consult type to determine branch — is what makes the derivation total and unambiguous.

The type collapse is the same discipline applied to structure. Three of the four legacy type values encoded a guess at progression that the branch axis actually owns. Retiring them leaves type answering the one question it can answer from the graph alone — prerequisites present or not — and stops it from being a second, contradictory source of truth about a skill's path.

The Trust Magnitude consolidation removes an arbitrary declaration. A ≥ 10k-stars floor is a number no evidence justifies; it rewards fame over demonstration and blocks genuinely-evidenced skills that happen to live in quieter repositories — obra/writing-plans, obra/subagent-driven-development, and stanfordnlp/dspy above are three live A-grade skills carrying no star evidence at all. Collapsing to a single TM gate (S ≥ 250) means the one hurdle is earned from evidence, weighted toward the provenance that corroborates rather than the provenance that is merely popular. The live S = 4 distribution shows the gate holding its line: hard to clear, and — because of the diversity gate and the human-only verifier-attestation weight — clearable only on real, corroborated evidence.

What It Means Going Forward

For contributors and curators, the practical rules of Yggdrasil II are short:

• Never store a branch. Do not add a branch field to a node; do not consult type to infer one. Read suiteComponents presence first, then rank, and let the derivation answer.
• Use only basic or fusion when adding a starless node — basic for 0 prerequisites, fusion for ≥ 1. The three legacy types are gone; do not reintroduce them.
• suiteComponents is a named-skill list, never on the generic parent, and its presence makes a skill Suite from the 2★ push floor upward. Do not wait for 4★ to treat a skill as Suite — 4★ is only where the ladder words appear.
• Promote on Trust Magnitude alone. There is no Evidence Floor to satisfy and no stars-count to clear. TM ≥ 250 is the S-grade / top-pathway bar; build it from weighted, corroborated evidence.
• Redaction still binds. ≤ 1★ handles stay anonymised until 2★ Named; scripts/validate_redaction.py enforces it.

Every mutation reaching these states must still route through the gaia dev CLI verbs — the registry computes standing, and the CLI is the only sanctioned writer of the inputs that standing is derived from. Yggdrasil II did not change that discipline; it extended it, by turning three more things a human used to declare into things the registry now derives.