Gaia Registry Meta Audit Series

The 2026-06-15 methodology answered the question "how strong is this piece of evidence?" — assigning each demonstration an Evidence Grade (Bronze, Silver, Gold, Platinum) from a per-row trust number on a 0–100 scale. G7 keeps that floor and adds the next floor up: how strong is the whole skill? A Trust Magnitude (TM) on a 0–500+ scale, computed from the evidence pool, drives the skill's Overall Trust Grade and an explicit nine-predicate Apex gate. Where the 2026-06-15 doc was a pole, G7 is the load-bearing wall built around it.

Status as of publication (2026-06-17): The G7 Trust Taxonomy RFC was ratified on 2026-06-16. Schema, CLI computation, registry backfill, apex demotion, and display surfaces are design-stage — not yet implemented. The live site continues to render the 2026-06-15 model: per-row Evidence Grade (0–100 trust number, S≥90/A≥80/B≥60/C≥40), Overall Trust Grade aggregated as the strongest row in the pool, two skills currently at 6★. This report shows what G7 will look like and what changes when the migration PR ships.

I. What is deployed today vs what G7 will change

The truth, told plainly. There is no point announcing a model that has not landed.

II. Two scales, one ladder

Both numbers are real, and both stay. They live at different layers of the trust model.

The grade letters (S/A/B/C) are the same on both axes. The thresholds are not.

Read the per-row column when you ask "is that one citation strong?" Read the skill column when you ask "does this skill clear the bar?"

III. From a row to a skill — the aggregation flow

   per-row evidence (Evidence Grade)            skill aggregate (Trust Magnitude)
   ───────────────────────────────              ──────────────────────────────────

   ┌─ arxiv paper, peer-reviewed ──┐
   │ trust number 87  → A  Gold    │ ──┐
   └───────────────────────────────┘   │
                                       │   weighted sum, type-diversified,
   ┌─ public repo, 2.3k stars ─────┐   │   sqrt-softened on fusion-recipe
   │ trust number 71  → B Silver   │ ──┼──▶  origins, capped per source
   └───────────────────────────────┘   │
                                       │           │
   ┌─ benchmark result, top decile ┐   │           ▼
   │ trust number 92  → S Platinum │ ──┘   ┌─────────────────────┐
   └───────────────────────────────┘       │   TM = 184          │
                                           │   diversity_types=3 │
                                           │   non_self = 2      │
                                           └──────────┬──────────┘
                                                      │
                                                      ▼
                                           ┌─────────────────────┐
                                           │  Overall Trust Grade│
                                           │       A  Gold       │
                                           │   (≥ 100, < 250)    │
                                           └─────────────────────┘

A single Platinum row no longer flips a skill to S. A skill earns S only when its evidence pool aggregates past 250 with at least three distinct evidence types and at least one non-self-producible row.

IV. What broke on 2026-06-15

The 2026-06-15 methodology was correct as written. Two things were missing.

Anti-auto-mint failure (caught 2026-06-16)

An apex-tier audit ran the regrader on mattpocock/skills (currently 6★). Its frontmatter carries evidence: [] — by design, an apex tier built entirely on its fusion components' shared standing. The regrader silently minted three rows that did not exist in the manifest:

  EXPECTED (frontmatter):                    OBSERVED (regrader output):
  ─────────────────────                      ─────────────────────────
  evidence: []                               evidence:
                                               - { type: github-stars-own,
                                                   trustNumber: 92, grade: S }
                                               - { type: repo-own,
                                                   trustNumber: 78, grade: A }
                                               - { type: self-attestation,
                                                   trustNumber: 60, grade: B }

       TM = 0 (correct)                              TM = 404 (inflated)
       Overall = ungraded                            Overall = S provisional

The same pattern would silently inflate any skill whose graded surface lived only in derived form (fusion recipes, suite roll-ups). G7 §10.14 closes this with a registry-wide anti-auto-mint clause: every grade is re-evaluated under strict-evidence at migration; phantom rows count zero.

No structural ceiling on apex

Pre-G7 the 6★ tier had no programmatic gate. Skills landed at apex by editorial judgement and stayed by inertia. G7 §10.12 introduces a 9-predicate hard gate that all five apex slots must satisfy.

   ┌────────────────────────────────────────────────────────────┐
   │   9-PREDICATE APEX GATE (all 9 must pass to hold 6★)       │
   ├────────────────────────────────────────────────────────────┤
   │   1. TM ≥ 250 (strict-evidence, no auto-mint)              │
   │   2. ≥ 3 distinct evidence types                           │
   │   3. ≥ 1 non-self-producible row                           │
   │   4. K = 2 cross-org cosigns from 4★+ Verifiers            │
   │   5. tenure ≥ 180 days from firstEvidenceAt                │
   │   6. no demote events in the last 6 months                 │
   │   7. system-wide cap = 5 apex slots filled                 │
   │   8. fusion components (if any) all graded ≥ C             │
   │   9. no provisional grade older than 6 months              │
   └────────────────────────────────────────────────────────────┘

At G7 cutover, both currently-6★ skills (mattpocock/skills, ruvnet/ruflo) demote to 5★ — they fail predicates 1, 4, or both under strict-evidence. The tier remains earnable. It is no longer earned. The system-wide 6★ count post-cutover: 0 of 5 slots filled.

V. The four-tier verification ladder (already shipped)

Every skill now sits on one of four verification tiers, computed by the CLI and shown on the skill report. This shipped with PR #709 (4-tier verification workflow) on 2026-06-16, before G7 ratification.

Tiers stack from the bottom — a benchmark-verified skill is also community-verified. The badge surfaces the highest tier passed. This is the floor. G7 builds the apex ceiling on top.

VI. What did NOT change

The 2026-06-15 methodology stays the law of the land for everything below the skill aggregate.

• Per-row Evidence Grades — still S/A/B/C from the same per-row trust number on the 0–100 scale.
• Evidence Types — still the orthogonal "what kind of demonstration" axis (arxiv, repo, github-stars, benchmark-result, peer-review, etc.).
• Inheritance — a named skill's effective evidence is still its own ∪ its starless parent's. The starless parent's grade still floors the named child.
• Star ranks — 0★–6★ still describes maturity. G7 puts a programmatic ceiling on 6★; the other five tiers are unchanged.
• Suite Ultimate Gate — three pillars (≥3 evidenced components, ≥1 S, ≥2 A, no row below C). Unchanged.
• Trust methodology page — the 2026-06-15 report stands. G7 sits on top of it, not in place of it.

VII. What a Verifier sees, before vs after G7

Before (2026-06-15 methodology only)

   garrytan/gstack                            6 evidence rows
   ────────────────                           ─────────────────────
   Overall: A                                   1 × Platinum  (paper)
                                                3 × Gold      (repo)
                                                2 × Silver    (mention)

After G7 cutover

   garrytan/gstack                            6 evidence rows
   ────────────────                           ─────────────────────
   Overall: A                                   1 × Platinum  (paper)      type: arxiv
   Trust Magnitude: 178                         3 × Gold      (repo)       type: github-stars-own
   Verification: enterprise-ready               2 × Silver    (mention)    type: external-cite
   Apex predicates passed: 4 of 9               
   ─────────────────────────                  Diversity types: 3
   Failed: cosign(K=2), tenure(180d),         Non-self-producible: 1
           apex-slot-vacancy(5/5),
           predicate-9                        ◯◯◯◯◯  apex slots filled (0 of 5)

The skill keeps its A. It is also clearly NOT apex-track. The Verifier reads the predicate checklist and knows what's missing.

VIII. Migration shape

One PR, three commits, no big-bang.

   commit 1                  commit 2                   commit 3
   ────────                  ────────                   ────────
   schema + meta             CLI computation            backfill regrade
                                                        + apex cutover
   ─ verification.tier       ─ trust_magnitude()        ─ run regrade
     enum on schemas         ─ apex_gate(9 preds)         (strict-evidence)
   ─ TM thresholds           ─ anti_auto_mint            on all 220 skills
     in meta block             enforcement              ─ stamp meta-report
   ─ apex_slots: 5           ─ K=2 cosign tracking      ─ demote 2 skills
                             ─ 180d tenure base           from 6★ to 5★
                                                        ─ post stamp report
                                                          via gaia-post

No skill loses evidence in the migration. Some skills change grade (the regrade is mechanical from existing rows under strict-evidence). Two skills change rank.

IX. Reading guide

If you are…

• a contributor opening a PR to add evidence — nothing changes. Same gaia dev evidence, same per-row grading.
• a Verifier reviewing a 4★+ skill — read the new predicate column on the report. Cosign with gaia dev evidence --cosign-with if you concur.
• a Maintainer of a currently-6★ skill — your skill demotes to 5★ at cutover. The 6-month grace window opens; if predicates 1–9 all clear before then, you re-promote.
• a reader comparing skills — you now have two numbers. The grade letter for "this evidence is strong" and the magnitude for "this skill clears the bar."

X. References

[1] GAIA Registry. (2026-06-15). The GAIA Trust Methodology: Evidence Types, Grades, and Inherited Standing. meta/reports/2026-06-15-the-gaia-trust-methodology-evidence-types-grades-and-inherited-standing.html

[2] GAIA Registry. (2026-06-16). G7 Trust Taxonomy RFC. Ratified on dev/orchestrator-phase1-closeout, file founder/handovers/G7TRUSTTAXONOMY_RFC.md. Sections cited: §0 (Executive Summary, headline thresholds), §4 (Overall Grade Thresholds and Diversity Gates), §10.11 (transitive-closure fusion-recipe origins), §10.12 (9-predicate hard apex gate), §10.13 (no grandfathering at G7 cutover), §10.14 (registry-wide anti-auto-mint clause), §11.12 (per-skill migration disposition table).

[3] GAIA Registry. (2026-06-16). Benchmark Framework RFC (PR #649, commit 0cbfc2fe). docs/architecture/benchmark-framework.md. Lines 108–113 (score-to-grade mapping) and 188–196 (G7 §0 cross-reference).

[4] GAIA Registry. (2026-06-16). 4-tier verification workflow (PR #709, commit 129ffd49). src/gaia_cli/verification.py. Tiers: community-verified, benchmark-verified, security-reviewed, enterprise-ready.

[5] GAIA Registry. (2026-06-16). Apex-tier audit workflow (wf_f14f7317-972). 7 agents, 595k subagent tokens; caught the auto-mint inflation on mattpocock/skills and motivated the §10.14 clause.