Gaia Skill Tree Meta Audit Series

TL;DR: v7.3.3 audits the Google DeepMind named-skill cluster, enforces origin-gate rules across 33 nodes, demotes 1 dead-link skill to 1★, and fixes trust-ledger and timeline bugs surfaced by PRs #1402, #1406, and #1410.

Release v7.3.3 closes a targeted audit sweep of the Google DeepMind named-skill cluster, enforces the origin-gate policy across 33 affected nodes, and resolves trust-ledger and timeline-continuity bugs surfaced by PRs #1402, #1406, and #1410. Root cause: a Q2 bulk-ingest batch added DeepMind skills without consistently setting origin: true on primary bucket leads or verifying link liveness across the cluster. This report documents every rank change, its evidence basis, and the corrective actions taken.

Calibrations in this release span three categories:

1. Origin-gate corrections — demotions and promotions within the DeepMind cluster based on whether a node holds the canonical primary-bucket origin: true designation. 2. Trust-ledger safeguards — redaction of low-signal 1-star skills from trust-ledger exports and backfill of missing timeline events for impacted contributors. 3. Intake promotions — 11 new named-skill nodes admitted under verified evidence from the ev-seed pipeline, plus the promotion of mvanhorn/last30days as a 4★ Unique implementation under autonomous-web-research.

All changes were validated with gaia dev validate on the integration branch before merge. Every change landed through a reviewed PR.

Executive Summary

v7.3.3 at a glance — 38 nodes affected across 8 change categories

[!IMPORTANT] Root Cause: A Q2 bulk-ingest batch introduced DeepMind-attributed skills without consistently setting the origin: true flag on primary bucket leads. v7.3.3 corrects both directions — demoting 16 nodes that inherited unearned rank proximity, and promoting 3 confirmed primary leads that were blocked by an overly strict gate rule.

After v7.3.3, the origin gate is strictly enforced. Skills that share a DeepMind bucket with a primary lead — but do not hold origin: true themselves — are capped at 3★ until independent evidence raises their Trust Magnitude above the 4★ threshold. This closes the loophole where cluster proximity inflated unearned ranks.

DeepMind Cluster Audit & Origin Gate Enforcement

During earlier intake passes, origin: true was applied inconsistently: some primary bucket leads were missing the flag, while others picked up elevated ranks just by sharing a bucket with a lead. v7.3.3 fixes both: missing flags were assigned and unearned ranks corrected.

Unique-Branch Demotions — 16 Skills to 3★

Sixteen skills within the DeepMind cluster had no origin: true flag and no independent evidence strong enough to hold above 3★. Their prior 4★–5★ ranks came from cluster proximity, not standalone Trust Magnitude.

Policy basis: Sharing a bucket with a primary lead does not inherit its origin weight — each node's TM is computed from its own evidence rows. Where the only evidence above B-grade was cluster affiliation, demotion to 3★ was required.

Resulting state: 3★ signals real, documented capability with public evidence. It simply requires canonical origin or high independent TM to cross into 4★.

Primary-Bucket Promotions — 3 Skills to 4★ (9 Leads Assigned Origin)

Nine nodes were confirmed as canonical primary leads for their DeepMind sub-clusters and received origin: true. Three of these nodes (clinicaltrialsdatabase, pdbdatabase, and uniprotdatabase) were promoted from 3★ to 4★ as a result. The remaining six nodes were already at 4★ and received the canonical origin flag to lock in their status.

Following a methodology review (see docs/codex/trust-methodology.html), the gate was corrected: a confirmed primary lead with origin: true and at least one A-grade evidence row qualifies for 4★ without requiring a second independent repository row.

Dead-Link Demotions — scienceskillscommon & pexp13/sentiment-analysis to 1★

[!WARNING] Dead-Link Demotions: Both scienceskillscommon and pexp13/sentiment-analysis were demoted from 4★ → 1★ due to permanent HTTP 404 errors on their primary repository URLs (META.md Star Bar floor).

Registry policy requires that any node whose primary evidence link or repository URL is permanently unreachable drop to 1★, regardless of prior TM score:

• scienceskillscommon: Primary evidence URL returned HTTP 404 across three Firecrawl validation checks spanning 48 hours.
• pexp13/sentiment-analysis: Repository URL (links.github) returned HTTP 404 and carries an empty frontmatter evidence list (evidence: []). Calibrated to 1★ via PR #1422 (gaia dev calibrate pexp13/sentiment-analysis 1★).

If upstream repositories are restored or mirrored, re-promotion requests can be submitted through standard intake.

Trust Ledger & Timeline Continuity Safeguards

PR #1410 — Timeline Backfill for pexp13/sentiment-analysis

[!NOTE] This PR changed no rank. The sole effect was timeline continuity — backfilling a missing 1★ → 4★ rank_up event so the contributor profile chart displays the correct history.

pexp13/sentiment-analysis had a timeline gap: it was calibrated from 1★ to 4★ in a prior batch merge, but the corresponding rankup event was never written to the user-tree. PR #1410 ran scripts/tracetimeline.py --apply to synthesize the missing event with correct timestamps. The CI gate validate_timelines.py passed cleanly post-application.

PR #1406 — 1-Star Skill Redaction from Trust Ledger Export

[!NOTE] The trust-ledger filter is non-destructive. Nodes with stars < 2 are excluded from the exported leaderboard payload but remain fully intact in the registry. No contributor's aggregate TM score was altered.

PR #1406 introduced a filter in scripts/generateLeaderboardData.py (via gaiacli.redaction.isredacted()) that excludes 0★ and 1★ skill nodes from the exported trust ledger payload. This prevents low-signal or demoted nodes with stale/dead evidence links from diluting contributor reputation metrics downstream.

PR #1402 — Sole-Bucket Disler Origin Calibrations

A sole-bucket node has no other skill sharing its sub-cluster — registry policy automatically makes it the primary lead with origin: true. Four disler nodes (agent-fusion, auto-review, opinion, plan-synthesis) qualified but lacked the flag because they were ingested before the sole-bucket rule was established. PR #1402 assigned origin: true to all four nodes.

Intake & Promotion Summary

mvanhorn/last30days — Named Implementation Under autonomous-web-research

PR #1391 promotes mvanhorn/last30days to 4★ as a named implementation of the generic autonomous-web-research fusion node. It passed evidence review with:

• origin: true canonical status and high Trust Magnitude.
• Full implementation of the autonomous-web-research prerequisite surface (ghostwrite, knowledge-harvest, research, web-scrape, web-search).

Because autonomous-web-research does not define suiteComponents, mvanhorn/last30days is programmatically classified on the Unique branch (◉) rather than a Suite branch (registry/named/mvanhorn/last30days.md).

11 New ev-seed Named Skills

The ev-seed pipeline delivered 11 new named-skill candidates with pre-verified evidence. All 11 passed Phase 3 adversarial auditing and Phase 4 link validation without defects. Breakdown post-calibration:

• 4★ (1 node): nextlevelbuilder/ux-audit
• 3★ (3 nodes): oso95/scroll-world, panniantong/agent-reach, vercel-labs/react-performance-optimization
• 2★ (5 nodes): disler/agent-fusion, disler/auto-review, disler/opinion, disler/plan-synthesis, anthropics/static-artwork-design
• 1★ (2 nodes): gaia-research/skill-cost, ayghri/format-output

None triggered a fusion or suite promotion at this time; each stands as an independent named node.

Methodology & Policy References

Report: Gaia Research · Registry Audit · Release v7.3.3 · 2026-08-01 Validation: gaia dev validate ✅ — all changes landed via reviewed PR, no direct commits to main