Gaia Skill Tree Meta Audit Series

Between early and mid-September 2026, the Gaia Skill Tree shipped releases v8.7 through v8.12 with five user-visible changes: two new named skills were admitted (including a 4★ promotion), four skills incorrectly caught in an unrelated purge sweep were disentangled and restored under new anti-purge safeguards, six flagship skill suites were synchronized to their upstream authors' latest releases, a new Frozen Rank policy now protects author-pinned skills from automated sweep demotions, and the Four-Tier Audit Taxonomy was ratified as the permanent framework governing how skills are evaluated, demoted, or removed. This report covers every rank shift, every new rule, every restoration, and every upstream version change that landed.

What Users Should Know

Five things changed since the September 9 Rank Calibration:

1. Two new skills were admitted. trailhq/graft entered at 4★ and yylo-dev/ledger-tasks-yylo entered at 2★. 2. Four wrongly purged skills were restored. rico-favor/implement-with-discernment and Matt Pocock's edit-article, obsidian-vault, and ubiquitous-language were disentangled from an unrelated imposter sweep and reinstated at their earned ranks. 3. Frozen ranks are now protected. Skills whose upstream authors have deprecated or archived them retain their earned rank permanently and cannot be demoted by automated recalibration sweeps. 4. Upstream skill suites are current again. Six suites — spanning 14 named skills — were synchronized to their authors' latest published releases. 5. The Four-Tier Audit Taxonomy is now permanent policy. Every future skill evaluation, demotion, or removal follows a clear, published four-tier framework with mandatory safeguards against collateral damage.

New Skills Admitted

trailhq/graft — 4★ Extra Skill

trailhq/graft was admitted at 4★ under the new generic capability codebase-graph-retrieval. Graft builds a deterministic context graph of symbol definitions, call sites, and architectural dependencies across a codebase, giving agents immediate structural awareness of any repository they enter.

This is the first skill catalogued under the codebase-graph-retrieval capability — a recognition that agents need structured graph-based code understanding, not just keyword search or file-by-file reading.

yylo-dev/ledger-tasks-yylo — 2★ Named

yylo-dev/ledger-tasks-yylo was admitted at 2★ under the existing project-management capability. It provides structured task accounting and ledger-based project tracking for agent workflows.

Skills Restored Under the Anti-Purge Safeguards

PR #1721's registry-wide "imposter purge" removed 12 entities in one sweep. Contributor mbtiongson1 flagged in Issue #1809 that the sweep had also caught legitimate skills — a self-hosted co-founder prototype and three skills that failed a Star Bar check because their upstream repository had gone dark, not because their evidence was bad. Both were disentangled from the bad-faith squatters they'd been bundled with and restored at their earned ranks.

rico-favor/implement-with-discernment — restored to 1★ Awakened

Authored by co-founder Rico Tiongson (rico-favor), this skill was swept up in the imposter purge as part of a blanket rico-favor/* removal, despite being a legitimate self-hosted engineering practice rather than bad-faith aggregation. Per the Four-Tier Audit Taxonomy this is a Tier 3 (Early-Stage / Under-Evidenced) case, not a Tier 1 imposter. It was restored in PR #1812 with a substantive, non-placeholder skill guide and co-founder self-attestation evidence (Trust Magnitude 5.7), and re-tagged installable: false per CONTRIBUTING.md §12 rather than expunged.

Matt Pocock's frozen trio — restored to 3★ Evolved

The same PR #1721 sweep also enforced a Star Bar check that demoted three of Matt Pocock's skills to 1★ after their upstream repository went dark — evidence that had been verified was discarded because the live blob could no longer be re-checked. Under the new Frozen Rank policy (below), all three were restored to their verified 3★ Evolved rank in PR #1820:

These skills earned their 3★ rank through verified evidence before their upstream repository was archived. The frozen rank policy ensures that historical achievement is preserved honestly.

Frozen Ranks: Protecting Author-Pinned Skills

A problem surfaced during the September 9 recalibration: skills intentionally frozen at a specific rank (because their upstream author deprecated or archived the repository) were being caught in automated recalibration sweeps and incorrectly demoted.

The New Rule

Skills with a documented upstream_deprecated lifecycle event now retain their earned rank permanently. Automated sweeps skip them entirely. If a frozen skill's rank is accidentally altered, the system detects the discrepancy and proposes a restoration command.

Upstream Suites Synchronized

When upstream authors release new versions of their skill repositories, the registry must reflect those changes. The following six suites were updated to match their authors' latest releases:

All version pins and installation instructions across these 14 named skills now reference the upstream author's current release. No star levels changed as a result of synchronization — version tracking is independent of trust evaluation.

The 48-Skill Rank Calibration (September 9)

Immediately before this cycle, a registry-wide Trust Magnitude recalibration adjusted 48 named skills to match their computed evidence grades. The full table was published in the September 9 report; the headline shifts were:

These demotions followed the evidence — skills with thin or stale evidence lakes dropped to their computed grade floor, while skills with verified independent witnesses climbed.

The Four-Tier Audit Taxonomy

Every skill in the registry will eventually be evaluated. Until now, the rules governing evaluation, demotion, and removal were scattered across internal playbooks. Starting with this release, a single, permanent framework governs all audit outcomes.

The Four Tiers

Tier 1 — Imposter / Squatter → Expungement. Skills that exist through bad-faith aggregation (hijacked star counts, falsified attribution, no real skill artifact) are permanently removed from the registry. This is the harshest outcome and is reserved for deliberate fraud, not honest packaging mistakes.

Tier 2 — Packaging Gap → Retained at ≤2★. The author and concept are legitimate, but the repository lacks an installable SKILL.md or runnable entry point. The skill stays in the registry for discovery purposes but is honestly flagged as non-installable. It is never purged — the path forward is for the author to add proper packaging.

Tier 3 — Early-Stage / Under-Evidenced → 1★ Awakened. The author is verified and the capability is real, but the evidence lake is thin — perhaps only a self-attestation or a preliminary repository. The skill holds the floor rank until real evidence (benchmarks, peer reviews, independent adoption) arrives.

Tier 4 — Product-Coupled / Non-Generalized → Attribution Required. The skill documents a specific third-party tool (like PyMOL or RDKit) without crediting the tool's actual maker. The fix is not removal but proper attribution: credit the upstream maintainer and map the skill to a generic capability rather than a vendor name.

The Mandatory Safeguard

All four tiers carry a mandatory disentanglement check: before any demotion or removal, a PR performing an evidence scrub or purge must itemize and categorize each target into one of the four tiers rather than blanket-labeling the batch, and outreach plus documented review are required before permanently deleting a skill authored by a known contributor or team member. This safeguard was formalized in PR #1814 and immediately applied retroactively to restore the skills described above.

Product-Attribution Detector

A new automated detector now flags skills that document a commercial or open-source library without acknowledging its actual maintainer. Six k-dense-ai library wrapper stubs (PyMC, PyTorch Lightning, Scanpy, scvi-tools, Stable-Baselines3, PyTorch Geometric) were identified as having zero attribution to their upstream authors. These remain in the registry at 1★ with needs-info status — the fix is for the contributor to add proper attribution and documentation, not for the registry to silently remove them.

New Evidence Types

Two new evidence types are now recognized across the registry:

• npm-downloads: Measures verified download volume from the npm package registry, providing an adoption signal independent of GitHub stars.
• engagement: Captures community interaction metrics (discussions, issue activity, conference mentions) that demonstrate active use beyond passive star-clicking.

These types expand the range of evidence that can contribute to a skill's Trust Magnitude score, giving skills with strong real-world adoption but modest GitHub visibility a fair path to higher ranks.

Version Timeline

As of v8.12, the registry contains 339 named skills across 294 catalogued capabilities, with all ranks aligned to their computed Trust Magnitude grades.