Intake Batch: 11 Named Skills Onboarded — ev-pipeline v2 Live

Gaia Research

July 30, 2026

Abstract

Eleven named skills across seven contributors cleared the evidence pipeline and landed in the registry. ev-pipeline v2 ships deterministic append scripts and a three-hop intake label ladder.

The July 2026 evidence-seed sprint closed eight intake issues, promoted eleven named skills from seven contributors, and shipped ev-pipeline v2 — replacing freehand agent writes with deterministic, append-only, URL-deduplicated scripts backed by a typed 13-field JSON schema.

New Named Skills

Eleven named skills landed across seven contributors, spanning UI/UX auditing, agent workflow orchestration, scroll-based interaction, React performance, and Anthropic tooling.

Contributor	Skill	Level	TM	Grade
nextlevelbuilder	ui-ux-pro-max	4★	194.3	A
pbakaus	impeccable	4★	127.4	A
disler	agent-fusion	2★	127.5	A
oso95	scroll-world	3★	61.0	B
panniantong	agent-reach	3★	53.9	B
vercel-labs	react-best-practices	3★	50.5	B
anthropics	canvas-design	2★	43.5	C
garrytan	design-review	2★	36.0	C
disler	auto-review / opinion / plan-synthesis	2★ each	36.3	C

Two bucket corrections shipped alongside: pbakaus/impeccable and garrytan/design-review were reclassified from ux-audit to the new ui-audit generic, freeing ux-audit for nextlevelbuilder/ui-ux-pro-max as Origin at 4★.

ev-pipeline v2 — What Changed

The prior evidence pipeline relied on agents to hand-write collector markdown blocks — correct in specification, fragile in practice. Three failure modes recurred across the first real seed run:

Field drift. Agents recalled delimiter rules inconsistently (Citations: vs citations, Views: vs views). The compiler accepted whatever it found, silently producing null TM signal for rows where numeric fields existed but were misnamed.

No URL deduplication. A retry after partial failure re-appended the same URLs. The data lake accumulated duplicate rows frozen by  with no repair path.

Freehand HTML stats. The verification dashboard was updated by agents mimicking prior output. Two runs on the same date produced double-counted cumulative stats and format drift that broke the next agent's pattern match.

Fixes shipped

scripts/ev_append.py — deterministic, append-only, URL-dedup script. Agents produce a 13-field JSON row; the script owns block delimiters, section numbering, date-stamping, and cross-run URL deduplication. A retry is a no-op.

scripts/evstatspatch.py — idempotent HTML stats patcher with five CLI flags and strict date-key deduplication. Agents never touch the HTML directly.

The ev-pipeline Post-Run Outputs checklist gained a fourth step: apply intake:evidence-ready on each intake issue to advance the three-hop label ladder (evidence-review → evidence-ready → evidence-approved).

Intake Label Ladder

The CI workflow (intake-approval.yml) requires both intake:evidence-review AND intake:evidence-ready before the evidence-approved gate opens a draft review/meta/ promotion PR. All eight intake issues now carry the correct labels and are closed.

References

[1] PR #1383 — ev-pipeline scripts + 11-skill evidence seed. gaia-research/gaia-skill-tree.

[2] PR #1387 — review/meta intake ev-seed ingest: named node creation, evidence ingestion, calibration. gaia-research/gaia-skill-tree.

[3] PR #1388 — CLI follow-up: UTF-8 safe validate output, trust_appraise named lookup fix, suiteComponents preservation. gaia-research/gaia-skill-tree.

[4] Postmortem: founder/handovers/2026-07-30-POSTMORTEM-ev-pipeline-intake-seed.md. Seven findings against ev-pipeline v1 and curate-v2 curation worker determinism.