waggle-os/docs/briefs/2026-04-30-cc-sesija-C-gaia2-setup-evidence.md

CC Sesija C — §0 Preflight Evidence (gaia2-setup-evidence.md)

Evidence ID: cc-sesija-c-gaia2-setup-evidence-v1 Date: 2026-04-30 Author: CC Sesija C Brief: briefs/2026-04-30-cc-sesija-C-gaia2-setup-dry-verification.md (LOCKED) Status: §0 PARTIAL — HALT-AND-PM REQUIRED before §1 begins Verdict summary: §0.1 PASS · §0.2 PARTIAL FAIL (brief contradiction) · §0.3 PAPER ESTIMATE at/above cap, actual probe deferred Working tree state at evidence time: waggle-os primary worktree on phase-5-deployment-v2 with 1,282 unstaged changes (-52,062 LOC) matching main's shrunk state — independent halt-and-PM required.

---

§0.1 — Gaia2 ARE platform availability — PASS

Sub-item	Brief requirement	Verified evidence	Verdict
1	ARE platform repo accessible	https://github.com/facebookresearch/meta-agents-research-environments — Public, MIT License, 26 commits on main branch. Install methods: uv / pip install meta-agents-research-environments / Docker. CLI are-benchmark gaia2-run --hf meta-agents-research-environments/gaia2.	PASS
2	Gaia2 paper anchor arxiv 2602.11964	Cited in brief §6 + benchmark portfolio brief 2026-04-29 §2.1 (Froger et al., 12 Feb 2026, Meta SuperIntelligence Labs). Anchor consistent across authority chain.	PASS
3	Gaia2 dataset license verified for research use	HuggingFace dataset card meta-agents-research-environments/gaia2: Creative Commons Attribution 4.0 International (CC-BY-4.0), SPDX cc-by-4.0. Direct quote: "The Data is released CC-by 4.0 and is intended for benchmarking purposes only." No commercial-use prohibition. Llama-attribution clause applies only if the data is used to train/finetune distributed models — not applicable to dry run (we benchmark prompt shapes, do not retrain models). Synthetic-data subcomponents are Llama-3.3 + Llama-4 Maverick outputs subject to those model licenses.	PASS
4	Gaia2 Search split task count	HuggingFace card lists 800 total scenarios across 6 configurations: execution 200 · search 200 · adaptability 200 · time 200 · ambiguity 200 · mini 200. Brief targets dry run N=10–20 (subset of search 200) with full N=200 deferred to Phase 3 sprint Week 6. Aligned.	PASS

§0.1 verdict: PASS. Platform installable, paper authoritative, license permits dry run + Phase 3 sprint research use, search split has sufficient task count.

---

§0.2 — GEPA-evolved variants accessible — PARTIAL FAIL (brief contradiction)

Sub-item	Brief requirement	Verified evidence	Verdict
1	claude::gen1-v1 + qwen-thinking::gen1-v1 shape definitions reachable	Files exist in repo at packages/agent/src/prompt-shapes/gepa-evolved/claude-gen1-v1.ts and qwen-thinking-gen1-v1.ts. Headers preserved (Faza 1 mutation oracle provenance, Phase 4.5 retrieval-engagement evidence, Amendment 2 §3 retrieval-engagement bonus, Cell-semantic baseline anchors via mutation-validator.ts SHA pins). BUT: git ls-tree -r main returns ZERO matches for prompt-shapes/ AND ZERO matches for retrieval-agent-loop.ts — files DO NOT exist on main. Files exist on feature/c3-v3-wrapper, gepa-faza-1, phase-5-deployment-v2.	FAIL on main (brief §1 mandate) / PASS on feature/c3-v3-wrapper
2	registerShape canonical API working (Amendment 8 native)	On phase-5-deployment-v2, packages/agent/src/prompt-shapes/index.ts exports registerShape from ./selector.js: export { selectShape, listShapes, getShapeMetadata, REGISTRY, registerShape, _resetConfigCache, type SelectShapeOptions } from './selector.js';. Comment block confirms "Public API (Phase 1.2 of agent-fix sprint)". Amendment 8 native confirmed via Faza 1 closure decision §G.1 SHA chain. BUT: API is reachable only on branches that contain prompt-shapes/. Same branch-base contradiction as sub-item 1.	FAIL on main / PASS on Faza 1 substrate branches
3	runRetrievalAgentLoop accessible from Gaia2 harness adapter	On phase-5-deployment-v2, packages/agent/src/retrieval-agent-loop.ts is the production entry point: "Two entry points: runSoloAgent — single-shot Cell A/C pattern; runRetrievalAgentLoop — multi-step Cell B/D pattern." Phase 5 canary wiring already integrated (routeRequestToVariant, WAGGLE_PHASE5_CANARY_PCT). 38.3 KB integration surface, consumes selectShape + MULTI_STEP_ACTION_CONTRACT from prompt-shapes/index.js. Same branch-availability scope as sub-item 1+2.	FAIL on main / PASS on Faza 1 substrate branches

§0.2 verdict: PARTIAL FAIL. Files + APIs verified existent and correctly structured, but brief §1 explicitly mandates branching from main, where these artifacts do NOT exist. The brief is internally contradictory:

• Brief §1 line 8: "Branch: Kreirati feature/gaia2-are-setup iz main (ne zavisi od Sesija A ili B grana)"
• Brief §0.2 sub-item 1: "shape definitions reachable u packages/agent/src/prompt-shapes/ ili monorepo migrated location"

The shapes are on Faza 1 substrate branches (per Faza 1 closure decision §G manifest v7 substrate anchor c9bda3d6dd4c0a4f715e09f3757a96d01ff01cd7 on feature/c3-v3-wrapper). They are NOT on main. The "monorepo migrated location" escape clause does not yet apply — Track C (CC Sesija B: hive-mind monorepo migration) has not yet executed, so no monorepo migration has happened.

Halt-and-PM required. PM must ratify branch-base before §1 can begin.

---

§0.3 — Cost projection probe — PAPER ESTIMATE (deferred actual probe)

Constraint: Brief §0.3 specifies "3-request dry run probe sa Gaia2 sample tasks". Actual probe requires ARE platform installed + Gaia2 dataset downloaded — both are §2.1 Day-1-morning tasks (Task C1+C2). §0 cannot literally execute a 3-request probe before §2.1 install.

Methodology adopted: Paper estimate using two anchors —

• (a) Faza 1 cost evidence (decisions/2026-04-29-gepa-faza1-results.md §F): 135 evaluative records / $43.49 = $0.32/eval avg (mixed: NULL baseline, Gen 1 mutation, Checkpoint C held-out).
• (b) Gaia2 task character: per-scenario data field 2.44–2.67M characters (HuggingFace card), 12 apps + 101 tools system overhead, multi-step async with dynamic events. Materially larger than Faza 1 LoCoMo-style analytical scenarios.

Per-task cost estimate band

Configuration	Per-task cost estimate	Source / reasoning
Qwen-thinking baseline (filtered tool surface, single-step retrieval)	$0.08–0.15	Faza 1 Qwen Checkpoint C $1.93/15 = $0.13/eval; Gaia2 multi-step adds ~30% premium offset by tool filtering
Claude baseline (full system prompt, multi-step retrieval)	$0.25–0.50	Faza 1 Gen 1 $15.02/120 = $0.125/eval; Gaia2 adds 2-4× for 12-app system overhead + multi-step async
GEPA-evolved variants (claude::gen1-v1 + qwen-thinking::gen1-v1, retrieval-engagement positive signal)	+15–30% premium over baseline	Faza 1 evidence shows GEPA variants emit more retrievals (Phase 4.5 mechanism: qwen-thinking 2.231 mean retrieval vs same-shape baseline 1.625, +37% relative)

Dry run total projection (4 shapes × 10 tasks = 40 invocations)

Scenario	Per-task avg	Total projection
Optimistic (all Qwen baseline rates)	$0.13	$5.20
Mid (mixed Claude + Qwen, baselines + GEPA average)	$0.25	$10.00
Pessimistic (Claude-heavy + GEPA premium + multi-step retrieval failures)	$0.45	$18.00

Brief cost cap: $10 hard / $8 halt / $5–8 expected. Projection: Mid-estimate $10 lands AT hard cap. Pessimistic $18 lands ABOVE hard cap.

Recommended scope adjustments (require PM ratification)

Option α — Reduce dry run scope to N=5 per shape (20 invocations total). Mid-estimate $5, pessimistic $9. Stays within $8 halt. Sacrifices statistical power but aligns with §0.3 cost discipline.

Option β — Keep N=10 per shape (40 invocations) with strict per-shape halt monitoring. After each 10-invocation shape batch, compute running cost; halt-and-PM if total > $8 before all 4 batches complete.

Option γ — Defer §0.3 to first-batch-as-probe. Run the 4-shape × 3-task probe (12 invocations) first, measure actual cost, project remaining 28 invocations from probe data, halt-and-PM if projection > $8.

Option δ — Authorize cost cap raise to $15 hard / $12 halt (still within Phase 5 cost amendment $75 hard / $60 halt envelope).

§0.3 verdict: PAPER ESTIMATE PARTIAL. Actual 3-request probe deferred to first execution batch in §2.1+§2.3 sequencing. PM must ratify scope adjustment (α/β/γ/δ) before §1 begins.

---

§0.4 — Independent halt-and-PM: working tree state inconsistency

Not part of brief §0 gates — surfaced because it blocks §1 branch creation.

git status --porcelain | wc -l on primary worktree D:/Projects/waggle-os reports 1,282 unstaged changes.

git diff --stat phase-5-deployment-v2 main reports 261 files changed, 89 insertions, 52,062 deletions. Direction phase-5-deployment-v2 → main = main is the SHRUNK version. Working-tree pending deletions match main's state, suggesting prior git checkout main -- . or equivalent operation that left HEAD on phase-5-deployment-v2 while replacing working tree files with main's content.

Examples of pending deletions: apps/web/components.json, apps/web/index.html, apps/web/package.json, apps/web/playwright.config.ts, all of apps/web/src/components/os/apps/, apps/web/src/assets/personas/, plus packages/core/tests/mind/scoring.test.ts, scripts/run-pilot-2026-04-26.ts (951 LOC), scripts/run-mini-locomo.ts (619 LOC), scripts/parity-check.sh, vitest.setup.ts (8 LOC), and 256 more files.

Filesystem confirms: apps/web/ exists as empty/stub directory; external/ does not exist (would be Task C1 install destination); packages/agent/src/prompt-shapes/ returns ZERO files on disk via Glob (consistent with main's tree which lacks the directory).

Faza 1 worktree referenced in Faza 1 closure decision (D:/Projects/waggle-os-faza1-wt) does NOT exist on filesystem. Per git worktree list: only one worktree, the primary at D:/Projects/waggle-os. The Faza 1 worktree was cleaned up between 2026-04-29 closure and 2026-04-30 Sesija C kickoff.

Resolution options for working-tree state:

• Option A — Reset working tree to phase-5-deployment-v2 HEAD (git checkout phase-5-deployment-v2 -- . then git clean -fd). Restores the Phase 5 canary state. Loses the deletions (which appear unintentional).
• Option B — Stash deletions and discard (git stash push --keep-index --include-untracked then git stash drop). Same effect as A.
• Option C — Investigate first. What happened between 2026-04-29 S2 handoff (clean tree) and 2026-04-30 evidence time? git reflog may explain.

§0.4 verdict: HALT. PM should pick A/B/C before §1 branch creation.

---

§1 prerequisites — what PM must ratify

Three independent ratifications required:

Ratification ask 1 — branch base (§0.2 contradiction resolution)

The brief mandates branching from main, but main lacks the GEPA-evolved shape files + retrieval-agent-loop.ts. Pick one:

Option	Branch base	Trade-off
1A	feature/c3-v3-wrapper (Faza 1 substrate anchor c9bda3d6)	Cleanest — substrate-pinned to Faza 1 closure SHA. Diverges from brief §1 literal text.
1B	phase-5-deployment-v2 (Phase 5 canary tip a8283d6)	Inherits Phase 5 canary infrastructure (manifest, monitoring emitters, feature flags). Slightly larger surface than needed.
1C	gepa-faza-1	Same shapes as 1A. Different branch label. Less canonical.
1D	First merge feature/c3-v3-wrapper → main, then branch from main	Honors brief §1 literal text. Adds prerequisite work + merge resolution risk. Couples Sesija C to merge work.
1E	First cherry-pick prompt-shapes/gepa-evolved/ + retrieval-agent-loop.ts into a fresh main-based branch	Surgical. Preserves brief §1 literal text. Risks substrate-isolation discipline if cherry-pick alters anchor SHAs.

CC recommendation: 1A (feature/c3-v3-wrapper) — Faza 1 substrate anchor preserved verbatim, no merge or cherry-pick risk, no working-tree resolution coupling. Brief §1 line 8 should be amended to substitute feature/c3-v3-wrapper for main.

Ratification ask 2 — §0.3 cost projection scope (Option α/β/γ/δ)

Option	Action	Risk
α	Reduce dry run to N=5/shape (20 invocations)	Lower stat power
β	Keep N=10/shape with strict per-shape halt	Modest cost-overshoot risk
γ	First batch as probe (12 invocations), project remainder	Probe-validated; aligns with brief §0.3 literal text
δ	Raise cost cap to $15 hard / $12 halt	Within Phase 5 amendment envelope ($75/$60)

CC recommendation: γ (first batch as probe). Honors brief §0.3 literal text most closely; produces a real probe-validated projection at minimal initial spend (~$1–4); halts before exceeding $8 if pessimistic estimate is realized.

Ratification ask 3 — §0.4 working-tree resolution (Option A/B/C)

CC recommendation: C (investigate first). git reflog is cheap and may reveal whether the deletions are intentional (e.g., a preparatory cleanup that should be committed) or accidental (e.g., aborted checkout). After reflog inspection, fall back to A or commit-and-move-on as appropriate.

---

Audit anchors

• Brief: briefs/2026-04-30-cc-sesija-C-gaia2-setup-dry-verification.md
• Authority chain:
  • decisions/2026-04-30-pre-launch-sprint-consolidation-LOCKED.md
  • briefs/2026-04-29-benchmark-portfolio-refresh-2026-venues.md (§2.1 + §2.3)
  • decisions/2026-04-29-gepa-faza1-results.md
• ARE platform: https://github.com/facebookresearch/meta-agents-research-environments
• Gaia2 dataset: HuggingFace meta-agents-research-environments/gaia2 (CC-BY-4.0)
• Faza 1 substrate anchor: c9bda3d6dd4c0a4f715e09f3757a96d01ff01cd7 on feature/c3-v3-wrapper
• Phase 5 substrate tip: a8283d6 on phase-5-deployment-v2 (tag v0.1.0-phase-5-day-0)
• This evidence: briefs/2026-04-30-cc-sesija-C-gaia2-setup-evidence.md

---

End of §0 evidence. CC HALTED at §0 gate. §1 begins only after PM ratifies asks 1+2+3.