outputs/readme_v4_diff.md

README v3 → v4 — Change Log

Date: 2026-05-26 v3: 653 lines · 6284 words v4: 699 lines · 7162 words Delta: +46 lines · +878 words (well under the +150-line / 800-line budget)

The v4 pass incorporates the overnight 14-sub-agent stress loop findings (outputs/MASTER_REPORT.md, outputs/BUG_BACKLOG.md) while preserving v3's Medium-quality voice.

---

What changed

1. Badges (top of README)

• Tests badge: 149 Py + 30 Sol + 15 FE → 219 Py + 36 Jest + 30 Foundry (real counts post-overnight)
• Slither badge: clarifies "0 High / 0 Medium" applies to first-party code (OZ Math.sol library noise excluded)
• New Smoke badge: Smoke 10/12 GREEN linking to MASTER_REPORT.md
• Tests + Slither badges now point at MASTER_REPORT.md instead of final_audit_summary.md (more current)

2. TL;DR paragraph 3

Added: builder code 0xa934…beb1 registered, Alchemy Polygon RPC bound, overnight stress loop reference, real-stack coverage figure (~85%), smoke 10/12 GREEN.

3. Real vs Mock: Honest Accounting

• Phase 1 chain glue + dispatch rows: PHASE 1 (landing) → REAL (smoke verified)
• New row: Alchemy Polygon RPC binding marked REAL with median latency
• Coverage estimate: 25–30% → ~85% with provenance (Phase 1 + overnight verification)
• WARNING callout rewritten: the gap now is BLEU/COMET reference-lookup wiring (HIGH-1 in BUG_BACKLOG), not chain glue. MQM (the most informative of the three) is real.

4. The Numbers table

Replaced "target 60–75 s" with measured values from perf_benchmark.md:

Row	v3	v4
Lifecycle wall clock	60–75 s target	p50 65.87 s measured, p95 ≥180 s on stalls
API p95	—	8.7 – 29.3 ms (/events, /leaderboard, /events/{id})
Backend cold start	—	1.65 s + Next.js FCP 90–760 ms
FAISS lookup median	—	16.07 ms vs 100 ms budget
Arc RPC eth_blockNumber	—	p50 590.6 ms · p95 828.3 ms
Test suites total	—	285 pass (219 + 36 + 30)
Slither verdict	post-hardening	post-hardening first-party clarified

5. NEW SECTION — "Stress Tested Overnight (2026-05-26, 04:30–08:00 SGT)"

Position: before "Audit + Hardening Pass" (chronologically the newer event).

Covers:

• 14 sub-agents launched in 3 waves over 3.5 hours
• 600+ check items across 9 domains
• 47 bugs catalogued, 27 auto-fixed
• Before/after table on 9 surfaces (smoke 4→10/12, mobile 47%→81%, etc.)
• 121 screenshots produced
• Demo readiness verdict: GREEN mechanism / YELLOW market with explanation
• Pointers to MASTER_REPORT.md and BUG_BACKLOG.md

Voice: "we put it through a stress test loop" — not boastful, just earned.

6. Audit + Hardening Pass

One-line transition update: "Before the overnight stress loop, an earlier 8-audit parallel pass ran…" — chronology now reads cleanly.

7. Roadmap

Added a new "Production hardening" phase (1–4 weeks post-ship) capturing the three concrete Agent H production recommendations:

• BackgroundTasks migration for /trigger/event (BLOCKER)
• LLM timeout + circuit breaker around the 4-provider fan-out
• gunicorn --workers 4 + reverse proxy
• BLEU/COMET reference-lookup wiring (HIGH-1)
• Firefox SSE CORS fix (HIGH-2)

Plus a follow-up paragraph explaining what each is, what surfaced it (Agent H perf benchmark), and that they are roughly day-of-work fixes, not architecture changes.

8. Arc capabilities table

Added Alchemy Polygon RPC row with app id ngx37mo60qae6ror and median latency.

9. "The Numbers" intro polish

Opener now lists the concrete corpus + event + bid + submission + test counts that back the "real data, not just simulated" claim.

---

What stayed the same (deliberate)

• 6 Mermaid diagrams (none added, none removed)
• All 24 unique §5.X cross-reference anchors still resolve
• Section order unchanged except the new "Stress Tested Overnight" insertion
• Voice / tone preserved from v3 (Medium-quality narrative)
• §5.30 honest-scope discipline maintained: no claim about proof-of-market
• Mechanism design defaults table (locked parameters) untouched
• License + Closing Thesis untouched

---

Anchor-resolve verification

Unique §5.X anchors in v4 (24 distinct): 50, 502, 503, 510, 515, 518, 521, 522, 527, 528, 530, 540, 5402, 541, 542, 543, 544, 546, 547, 548, 55, 551, 56, 57

All identical to v3 anchor set. No anchors added or dropped.