Skip to content

Build log

Reverse-chronological entries. Each entry links the commit + a short why.

2026-04-20 — M4 scraper operational (6 RSS feeds live)

Korean news scraper running end-to-end on real feeds. M4 (90-day blueprint's "scraper + oracle ingest") is now complete.

Shipped:

  • apps/kfish-core/src/kfish_core/pipeline/ingest_articles.py — orchestrator: 6 sources → SimHash-48h-dedup → Kiwi tokenize → UPSERT → lazy FTS rebuild. Every step independently failure-tolerant; missing credentials degrade gracefully.
  • kfish-news CLI with --dry-run, --stats, per-tier --sources, --translate-bulk (Papago), --translate-high-signal N (Claude).
  • Nightly pipeline now runs news ingestion alongside markets + oracle, with KFISH_NAVER_QUERIES env override.
  • 7 unit tests + 7 live-network tests (all RSS feeds probed on 2026-04-20).
  • Bug found + fixed: SimHash64 is unsigned; DuckDB BIGINT is signed int64; added _to_signed_64 reinterpret at the DB boundary.
  • RSS URL correction: Hankyoreh redirects /rss/ → /rss (saves one round-trip per run).
  • 6 canonical feeds: yonhap, yonhap_econ, hankyoreh, blockmedia, dailynk, donga.
  • Full evidence on the news-scraper page.

First live run: 60 articles fetched, 4 near-duplicates caught, 56 inserted with Kiwi tokens, FTS built — 13.5 s total.

Naver tier remains pending your Developer-app registration (operator task, free tier, instructions in runbooks/scraper-ops.md in the private repo).

2026-04-19 — Live-trading hardening (18/18 review items fixed)

After the live-API validation, spawned independent code + security review agents plus my own coverage / pyright audit. They flagged 18 issues across CRITICAL / HIGH / MEDIUM / LOW. Every single one is fixed — see the full review.

Headlines:

  • /close, /positions, /balance now hit the real HL info endpoint. No more hardcoded 100.0 reference price, no more placeholder strings, no more "order accepted" on an order that was never sent.
  • HYPEKR_APP_SIGNING_SECRET is now a required ≥32-char value — the silent "dev-only-insecure-secret" fallback that would have let attackers forge wallet approvals from public source is gone. 9 regression tests enforce this at build_app() time.
  • HKDF is now real RFC-5869 HKDF with a static domain-separation salt + per- user info binding, not a single HMAC-SHA256 step.
  • Mondrian conformal stratifies by true class. Per-class coverage test added — passes at ≥ 0.86 for both classes on out-of-sample data.
  • Delphi convergence checks absolute spread, not delta. No more early-exit at spread=0.29 when target is 0.02.
  • WalletStore DuckDB connection now guarded by threading.RLock so concurrent aiogram handlers serialize cleanly.
  • pypa/gh-action-pypi-publish + all publishing-path actions SHA-pinned (6733eb7d… for v1.14.0).
  • Raw Ethereum addresses never appear in logs — only 12-char SHA-256 fingerprints.
  • Pyright: 14 errors → 0 errors. Ruff: clean. Tests: 140/140 pass.

2026-04-19 — Live-API validation + Brier-parity + quant-notes

  • Probed live Polymarket Gamma + UMA Goldsky OOv2 + Managed OOv2 — all three reachable; Goldsky project id rotated (clcxirpj… → clus2fnd…), updated in config.py and .env.example.
  • Rewrote UMAGoldskyClient to use the current OptimisticPriceRequest entity with per-stage timestamps. Added bond_usdc property that handles the 6-decimal USDC fixed-point.
  • Added tests/test_live_apis.py under -m network — runs on demand against the real endpoints. All 4 live tests green.
  • Wrote scripts/brier_parity.py — validates the new kfish-core calibration math against legacy retrodiction JSON. Brier parity exact to 4 decimal places on all three archived legacy runs (N=100, 200, 30). First of the 7-night ADR-0002 cutover gate is green.
  • Wrote scripts/hip4_watch.py — polls HL mainnet + testnet /info meta for HIP-4 symbols; exit code 2 fires when mainnet goes live.
  • Wrote scripts/bootstrap_live.py — one-command launcher for repos + subtree split + branch protection + PyPI environment. All three irreversible steps gated on --confirm.
  • Created ksk5429/kfish (private) + ksk5429/polymarket-oracle-risk (public, MIT). Subtree-split the oracle package into its own history + pushed. Applied protect-main ruleset with required signatures, linear history, and ci-pass status-check gate on both repos.
  • Created GitHub Environment pypi on the public repo. Local uv build && twine check --strict both pass. Runbook runbooks/pypi-first-release.md describes the one remaining browser-only step (pending-publisher registration on pypi.org).
  • Wrote scripts/langfuse_smoke.py — emits one generation span to the configured Langfuse host; returns exit 0 only when the trace flushes.
  • Built this site (quant-notes) and pushed to ksk5429/quant-notes.

2026-04-19 — Production build

  • Full workspace implementation per blueprint. 123 tests green, ruff clean, uv.lock committed. See the feat commit e804d6e.
  • Wire-level: kfish-common DuckDB + ASOF helpers + Claude adapter with prompt caching + OpenAI independent evaluator.
  • 9-persona swarm + Delphi multi-round + asymmetric extremization.
  • Isotonic + Venn-Abers + Mondrian conformal + empirical-Bayes shrinkage ensemble. Atomic CalibratorBundle pickle for hot-swap.
  • Polymarket Gamma + CLOB + UMA Goldsky + Kiwi tokenizer + SimHash dedup.
  • hypekr-bot: Fernet-encrypted two-layer wallet envelope, aiogram 3 with Korean aliases, FastAPI Mini App with HMAC-signed nonces + full ECDSA signature verification on ApproveAgent, pure-functional order construction with builder-fee.
  • polymarket-oracle-risk: NumPyro NUTS posterior fit, out-of-time backtest, Streamlit dashboard, CLI with score / subjectivity / demo subcommands.

2026-04-19 — Scaffold

  • Initial hybrid-monorepo scaffold (pre-build): ADRs, runbooks, CI workflows, migrations, dbt project, Langfuse docker-compose, systemd units.