Skip to content

Open questions

Currently-ambiguous design choices. PRs and issue comments welcome.

Resolved in 2026-04-19 live-hardening pass

  • /close, /positions, /balance, _reference_price stubs — all now hit live HL info endpoints
  • HYPEKR_APP_SIGNING_SECRET silent-default risk — now required ≥32-char
  • Mondrian conformal stratification bug — fixed with per-class arrays
  • See 2026-04-19 Live-Hardening for the full 18-item changelog

1. Should ksk5429/quant (the old public repo) be archived?

The old public repo at https://github.com/ksk5429/quant was the scratch space before we split into kfish (private) + polymarket-oracle-risk (public) + quant-notes (this site). Three options:

  • Mark ksk5429/quant read-only; add a deprecation notice pointing to the three new repos. Keep history for academic citation.
  • Pros: stops accidental alpha leaks through old commits.
  • Cons: breaks any existing inbound links.
  • Keep as-is; let it be the landing-page / README-only repo that points at the three new ones.
  • Pros: clean "overview" top of funnel.
  • Cons: risk of drift; two places to keep current.
  • Force-delete. Blueprint-literal action.
  • Pros: no alpha-leak surface at all.
  • Cons: destroys history that's already public; rude to inbound links.

Current lean: archive with a deprecation notice.

2. Korean bot — launch criteria

The blueprint targets a Q3 2026 soft launch into 2-3 trusted KR Telegram groups. Pre-launch gates that are currently only partly-ticked:

  • Attorney review of Korean ToS + PIPA privacy policy (budget ₩1-3M)
  • Verified agent-wallet key rotation drill (target: < 15 min)
  • Fly.io NRT p95 /healthz < 300 ms under 10 rps load
  • First real builder-fee payout verified against testnet

Honest: without legal review, launching even "KR-niched" is inadvisable. Paper trading stays default until those gates are green.

3. ECE calibration convention — ours vs netcal's

The Brier-parity validator shows ECE diverges 0.01-0.03 from legacy because legacy uses netcal's bin assignment and ours uses np.digitize with open right boundaries. Two options:

  • keep netcal as a dev-dep and delegate to it in compute_ece — tightens parity to near-zero but adds an optional heavy dep.
  • document the binning convention, gate on Brier only (current stance).

Current lean: document, gate on Brier. ECE is derived, not primary.

4. Sharing strategy with collaborators

This site is public. Three modes of collaboration that stay safe:

  1. Discuss architecture — ADRs, calibration math, oracle-risk methodology. All of polymarket-oracle-risk is MIT.
  2. Discuss performance — Brier decomposition, skill score, reliability diagrams. All on aggregated metrics; no individual-market alpha.
  3. Discuss roadmap — the 90-day plan, HIP-4 readiness.

What never goes on this site:

  • individual market analyses where we hold or will hold a position
  • per-persona prompt versions still in active testing
  • Fernet master keys, bot tokens, API secrets (obviously)
  • whale-address watchlists (could leak via aggregation)

5. What to track on this site vs research-notes?

Clean line:

  • quant-notes (this site) — anything about the K-Fish engine, calibration, oracle risk, the Telegram bot, trading methodology.
  • research-notes — offshore-wind PhD work, papers, thesis chapters.

Cross-links only when a technique genuinely transfers (rare).

How to add a question

Edit this file directly via the "Edit this page" pencil, or open an issue on quant-notes.