Levandor

BNB 5m Expansion — Evaluation Groundwork (2026-07-05)

8 min read

Groundwork for expanding the crypto 5m algo (currently btc/eth/sol/xrp/doge) to BNB. Prompted by the historian noticing BNB passes every documented gate yet was skipped by inheritance (the Asset enum copied cohort_algo’s 4 majors + doge), not by any decision. This note captures the universe survey, feed/market verification, the observe-only code that shipped (now merged to main + live as an observe container, recording verified — 2026-07-05), and the one open go/no-go (calibration transfer).

For Agents — status

UPDATE 2026-07-05 — the BNB observe lane is LIVE. feat/bnb-observe is merged to main (@deploy, rebase-merge 336bebea, “PR #1”) and the observe-only ASSET=bnb container is rolled + recording is verified end-to-end (see the LIVE section immediately below). BNB remains observe-only, NOT armed, and cannot arm by construction (the emit_supported allowlist). The go/no-go is still the calibration-transfer check — now unblocked: BNB p_fair is accruing, so the clock has started. Next = bank ~a few days, then run the frozen-BTC-calibration ~2 SE transfer check vs the 1009-window label set. babylon pmv2 #1052 (deploy) / #1053 (crypto confirm).

LATER 2026-07-05 (13:22 CEST): BNB was ARMED LIVE (bnb_5m_algo_up/_down = live/enabled in pmv2_autotrade_sources) — before the calibration-transfer confirmation this note recommends. Early ~breakeven (~21 resolved) + an early adverse-fill hint. See the ARMED LIVE section below.

LIVE — observe container rolled + recording verified end-to-end (2026-07-05)

@deploy merged feat/bnb-observe to main (rebase-merge, commit 336bebea, “PR #1”) and rolled the observe-only ASSET=bnb container. Recording is VERIFIED end-to-end from the crypto_shortterm read path:

CheckValueMeaning
asset='bnb' observations786 rows / 4 windows (landing 07:55–08:11Z)pipeline is writing BNB rows
avg K (strike)575.08matches the on-chain Chainlink BNB/USD price I verified (~$571–575) → aggregator 0x82a6c4AF830caa6c97bb504425f6A66165C2c26e is correctly wired
avg p_fair0.568, ZERO nullsthe fair-value model is computing on BNB (no null/NaN gaps)
book asks0.53 up / 0.41 down capturedboth token books are polling

Observe-only confirmed — there is no bnb emit container (by construction; the emit_supported allowlist hard-refuses BNB at boot). So the code change is validated live, and the calibration-transfer clock has started.

Coordination: babylon pmv2 #1052 (@deploy rolled the container) / #1053 (crypto confirmed recording live).

UPDATE 2026-07-05 (13:22 CEST) — BNB ARMED LIVE (before the calibration-transfer confirmation)

BNB was armed live BEFORE the go/no-go this note recommends

After the ~1h dry period, bnb_5m_algo_up / bnb_5m_algo_down were set live / enabled in pmv2_autotrade_sources at 2026-07-05 13:22 CEST. BNB was therefore promoted past the observe-only-by-construction state — for emit to fire BNB it must have been added to the emit validated-asset allowlist {Btc,Eth,Sol,Xrp,Doge} from the safety-hardening section below. This happened before the frozen-BTC-calibration ~2 SE transfer check (still the open validation in “Next — the calibration-transfer go/no-go”).

Early live read: ~breakeven, ~21 resolved — far too small to conclude anything.

Early adverse-FILL hint — WATCH (the btc-up signature)

In the 17–19 CEST window, BNB signal won 10/11 (91%) but fills won only 7/11 (64%), filled at 0.668 (the cheapest entry). This is the btc-up adverse-fill signature (crypto-shortterm-lane-state-adverse-fill-2026-07-05) — we fill the cheap (about-to-lose) asks and miss the winners whose ask ran away. n is tiny; if it holds, BNB has the same execution problem the speculative-upside note below hoped it might avoid (less-efficient venue ⇒ cheap asks not pre-informed).

Attribution caveat — manual BNB contamination. The proxy wallets carry some manual BNB trades (Balu). The algo majors are clean/algo-dedicated (836/837 on-chain conditions matched PM-placed; only 1 manual Paraguay bet — see crypto-shortterm-pnl-onchain-cash-truth-2026-07-06), but a live BNB lane shares the wallet with manual BNB → BNB P&L attribution will be muddied unless separated. Factor this into any BNB go/no-go read.

The 5m Polymarket universe is exactly 6 assets

The Polymarket 5-minute Up/Down universe is btc / eth / sol / xrp / doge + bnb — and nothing else. Confirmed empirically via the Gamma API: 38 candidate coins (all top-30 by market cap + the major memecoins) were probed; only these 6 have active 5m up/down markets. Polymarket runs 5m up/down only on the most liquid majors. This matches the backtest-harness spec’s documented candidate list.

Discovery method (reusable)

  • Query events?slug={asset}-updown-5m-{ts} on the Gamma API — the working probe.
  • The series pattern is {asset}-up-or-down-5m.
  • The tag_id endpoint is a dead end — it returns a schema blob, not market data.

Consequence: BNB is the only untapped 5m asset — there are no others. Any further 5m breadth expansion is impossible on Polymarket’s current listing; the growth axis after BNB is the 15m series (separate track, see crypto-shortterm-competitor-wowitsamazing-refuted-2026-07-03), not more 5m coins.

BNB feeds verified

Both data legs the pipeline needs exist and were verified.

  • Binance: bnbusdt spot ticker exists (the ingest lead feed).
  • Settlement (on-chain Chainlink, Polygon): BNB/USD aggregator = 0x82a6c4AF830caa6c97bb504425f6A66165C2c26e.
    • Verified on-chain via eth_call description() → "BNB / USD", returning a sane live price (~$571).
    • Method cross-checked against the known BTC feed (same call shape).
    • Sourced from the official Chainlink registry (reference-data-directory.vercel.app/feeds-matic-mainnet.json), whose BTC and DOGE addresses exactly match the codebase’s existing aggregators — so the registry is trustworthy for BNB.

BNB market structure is healthy

Backfilled 1009 windows over 3.5 days (Gamma winners harvest, same pattern as the BTC backfill):

  • A market in every 5m slot — continuous, no gaps.
  • Balanced 51/49 up/down — no directional skew.
  • 1007 / 1009 resolved.
  • Series **24h volume ~153k), our current smallest lane.

BNB’s book is deeper and more continuous than expected for a smaller-cap name. Labels persisted to data/bnb5m_actual_winners_2026-07-05.csv — the ground-truth set the calibration-transfer check will score against.

Code shipped — feat/bnb-observe (observe-only + safety hardening)

Branch feat/bnb-observe (5673d6e) — merged to main 2026-07-05 by @deploy (rebase-merge, commit 336bebea, “PR #1”). Adds Asset::Bnb, observe-only. All 4 build gates clean.

Safety hardening — emit_supported allowlist (critical)

emit_supported was gating only on window_secs == 300. That check refuses the 15m series (900s) but would NOT have refused a new 5m asset like BNB — so adding Asset::Bnb to the enum would have silently made it emit-eligible.

Changed to require an explicit validated-asset allowlist {Btc, Eth, Sol, Xrp, Doge} AND window_secs == 300. Consequences:

  1. BNB is observe-only by construction — emit hard-refuses at boot.
  2. No future 5m asset auto-arms on being added to the enum — it must be explicitly promoted (added to the allowlist) after passing its own gate.

Test: emit_arms_only_validated_5m_assets.

This mirrors the 15m window_secs-refusal invariant (crypto-shortterm-algo-accuracy-audit-2026-07-02, 15m benching) but closes the gap that invariant left open for same-interval assets.

Coordination

  • babylon pmv2 #1049 — asked @deploy to merge feat/bnb-observe and roll ONE observe-only ASSET=bnb container (no migration, no NATS, no PM rows), explicitly NOT arm BNB.
  • babylon pmv2 #1052 / #1053 — DONE (2026-07-05). @deploy merged (336bebea, “PR #1”) + rolled the observe-only ASSET=bnb container (#1052); crypto confirmed recording live end-to-end (#1053). See the LIVE section at the top of this note.

Next — the calibration-transfer go/no-go (open)

The actual go/no-go is the calibration-transfer check: does the frozen BTC 20-bin calibration map transfer to BNB within ~2 SE? (The same test that cleared eth/sol/xrp/doge — Cross-asset transfer — BTC calibration generalizes.) It has not run yet, but is now unblocked: as of 2026-07-05 the observe container is live and BNB p_fair is accruing (Path 1 below is underway). Two paths:

  1. Deploy the observe container (few days of accrual) → then a one-query check of live p_fair against the 1009-window label set. Cleanest. ← now underway — container deployed 2026-07-05 (336bebea); bank ~a few days, then run the check.
  2. Binance + Chainlink harness backfill of p_fair offline → a faster read without waiting on deployment.

Economics. BNB is expected to be a doge-sized lane (~$5–30/day) — small, but near-free if the calibration transfers (feeds verified, market healthy, observe-only code done).

Speculative upside (untested)

BNB is less efficient than BTC, so our model’s disagreement against the BNB book may be less adversely-selected than BTC’s. This ties directly to the btc-up adverse-fill finding (we systematically fill the losers on BTC because the market already knows). A less-efficient venue could mean the cheap asks we fill are not pre-informed — i.e. the adverse-fill inversion may be weaker on BNB. Unproven; a reason to actually run the check rather than assume doge-parity.

Graph View

Backlinks