ThesisResearchIdentityViewsMockupsPublic pageWhitepaperPlanDecisions
Proposal · 2026-10-13 · for the operator to decide

Steel, as a crypto desk would build it.

The dashboard was laid out for stocks and then had a crypto terminal bolted on: seven views that grew by accretion, a light/dark toggle that now wraps, and no identity of its own. This is a proposal to reorganize it around what a crypto trader looks at and in what order, to give Steel a visual system that is not Flint's, and to add a public page and a whitepaper that read differently from Flint's.

The thesis

Steel is the desk's second opinion: it watches every venue around the clock, forecasts the next hour as a distribution, shows its work, pays every real cost in simulation, and never trades unless armed. The interface should be built on the trader's routine, not on where the code lives.

What changes

Views regrouped as Desk, Chart, Flow, Book, Model, Venues, Ops. A cold-steel identity with one accent and a heat scale taken from tempering colours. A public page that opens on the live tape. A whitepaper organised around measurement and cost honesty.

What stays

The engine, the API, the panels as widgets, the chart engine, the seven-key navigation. Nothing here changes a forecast or a fee; it changes where things sit and how they look.

What crypto traders watch, and what the terminals give them

Independent material on what professionals prioritise is thin; most of what is published is vendor positioning. What does recur, across the educational guides and the quant desks' own daily notes, is a routine, run before the session and repeated through the day:

StepWhat they readSteel todayProposal
The clockScheduled events (FOMC, CPI, jobs, expiries); decide in advance whether to trade through them.Next events panel, event_hours feature.A countdown in the status strip on every view.
Funding sweepFunding across many pairs; extremes read as a fragility map, not a signal.Per pair in the screener and the derivatives panel.A market-wide funding and OI strip on Flow; the screener sorted by it in one click.
Open interestWhether price moved with OI rising (new positions) or falling (closing), by side.Positioning pane, OI pane, 4-venue aggregate incl. Hyperliquid.Lead with the OI-confirmed net positioning, not raw OI.
Liquidation mapWhere leverage would be force-closed; forced orders are market orders, so price overshoots clusters.Modeled map with a calibration score against real liquidations.Keep the honesty label everywhere the map appears; show the score beside it.
CVD biasSpot vs perp delta; a persistent imbalance has to rebalance.Tape CVD by size band, spot/perp venue toggles.Spot CVD and perp CVD as two lines by default, the divergence named.
LevelsDay, week, month opens and the previous ranges; VWAP; the volume profile.All present on the chart.Unchanged.
Positioning of othersWhale positions, the crowd (funding, F&G, Reddit), the options dealers.Crowd overlay, GEX, Polymarket.One "who is positioned where" row on Desk.

The terminals

Market Monkey: footprint, book heatmap, TPO, CVD, liquidations, dockable widgets, saved layouts, scripting. Coinglass: the free liquidation heatmap and OI/funding tables, derivatives-first, little spot. Velo: derivatives charting with deep history and a clean chart. Laevitas: options-first, institutional. Bookmap and TradingLite: book heatmap on the chart with footprint clusters. aggr.trade: open-source multi-venue tape with audio and large-trade grouping.

None of them forecasts. None of them keeps a paper book that pays fees, funding and on-chain transfer waits. None of them says how its liquidation map was made or scores it against the real forced orders. That is the identity.

The visual conventions

Dark, number-first, hairline-separated dense panels, tabular figures so values do not shift as they update, green for longs and red for shorts always paired with a label, one accent for live data and primary actions. Hyperliquid's reading of this is a blue-shifted near-black canvas with a single mint accent and white text at stepped opacities. Light themes exist mainly for bright rooms and screenshots.

Steel follows the conventions and differs in the accent and the heat scale, which are its own.

Sources: BloFin, liquidation heatmaps and order flow · a 15-minute order-flow routine · CryptoQuant on funding and OI · a futures-indicator primer · Market Monkey features · a Market Monkey review · Coinglass alternatives (vendor) · Velo (vendor comparison) · Laevitas (vendor comparison) · Bookmap · TradingLite and others · aggr (GitHub) · a reading of Hyperliquid's design system · on light themes (vendor).

Identity: cold steel, one accent, a heat scale

Flint is warm: a warm black, fire orange, gold, a system sans. Steel is its cold sibling. The canvas is a blue-black; the accent is tempered-steel blue; and the one detail only this subject has is the tempering scale, the oxide colours steel takes as it heats (straw, bronze, purple, blue), which becomes the heat scale for every map and the pulse under the wordmark.

STEEL

The wordmark is Archivo at its narrow width, wide-tracked. The temper line under it is the live-data pulse: it drifts while feeds tick and freezes on a stall. Tagline: the market never closes; neither does the model.

Archivo, narrow and heavy, for headings and the wordmark
Archivo at normal width for everything read: panels, notes, the brief. It is a grotesk with a real width axis, so the display and the body are one family at two widths.
JetBrains Mono, tabular, slashed zero: 83,831.06 · +12 bps · 0.0069 %

Tokens

canvas#0a0d12
surface#11161e
surface-2#171d27
temper#6fb3ff · the accent
straw#f2c14e · heat, alerts
up / long#45d483
down / short#ff5c6a
temper scalehot → cold, the heatmaps

Text is white at 94, 72 and 50 percent; hairlines at 9 and 16 percent. Radii 6 on controls, 10 on panels. The light theme keeps the same roles on a cool grey ground (shown if your viewer is set to light). Green and red are semantic and never the accent. Modeled things carry a dashed outline and the word "modeled"; measured things carry a solid one.

Voice. Short sentences, the number first, the source named, no adjectives about the market. "Modeled" and "measured" are always said. It never says what to do.

The views, regrouped by what a trader does

Seven views stay (the keys 1 to 7 work the same), but they are renamed and regrouped so each one answers one question. Every panel remains a widget: drag, collapse, export.

KeyViewThe questionWhat moves here
1DeskWhat is the market doing right now, and what does Steel think?The status strip (regime, breadth, funding, OI change, liquidations 1h, next event, dislocations open, trust), the screener heatmap, the model's live calls with their why, the alerts feed, dislocations, news and the brief. Replaces Market and the Names cards.
2ChartWhat is happening in this pair?The chart with a right rail: tape, book, derivatives, options for the pair; the docked ticket with its rays; replay. The pair detail folds into this view.
3FlowWhere is the leverage, and who is positioned where?Market-wide OI, funding, liquidations and net positioning by pair; the liquidation maps with their calibration; the options surface; the prediction markets; the crowd. New: these are scattered across three views today.
4BookWhat do I hold, what did it cost, what is in flight?Paper and live positions, fills with their fees, funding paid, transfers and their waits, the exchange balances, the algorithm runs, exports.
5ModelCan the forecasts be trusted, and why?The scorecard, calibration by pair, feature drift and net weight, the backtest and sweep, the universe, the indicator editor. Replaces Research.
6VenuesWhat is feeding, and with which keys?Sources and their health, the simulator switch, the signal providers, keys and fee tiers, the probe.
7OpsWhat is Steel allowed to do?Controls, the execution gate and kill switch, series alerts, the GPU handoff, backups, the consoles.

The status strip

One row on every view: the routine's numbers (next event in h, funding median, OI 1h, liquidations 1h by side, breadth, F&G, dislocations open, trusted or not). It is the 15-minute checklist made permanent.

The pair rail

Tape, book, derivatives and options for the selected pair live beside the chart, not in a modal. Selecting a pair anywhere (screener, calls, book) moves the rail and the chart with it; the link is on by default.

The phone

Desk and the alerts work at phone width: the strip stacks, the heatmap goes four wide, the calls list is the page. The rest stays desktop.

Mockups

Built from the tokens above, with numbers from the 2026-10-13 smoke run so the density is real. These are layout studies, not final pixels.

Desk

STEEL
feed: 6 venuestrustedGPU sharedCPI in 28.9 h
regimerisk-off
breadth22 %
funding median+0.0069 % / 8 h
OI · 4 venues$18.2 B +0.3 %
liquidations 1 h$1.5 k L · $0 S
fear & greed57
dislocations open3
GEX BTC+172 M / 1 %
screener · colour by tape delta 1 h96 pairs · table
BTC 83,831 +12
ETH 2,527 −10
ZEC
NEAR
HYPE
RLC
PEPE
UNI
RAY
XRP
OGN
CT
BLUR
SKL
SOL
SUI
LINK
TAO
ONDO
PENGU
AAVE
+ 74
dislocations · opennet of fees
XRP-USD · kraken → binance+6.1 bps12 s
SOL-USD · okx → bybit+4.4 bps3 s
ZEC-USD · coinbase → kraken+3.9 bps41 s
the model's calls1 h · P(up)
BUY SOL+42 bps · 61 %0.08
SELL ETH−38 bps · 36 %0.06
HOLD BTC+12 bps · 46 %·
faded: TRUMPcrowded long·
alertslast hour
liquidation $250 k BTC longs on bybit07:51
rule: SOL funding above 0.01 %07:40
trust: held-out hit 54 % on 212 calls07:00
who is positioned where
perp book net+$310 M longs opened 24 h
dealers (Deribit)long gamma · pinned 80–88 k
crowdleaning long · fade on

Chart, with the pair rail and the docked ticket

STEEL
BTC-USD5m6 venueslayers 14replay
candles from 6 venues' prints · forecast fan +12 bps · 4h +9 · 1d +31 · liquidation map (modeled, 0 of 0 real liquidations scored)
HOLD · entry
83,674.44
—
—
show rayscopyTWAP 6 × 1 minsend via gate · dry run
tape · 1 h3,400 prints / min
delta−2.53 M
$100 k+ prints14 · 9 buys
venue sharebinance 41 · bybit 23 · hl 19
book · USDT classmid 83,696
within 1 %$13.7 M
imbalance 0.5 %−35 %
derivatives · 4 venuesOI $18.2 B
funding (OI-weighted)+0.0069 % / 8 h
basis (perp − spot)+2.1 bps
HIP-3 builder markets159 on 10 dexes
options · deribit592 contracts
zero gamma88,000
max pain80,000
walls90 k · 95 k · 85 k

Desk on a phone

STEEL
trustedCPI 28.9 h
funding+0.0069 %
OI 1 h+0.3 %
liq 1 h$1.5 k L
calls1 h
BUY SOL+42 bps61 %
SELL ETH−38 bps36 %
HOLD BTC+12 bps46 %
screenerdelta 1 h
BTC
ETH
ZEC
NEAR
HYPE
RLC
PEPE
+ 89

The public page

Flint's page explains: what it does, how it works, run it. Steel's should show: it opens on the market itself, and the sections follow the trader's day. Hosted like Flint's, from docs/ on GitHub Pages, at steel.thingg.co (a CNAME beside Flint's flint.thingg.co).

Structure

  • Hero. The wordmark, the temper line, and under it a live-looking tape: a canvas replaying a bundled minute of real prints from six venues, with the forecast fan drawn ahead. Headline: The market never closes. Neither does the model. Two buttons: run it, read the paper.
  • 07:00 UTC. The routine as sections: the clock, the funding sweep, open interest, the liquidation map, CVD, the levels, who is positioned where. Each with one real screenshot and one sentence on where the number comes from.
  • It shows its work. Four claims, each a rule in the code: every map says how it was made; every simulated dollar pays real fees, funding and transfer waits; keys are read-only; nothing is sent unless armed, and the kill switch refuses everything.
  • What it is not. Not advice, not a signal service, not a hosted product. A research toy that runs on your box.
  • Beside the terminals. A short table: footprint, heatmaps, scripting, alerts, API, MCP, and the three things only Steel has.
  • Run it. The uv commands, the Spark service, the token, the data API, the MCP server.

How it differs from Flint's

  • Flint leads with a screenshot; Steel leads with motion (the tape) because the subject is a market that never stops.
  • Flint's sections are about the system; Steel's are about the trader's hours.
  • Flint is warm and explanatory; Steel is cold and declarative: numbers first, every claim with its source.
  • Flint has one page and a paper; Steel adds a live status strip at the top of the page fed by a small JSON the box publishes (optional, off by default), so the page can show "trusted since", "venues feeding", "last backtest" without exposing the dashboard.

The whitepaper

Steel's current paper mirrors Flint's outline (problem, data, model, learning, calibration, policy, pricing, evaluation, system). The proposal is to reorganise it around the two things that make Steel different: it measures a market that never closes from many venues at once, and it refuses to pretend. The outline:

  1. A market that never closes. What changes: no session, no overnight gap, a UTC day that is only a bookkeeping line, a stall that is always an alarm.
  2. Measuring the market. The tape (every print from every venue), the books (one dollar class at a time), the derivatives feeds (OI, funding, liquidations, four venues), the options surface, the calendar, the prediction markets. What is measured and what is modeled.
  3. From measurement to features. The 40 columns; the horizon axis; why a feature-set change resets the net and keeps the replay.
  4. Learning around the clock. Warmup, the held-out guard, the idle trainer, the weekly backtest in the thinnest tape, the GPU shared with Flint.
  5. Trust. The held-out scorecard, the live labels, calibration by pair, what "trusted" means and when it is withdrawn.
  6. Every cost is real. Venue fee profiles, funding by side, the half-spread, impact and Amihud, on-chain transfers with their fees and waits, bounded risk by construction (the isolated-margin short).
  7. Cross-venue structure. Dislocations and their windows, identity, the impossible-gap rule, the arbitrage replay by latency rung.
  8. Honest maps. The modeled liquidation map and its calibration against the venues' real forced orders; the book heatmap's method.
  9. From paper to live. The promotion checklist, the gate, the algorithms, the dead-man rule, the kill switch.
  10. Evaluation. The backtest's totals and the first paper weeks, reported with their caveats.
  11. Limitations. Vendor data, survivorship in the universe, the simulator's history, what the net cannot see.
  12. Future work. Pairs as a conversion graph (K20), real exogenous history, the model's own features.

Different from Flint's in kind

  • Each section opens with a measurement from a real run and its date (the dislocation day, the four-venue OI, the calibration score once there are liquidations to score). Flint's paper argues from design; Steel's argues from the log.
  • A "modeled / measured" margin tag on every figure and table.
  • The same type and tokens as the dashboard and the public page, so the three read as one product.
  • An appendix of every API route, every feature column and every setting, generated from the code at build time so it cannot drift.

Plan

1 · identity

Tokens, type, wordmark, the temper line; the header fixed. The dashboard's palette and fonts move to the new tokens (one stylesheet change; every panel inherits), the theme toggle and the Layout menu stop wrapping, the wordmark and the pulse land in the header. About a day. Visible at once on the smoke instance.

2 · public page

docs/index.html with the live-tape hero, the routine sections, the honesty section and the run-it section, plus the CNAME and the Pages workflow. Screenshots rendered from the restyled dashboard. One to two days.

3 · views

Desk, Chart with the rail, Flow, Model. Panels move between views (they are widgets already), the status strip becomes global, the pair detail folds into Chart, the derivatives panels gather on Flow. Old hashes keep working through the legacy map. Two to three days.

4 · whitepaper

The restructured paper with run data in every section and the generated appendix. One to two days, after the views settle so the screenshots are final.

5 · phone

Desk and alerts at phone width, a web-app manifest, and the sound and notification toggles that make a phone useful. One day.

Decisions needed before building

The accent. Tempered blue as proposed, or something further from the crowd (the straw as the accent and blue as heat would invert the scale and read warmer).
The names. Desk / Chart / Flow / Book / Model / Venues / Ops, or keep Market and Research as names people already know.
The domain. steel.thingg.co beside flint.thingg.co, and whether the page's optional live strip (a small JSON the box publishes) is wanted at all.
The order. Identity first (fast, visible, no behaviour change) and the public page second, then the views; or views first so the public page's screenshots show the final layout.

Answer these four and the build starts with the identity on the running smoke instance, so each step can be checked in a browser as it lands.