2026-07-28 — verex history

Source docs: docs/tasks/jul-28-verex.md (task spec) → docs/tasks/jul-28-verex-design.md (design written today).

Jul-28 features implemented — CLOB, multi-outcome groups, async settlement, create-market, how-to guide

All four tasks landed on claude/jul-28-features (11 feature commits, per the approved rev-2 design). Technical summary:

CLOB (api/src/book.ts, Order table): off-chain book, price-time-priority matching in one row-locked DB transaction (market-row lock + FOR UPDATE on crossable makers); resting limit orders are EIP-712-signed at placement (they become makers), IOC taker sub-orders are signed per fill at the maker's exact ratio (ceil buys / floor sells) so CTFExchange.matchOrders' crossing checks pass and chain state mirrors the DB to 1e-6 dust.
ChainJob queue (api/src/worker.ts): single serial in-process lane (= operator nonce management), atomic PENDING→RUNNING claim, 5/25/125s backoff, stuck-RUNNING recovery, per-fill idempotent settlement, terminal-failure compensation (fills reversed, groups cancelled). FIFO order guarantees redeem never lands before its resolve.
Operator MM (api/src/mm.ts): 5-level 1¢ ladders weighted 5:4:3:2:1 (≤2k tokens, asks bounded by minted inventory). After each user fill, the traded market's quote center follows the last price and the group's centers renormalize to Σ=1 — deterministic coherence, the clean-room replacement for nostra's ±5% arb-bot band.
Groups: MarketGroup + N binary member markets (CTFExchange's registry is structurally binary). Seeded HR Derby (7), World Series (8), TIME PotY (6). Group endpoints, group resolve (N payouts in one job), multi-series chart, selectable outcome rows, depth widget.
Create market (api/src/group-create.ts, /create): solvency pre-flight, CREATING→OPEN via CREATE_GROUP job (idempotent resume per member; failure deletes members + cancels the group), operator funds L USDC/outcome (cap 1,000), Yes/No⇒standalone binary. Found live: the seed's exact-amount CTF allowance is exhausted at runtime — the job now tops it up.
Web: settlement chips (2s poll of /jobs/:id) on trade/resolve/redeem + activity badges; /how-to guide with 6 sections and real screenshots (headless Chrome capture), including faucet + checking results per jay's request.
Verified end-to-end on a fresh reset: group trade fills and settles on-chain, renormalization holds Σ=1, group resolve = 7 payout txs, queued redeem pays exactly the expected amount, created 4-outcome group opens with quoted books at mid 1/N.
Left pristine: final ./scripts/reset.sh run — 10 binary markets + 3 groups, wallets #1–5 at 1,000 USDC.

Jul-28 features: design doc drafted, awaiting jay's approval

Explored verex + the nostra-server reference and wrote the full implementation plan into docs/tasks/jul-28-verex-design.md (branch claude/jul-28-features). No code changes yet — the task spec requires jay's sign-off first.

Decision: multi-outcome = N grouped binary conditions, not one N-slot condition

CTFExchange's registry (Registry.sol:41-51) stores exactly one complement per token and enforces it on every fill — a single N-slot CTF condition is untradeable on our exchange. So "who wins the World Cup" = N binary markets + a DB-only MarketGroup, same shape as Polymarket grouped markets and the reference project.

Decision: group prices renormalized deterministically (Σ = 1), diverging from the reference

nostra-server leaves group outcomes unnormalized (an arb bot merely keeps the sum inside ±5%) and its display price is a lifetime VWAP. Verex instead proportionally rescales the other outcomes after each trade so probabilities always sum to exactly 1 — simpler, always coherent, and a clean-room-different algorithm (also serves the copyright-differentiation requirement).

Decision: async settlement via a DB `ChainJob` queue; DB is the UX source of truth

Trade/resolve/redeem answer instantly from the DB and the chain settles behind a status chip; serial worker (concurrency 1) doubles as operator-wallet nonce management; atomic job claim fixes the double-execution flaw observed in the reference's queue. Supersedes the jul-22 doc's "job system would be over-engineering" stance — jay explicitly asked for server-side async. Caveat for jay: the Cloud Run worker needs --no-cpu-throttling + min-instances 1 (cost).

jul-28 design rev 2: CLOB adopted per jay's review; 4 other decisions recorded

jay reviewed the design doc inline. Big change: execution model switches from the existing fillOrder maker/taker flow to a CLOB (order book) — "other platforms use CLOB". Doc revised: new Order table + price-time-priority matching engine (fills instant in the DB), on-chain settlement via CTFExchange.matchOrders through the ChainJob queue, operator becomes a ladder-posting market maker, and group Σ=1 now holds at the MM's renormalized quote centers. Also recorded: option (a) grouped binaries ✅, Cloud Run --no-cpu-throttling+min-instances 1 ✅, anyone can create markets ✅, fee policy deferred to jay ✅. (Source: jul-28-verex-design.md, jay's inline comments.)

Root cause: local seed failed with `returned no data ("0x")` — Sepolia addresses leaked into the local backbone

seed.ts:51 loads packages/contracts/.env as a fallback, which still held USDC_ADDR/CTF_ADDR/EXCHANGE_ADDR from the jul-22 Sepolia deploy (identical to the test manifest in deployments.json). With all three set, the local seed took the "reusing backbone from env" branch and called Sepolia contracts on a fresh anvil (no code → empty return data). Fix: commented the three lines out (with a warning comment) — the seed now falls through to a fresh forge deploy; verified ./scripts/reset.sh end-to-end (10 markets, app renders). Note: the seed's code-exists preflight only guards the test/prod manifest path — adding it to the local env branch is a candidate hardening for the jul-28 branch. (Source: jay's failing ./scripts/reset.sh run, no task file.)

README: "Run locally" made concrete for the anvil lifecycle

Split into first-time setup vs. daily start (anvil is in-memory ⇒ every restart requires ./scripts/reset.sh before using the app) and documented the stale-env-address gotcha above with its exact error message and fix. (Source: jay's request in conversation, no task file.)

jun-19 task map closed out (all 5 tasks ✅)

Per jay: marked tasks 4 and 5 complete — task 4's leftovers (#3 ResolvePanel optimism, #4 SSE) are superseded by the jul-28 async-settlement design (Task C); task 5 is satisfied by the committed deployments.json manifest + VEREX_DEPLOY_TARGET preflight; task 2's domain gap closed via Firebase Hosting (5278191, 9921e23). (Source: jun-19-verex-design.md.)

Replaced the footer "powered by" image with jay's profile avatar (from the rabbit repo): old logo kept as jaylabs-old.png, avatar now jaylabs.png, layout.tsx unchanged in the end. Verified in the browser. (Source: jay's request in conversation, no task file.)

Environment rename: deploy target `test` → `staging` (docs first, then the technical id)

jay renamed the pre-production environment to "staging". First pass changed docs prose only; jay then asked for the technical identifier too, across all scripts: the deployments.json manifest key, VEREX_DEPLOY_TARGET, scripts/deploy.env, deploy.sh/setup-chain-secrets.sh, and the api helper scripts (save-deployment, check-deployment, gen-demo-mnemonic). Every entry point now fails fast on the old test value with a rename hint, so a stale habit or env file can't silently misfire. GCP resource names are unsuffixed and unaffected. (Source: jay's requests in conversation, no task file.)

Noted: pre-existing uncommitted MarketCard.tsx polish left untouched

The working tree already had cosmetic hover/probability-bar changes to packages/web/src/components/MarketCard.tsx from an earlier session; carried on the new branch unmodified, kept separate from jul-28 work.