Sports and exchanges, 48 hours: backfills into the web app, then a public repo with CI (29 Sep – 1 Oct 2026)¶

From 2026-09-29 16:00Z to the morning of 1 Oct (UTC), most of the work went into one problem: getting the backfills for every sport and exchange to reach the production web app. NASCAR and MotoGP went from tape-only listings to sports the app shows end to end: results, identified Kalshi markets, a simple model price for NASCAR's next race, and a demo paper P&L history on the Strategy and Positions pages. The owner's 3 × 3 soft-launch grid (F1, NASCAR, MotoGP × Polymarket, Kalshi, OG.com) has 7 of 9 cells with something in them; MotoGP on Polymarket (no linked markets) and MotoGP on OG.com (not listed) are the two empty ones. The last stretch moved the code to the public repo lalligagger/racinglines-public, where merging to main now deploys to the VM and deploys carry code only, so data and the database stay on the VM. The honest bottom line has not moved: the only held-out-robust strategy is still F1 profile A on Polymarket, which has listed no F1 race since 28 Aug. Every NASCAR and MotoGP dollar below is in-sample paper, and almost all of it is NASCAR 2026.

This replaces the 24-hour report of 30 Sep (kept in the old private repo, first version PR #104, final #113). PR numbers with a # (#85, #110 …) are the old private repo's; "PR 2" to "PR 9" are the public repo's, where numbering restarted.

Piece	Result on the morning of 1 Oct	Label
Production app	Markets shows every sport's status, NASCAR and MotoGP sections with model prices, next races and recent results; Strategy splits the record by sport and maker/taker; the OG.com column is on	Live, behind `RACINGLINES_SPORT_STATUS=1`, `RACINGLINES_SPORT_PAPER=1` and `RACINGLINES_OG_VENUE=1` (all on on the VM)
NASCAR paper record (Kalshi)	Pro demo 2025 −$278.31 (8 races), 2026 +$6,525.45 (32); Basic demo −$357.23 and +$2,657.23	In-sample, demo only
MotoGP paper record (Kalshi)	2026 only, 11 races: Pro +$31.25, Basic +$7.70; no 2025 markets	In-sample, no held-out season
NASCAR next-race price	Forecast run #9068 stored; the board and race pages show a fair price	Simple baseline, uncalibrated
MotoGP next-race price	None: no MotoGP calendar is ingested, so there is no next race to price	Gap
F1	Profile A stays (106-job sweep); demo taker paper 2025 +$450.46, 2026 +$353.01	Robust on Polymarket; paper
Champion replays (#95, #96)	Merged 30 Sep 22:41Z; the VM runs (handoff steps 5 to 7) not done yet	Gap
Kalshi and OG.com books	Recorded every 5 minutes on the VM since 30 Sep 23:39Z	Live; books only, no trade tape
Code delivery	Public repo; merge to `main` = staging smoke, VM deploy, production smoke. First full green run 08:14Z; PRs 2 to 9 merged	Live; tests in CI since PR 9
Data safety	The VM's disk is the only copy of the newest database dumps, backups and archived market history	Open: STG-5
Beta sign-up	Built (PR 7) behind `RACINGLINES_SIGNUP`, off by default; switched on by `vm.sh accounts`	Built; switch-on not verified here

Screenshots. The 24-hour report's screenshots were placeholders, and this version has none: the cloud session that wrote it can't reach the site. Retaking them (1024 px wide, 150% zoom, with browser_smoke.py) is an open item.

1. Backfills into the web app: what a visitor sees now¶

The backfills became visible through #105-#112 (deployed on 30 Sep, VM on 306a147) and then through the public repo's CI deploys on 1 Oct. The /markets timeout that failed the 16:05 smoke check on 30 Sep (#101 counted the Parquet archive inside the request) is fixed by #103.

Page	What changed	PR	Switch (VM)
Markets	"Every sport: status" table: one row per sport schema with race data, exchange markets, model, backtests and (Pro) the demo paper record; missing pieces say so	#110	`RACINGLINES_SPORT_STATUS=1` (on)
Markets	NASCAR and MotoGP sections: the model's replay-save prices, next scheduled races, recent results (MotoGP counts the RDR category)	#111	same
Markets, race pages	NASCAR next-race fair price from the stored forecast run	#112	— (data)
Strategy (Pro)	"Where the P&L came from": the record split by sport and by maker/taker	#110	—
Positions, Signals, nav P&L	NASCAR and MotoGP demo positions with race names, dates and a "demo replay, in-sample" pill; buy-all rows never shown	#110	`RACINGLINES_SPORT_PAPER=1` (on)
Strategy pages	NASCAR and MotoGP sport filters; the sport catalog dropdown lists every modeled sport	#108, #109	—
Docs	`docs/coverage.md` is the status page (sport × exchange matrix, honest model status)	#105, #106	—
Markets	Per-venue "Order books recorded" line (live / stale since / no data yet); the OG.com column	#119	`RACINGLINES_OG_VENUE=1` (on)
Every page	Maintenance popup on sign-in and app pages, dismissed once per browser session (`MAINTENANCE_NOTICE` in `racinglines/web/app.py`; `""` turns it off; changing it is a deploy)	PR 2	—
Sign-up, sign-in, admin	`/signup`: scrypt-hashed account plus 1,000 fantasy bucks in `accounts.fantasy_ledger` in one transaction; every sign-up starts as pro; no email collected; `/admin/users` one-time password reset with a ready-to-send email, Remove and balances; "Forgot password" opens a prefilled email to [email protected]. The `accounts` schema is created by `vm.sh accounts` (backup first), never by a deploy	PR 7	`RACINGLINES_SIGNUP=1` (off by default)
/docs	Built on the VM at every deploy (was answering "Docs not built"); `vm.sh docs` after a docs-only merge	PR 7	—
README	CI, Python, license and Tests badges; contributor guide and an invite to collaborators	PRs 7, 9	—

browser_smoke.py (PR 7) signs in with the demo buttons as pro and basic, follows same-site links two levels deep, reports 4xx/5xx and JavaScript errors, and saves 1024 px screenshots. It is read-only and hasn't been run against racinglines.bet yet.

2. Data status per sport and exchange¶

What the production database held on the evening of 30 Sep, and how far each cell can be trusted. "Paper" means the demo accounts' backfilled record, never real orders: both trading flags are unset and OG.com has no order code.

Sport × exchange	Data on the VM	Model price	Backtest	Paper record	Verdict
F1 × Polymarket	2025-26, 7,285 links, prices and trades	Real model	Robust (A: +$1,389 / +$1,432 at 16k, #85)	Taker 2025 +$450.46, 2026 +$353.01	Only tradeable cell; nothing listed since 28 Aug
F1 × Kalshi	2025-26, 3,793 links, trades, last-trade candles; no books	Real model	Maker K +$490 / +$659 (frozen); taker not usable	Maker 2025 +$193.39, 2026 −$118.84	Needs bid/ask candles
F1 × OG.com	20 season futures since 29 Sep	Real model (season)	None	None	Column on since #119 (`RACINGLINES_OG_VENUE=1`); books recorded every 5 min
NASCAR × Kalshi	408 events 2016-26, 7,544 links with a race, tape	Simple baseline; next race priced (#9068)	Not robust: every setting loses 2025	Pro −$278 / +$6,525; Basic −$357 / +$2,657	Demo only
NASCAR × Polymarket	Backfilled on the VM with Kalshi's selection (`vm.sh demo extra`)	Same baseline	No own grid	+$4.82 paper history, 2 positions (26 Jul 2026)	Demo only
NASCAR × OG.com	17 Cup Champion contracts since 29 Sep	Champion sim (#93), uncalibrated, offline	None (buy-all is a plumbing check)	None	Mechanical only
MotoGP × Kalshi	202 events, 289 race-winner links, tape	Simple baseline, race win only; no next race (no calendar)	No held-out season (no 2025 markets)	2026 Pro +$31.25, Basic +$7.70	Demo only
MotoGP × Polymarket	No linked markets	—	—	—	Empty
MotoGP × OG.com	Not listed by OG.com	—	—	—	Empty

How to read it: the paper record picks each market kind's best setting from the same seasons it then replays (--book best, docs/paper-trading.md), so it shows what the demo would have done, not an edge. 2026 is 32 of NASCAR's 40 races and all 11 of MotoGP's. The one test that separates settings from seasons, the NASCAR taker grid, is below: the same model loses 2025 at every setting.

Realistic model status. F1 has a real model and one robust strategy. NASCAR and MotoGP run a result-only baseline (recent finishes, a team prior; no qualifying, practice, track type or sprints) that has not been compared with the market or a naive baseline; the replay's race spread (2.0 places) is tighter than NASCAR's measured ~7. Treat every NASCAR and MotoGP number as a working demo of the pipeline, not a trading signal.

What is still missing. A MotoGP calendar ingest (so MotoGP gets a next-race price); Polymarket MotoGP links; runs of the champion replays (#95 F1 on Kalshi and OG.com, #96 NASCAR and MotoGP; code merged 30 Sep 22:41Z); forecasts refreshed after each race (today a forecast run is stored once, by hand; the timer is a later step); Kalshi bid/ask candles before any Kalshi taker result counts.

3. Getting it to production: the public repo, CI and code-only deploys¶

Once the backfills showed on the app, the question became how changes reach production without risking the data behind them.

The public repo. The owner rewrote history on 1 Oct (~06:10Z) and made lalligagger/racinglines-public the live repo. Raw data, archives, run outputs, reports and tests/fixtures/ are not in it (.gitignore keeps it code-only); the ten /pitch images and the fixture bundle live only on the VM's disk and the owner's Mac, and the fixtures are offered to contributors on request. Doc links point at the public repo (PR 5). 62 old branches were deleted at 08:08Z; since then one feature-track branch per night or sprint, head branch deleted at merge, the owner squash-merges.

How a merge reaches production (main-merge-gate, VM deploy: The flow):

detect-docs-only skips everything when a change set is only docs/ and *.md files.
predeploy-staging smoke-checks staging.racinglines.bet (whatever staging serves, not the PR's code).
On a push to main, deploy-production runs vm.sh deploy <sha>: pause the live timers, untrack what the new commit doesn't track (backed up to data/backups/files/), checkout, pip, alembic upgrade head, db seed, build /docs, restart, resume the timers, print each timer's state.
A production smoke check of racinglines.bet.

New vm.sh subcommands: repoint (point the VM checkout at the public repo and fetch; no file touched), backup <purpose> (a database dump on the VM), docs (rebuild /docs), accounts [off] (backup, accounts schema, sign-up switch). A deploy that fails before the checkout moves resumes the paused timers; one that fails after leaves them paused on purpose.

It took four failed CI deploys to get there. Each deployed nothing and left or put the timers back:

PR	Commit	Why its deploy failed	Fix
PR 2 (maintenance popup)	`cd7bf32`	`fatal: Needed a single revision`: the VM still fetched the private repo	PR 3: `vm.sh repoint`; deploy refuses until repointed
PR 3 (code-only deploys)	`a449e18`	Refused until repointed; deployed by hand at 07:19Z	—
PR 4 (repoint as app user)	`59a66af`	`scp` Permission denied on a `/tmp` file left by another ssh user; a fresh runner's ssh-keygen output on stdout fooled the origin check	PR 4 reads only the last line; PR 6 pipes `untrack.sh` over ssh
PR 6	`4469100`	First full green run (run 19, 08:14Z)	—

Code only, and where the data lives (owner rule, 07:00Z; audit in VM deploy: Code from GitHub, data on the VM). A deploy never restores, overwrites or deletes data/, the pitch images or database rows; those change only by ssh and scripts (backup first) or the site's admin views. The audit found:

The VM's disk is the only copy of the newest database dumps, data/backups/ and data/archive/markets/. No script or timer pushes them off the VM. A nightly off-VM backup is STG-5, waiting on the owner.
Database dumps are no longer a full backup on their own. The recorder's hourly archive pass moves cold market rows from Postgres to Parquet under data/archive/markets/ (dumps went from 118 MB to 56 MB at 01:56Z on 1 Oct, row counts checked).
Cloud sessions read stale data. Their bucket access reaches only the old bucket, last pushed on 28 Sep.
Re-committing a path the VM keeps untracked (pitch images, tests/fixtures/) makes the deploy's checkout fail after the untrack step, and the timers stay paused.

What CI still doesn't do (full table and STG-1 to STG-11 in todo: Staging and CI deploys): about 50 tests, all goldens included, skip on a public checkout because tests/fixtures/ is absent (STG-7); the staging smoke doesn't run the PR's code (STG-1, STG-3); a merged migration runs on the VM at once with no backup step (STG-8); there is no concurrency group (STG-6); nothing stops a code merge during a live window (STG-9); the Claude GitHub App isn't installed on the public repo, so PR and CI events don't reach Claude threads (STG-11); cloud sessions can't push .github/workflows/*, so workflow changes ship as patches. Tests in CI (STG-2) arrived with PR 9.

4. Recorders and round 16¶

The Kalshi and OG.com book recorder (vm.sh record, timer racinglines-record-venues, every 5 minutes) has run since 30 Sep 23:39Z (first pass backed up the database first); the owner confirmed rows landing at 04:04Z on 1 Oct. It has no end date; vm.sh record off stops it.
Round 16 (Sepang): the book opens Fri 2 Oct 03:30Z and closes Sun 4 Oct 07:00Z. At 08:14Z on 1 Oct the Polymarket recorder, the signals timer, [email protected] and the record-venues timer were active. Because merging deploys, no code merges in that window unless the owner says so. Still owed by the owner: the tier call (Thu 18:00 PDT), the loss cap, hype picks and crowd pace.

5. Lessons learned¶

#	What happened	Cost	Lesson (now a rule where noted)
1	The overnight run stopped at the NASCAR Kalshi tape pull: 13,599 links, 0 with a race, because NASCAR results and linking had only ever run on the Mac	~3.5 h and a manual fix	A dry run checks per-sport readiness (events, races, links with a race), not row counts. #98 added the preflight
2	#101 counted the Parquet archive inside the /markets request (~30 s); the deploy smoke check read `000`	A failed smoke, a second deploy	Anything that scans the archive runs in the background; read the deploy smoke before calling a deploy done (#103)
3	Overnight jobs under `systemd-run` printed nothing for long stretches: Python buffered stdout, and some steps had no progress line	"Is it stuck?" questions, relay steps for the owner	Every long-running job prints flushed progress at least every 5 minutes (owner rule, 2026-09-30 22:36Z)
4	The deploy from the Mac Remote Control session was blocked by its permission rules: ssh to the VM needs an exact-command allow rule, and only the progress tail had one	Deploy steps fell back to the owner pasting	Give the owner one command that does everything; add the exact allow rule before a planned remote step
5	About 30 PRs in 24 hours, several stacked; after a squash-merge of the base, the stacked ones conflicted, and a local squash left GitHub PRs open	Merge-order churn, stale drafts (#97)	Group work into feature-track branches (owner, 22:36Z); check PR state before giving merge commands
6	The overnight "Profile A" row was A's settings with the `early` taker, not A's `update` taker (inferred from the candidate id)	A false gap (+$2,607 vs +$1,389)	Label every row with its strategy family and candidate id; never compare across families
7	MotoGP's grid ranking crashed on a season with no markets (`KeyError 'net'`)	A re-run	Code for a new sport must handle an empty season; tests should include one (#102)
8	Report numbers were copied by hand from logs and PR bodies	Each number traced by an agent; the paper totals for Polymarket NASCAR are still unread	See recommendations below: reports should be built from saved run outputs
9	Merging became a deploy on 1 Oct; the first CI deploys of PRs 2, 3 and 4 all failed on the VM side (wrong origin, refusal, `/tmp` permission, ssh-keygen noise on stdout)	Three failed deploys; the round 16 timers paused once	Before telling the owner to merge, say what the merge deploys; read the deploy run, not just the PR check
10	Check and smoke results were read from an earlier commit or run	"Green" claims that didn't hold for the latest push	Read the check run on the PR's current head, and whether a run is still going after the last push
11	The workflow has no concurrency group	Two merges close together would run two deploys over each other	Merge one PR, wait for its Actions run, then the next (STG-6)
12	Two versions of the same command, and stub or duplicate messages, reached the owner	Re-pasted commands; a command already run was pasted again	One command that does everything, sent once; check what already ran before advising
13	`as_app "a && b"` ran only `a` as the app user	A repoint that half worked (PR 4)	Wrap chains in `bash -c` when switching users over ssh
14	Deploys didn't build mkdocs, so `/docs` said "Docs not built"	A broken docs link on the live site	The deploy builds `/docs`; `vm.sh docs` after a docs-only merge, which deploys nothing
15	Golden tests skip silently on the public checkout	`pytest -m "not live"` looks green with ~50 skips	Run goldens on the Mac for any pricing or model change (STG-7)

6. Report consistency and reproducibility: recommendations¶

Today a report is a Markdown file an agent writes from logs, PR bodies and memory; nothing re-runs it. The 24-hour report's figures were the first step: its make_figures.py (old private repo) rebuilds every chart from numbers typed in with their source, without a database. The recommendations, in order of value (the owner asked for these as roadmap and todo items):

Every run writes a machine-readable summary. Replays, grids, demo-history and forecasts write a summary.json (inputs, settings, git commit, seasons, net P&L per season, counts) beside their Markdown output, and data_changes rows point at it. Reports read those files instead of copying numbers.
A racinglines report build <folder> command. Reads a report.toml (which run summaries, which queries, which figures), renders tables and charts with the same code every time, then the existing markdown_html.render to HTML and PDF. The prose stays human- or agent-written; every number and chart is generated.
Screenshots by script. A Playwright script with the fixed viewport (1024 px wide, device scale 1.5 for the 150% zoom) and a page list per report, run against staging or production with a demo login.
One template for large-update reports: scene-setter, top-line table, a section per workstream, what-shows and what-doesn't, opens with owners. The two existing templates in CLAUDE.md stay; this adds the third kind.
A consistency check in CI: report build --check fails if a report's numbers differ from the summaries it cites.

7. Open items¶

Open	Who	Blocker
Champion replays (#95, #96): VM steps 5 to 7 of `handoffs/2026-10-01-season-replays.md`	Owner	Backup first; not during the round 16 book
Off-VM copy of the VM's data (STG-5)	Owner decision	Yes or later, unanswered
Migration gate (STG-8), staging host (STG-1: same VM or its own, keyless auth or key, Cloudflare Access or password)	Owner decisions	—
Sign-up switched on (`vm.sh accounts`) and checked on the live site	Owner	Unverified from here
Fantasy tiers (everyone starts as pro)	Owner decision	—
MotoGP calendar ingest, so MotoGP gets a next-race price	Next thread	Needs a Mac probe of the calendar endpoint first
Polymarket MotoGP links	Next thread	Tag slug unverified; Mac probe first
Forecast refresh after each race (timer)	Next thread	VM sign-off
Kalshi bid/ask candles re-pull	Owner decision	Needed before any Kalshi taker result counts
MotoGP grid ranking re-run, OG.com buy-all output, NASCAR spike share	Owner or Copilot	Written on the VM, not yet read
Report screenshots	Owner or a Mac session	Retake at 1024 px, 150% zoom (`browser_smoke.py`)
Round 16 calls: tier, loss cap, hype picks, crowd pace	Owner	Book opens Fri 2 Oct 03:30Z

What this does and doesn't show¶

It shows what the repo history (old #85-#121, public PRs 2 to 9), docs/todo.md, docs/vm-deploy.md and the project's notes record. No number from the 24-hour report was recomputed; they are as of the evening of 30 Sep. It doesn't verify the VM after 08:14Z on 1 Oct: PR 7's and PR 9's deploy results, whether vm.sh accounts has run, and the timer states come from notes or the owner's word, not from output read here. Every NASCAR and MotoGP P&L is in-sample demo paper, not an edge.

Appendix: the morning report of 30 Sep (as of 16:00Z), kept as the record of the overnight run.

The owner's 3 × 3 grid (morning)¶

F1 is the only full row, and only F1 on Polymarket has a held-out-robust backtest. Every other filled cell is data plus either a backtest that failed its held-out season, a mechanical check (buy-all), or an uncalibrated champion price.

Sport	Polymarket	Kalshi	OG.com
F1	Backtested, robust. Profile A (taker) and C (maker), 2025-26 races; overnight 16k re-confirmed A. No F1 race listed since 28 Aug 2026	Backtested, not usable as a taker. Maker K tuned (2026 +$490, 2025 +$659, frozen pending review). Taker replay reads last-trade prints, which spike (1.67% of points >15 pts off median)	Priced, not backtested. 20 season futures, model fair via the fair-price indicator (`RACINGLINES_OG_VENUE=1`, off on the VM). Buy-all ran overnight, output not yet read. #95 champion replay drafted
NASCAR	Gap. 1,101 links on the Mac, champion priced offline by #93; no race markets we can trade, no backtest	Backtested, not robust. 40 races with markets; 2026 +$4.2k to +5.5k, 2025 −$242 to −$390	Mechanical only. 17 Cup Champion contracts; buy-all ran, output not yet read; champion priced offline by #93 (uncalibrated)
MotoGP	Gap. Champion market (25 outcomes) is quote-only; per-race winners seen 29 Sep but not synced or backtested	Replayed, no verdict. 289 race-winner links; tape pulled; 16 grid runs done; ranking re-run pending the #102 deploy	Not listed by OG.com

Wider coverage outside the 3 × 3 (from docs/coverage.md, unchanged in this window): IndyCar has 1,374 closed Kalshi links but its results source's terms forbid automated use; road cycling, Le Mans and SailGP have tape-only schemas (#50) and no results source; downhill has no venue.

What each exchange offers historically¶

Kalshi is the only venue with deep, uncapped history for the new sports; OG.com keeps about a month, so its history starts on 2026-09-29, when we began pulling. PR #101 fixed the board's Exchange data counts, which had read only the Postgres buffer (Polymarket showed 0 trades, Kalshi 37,416).

Exchange	Trades	Price history	Order books	New-sport listings
Polymarket	Taker trades only, newest ~100,000 per market	`/prices-history`, hourly by default	No history API; only what we recorded live	NASCAR race winners + champion, MotoGP champion and some Grands Prix
Kalshi	Every trade, no cap (older than ~2 months via `/historical/trades`)	Hourly candles: last trade, else bid/ask mid. An empty side is now no price, not 0.50 (#89, #90)	No history API; no Kalshi book recorder yet	NASCAR 13,599 links (race, top-N, h2h, pole, fastest lap, champion), MotoGP 322, IndyCar 1,374
OG.com	About one month	1-minute bid/ask/last, 31 days per call	One snapshot per `books` run	F1 20 futures, NASCAR 17 Cup Champion, SailGP 13; no MotoGP, IndyCar, cycling or Le Mans

Counts on the cloud archive copy (mostly F1, from #101): Kalshi 1,065,605 trades and 520,969 hourly price points; Polymarket 447,274 trades and 3,983,682 price points. The VM's counts, which include the overnight NASCAR and MotoGP pulls, have not been read yet. NASCAR Kalshi hourly prices on the VM went from 152,257 before the dead-book cleanup to 295,553 during the overnight pull (docs/data-changes.md).

What was built¶

About 30 PRs landed in the window, and three more are open drafts. Grouped by what they unlocked:

Area	PRs	What it does	State
Ingest plumbing	#57, #58	Syncs seed their own sport rows; `vm.sh deploy` runs `db seed`; exchange schemas state API caps and fail to load if exceeded; one exchange data contract for charts	Merged
NASCAR results and identity	#60, #62, #63, #64, #67, #69	`cf.nascar.com` feed adapter (probed read-only from the Mac first), scheduled events, driver and race resolver, links carry driver, race, kind and h2h pairs for Kalshi, Polymarket and OG.com	Merged
MotoGP results and model	#75, #77, direct commits 72f16c0, 6e50e57, adac98d	Results API probe (MotoGP allowed, IndyCar terms forbid scraping), result-only race model (race-win log loss 0.0796), pricing and board parity	Merged
Taker replay for new sports	#82, #84	`nascar replay` / `motogp replay`: taker P&L on Kalshi and Polymarket tapes, per-race tape probe and pull, NO TAPE flags, `--save` with undo. New sports' defaults switched on (owner, 03:07Z)	Merged
Price hygiene	#89, #90, #91	Empty Kalshi book side is no price (was a 0.50 mid); dead-book quotes rejected; MotoGP UUID athlete ids kept; 27,024 bad Mac rows deleted and re-pulled, 157,049 on the VM	Merged
OG.com as a venue	#88	OG.com is a replay venue; debug `buy-all` buys one YES and one NO of every listed market to prove each is found, priced and settled	Merged
F1 taker	#81, #85, #87	Thin-market exception, Kalshi taker fee, re-sweep (A stays, blend TB offered to Pro), demo taker walk-forward story	Merged
Champion markets	#93; #95, #96 drafts	NASCAR Cup season sim through the 2026 Chase (read-only, uncalibrated); F1 champion replay on Kalshi and OG.com; NASCAR and MotoGP champion replays	#93 merged; #95, #96 wait on a Mac spot check
Overnight VM run	#86, #92, #94, #98, #99, #100, #102	Runner, 3 × 3 plan, n2-standard-2 sizing, per-sport preflight that loads results and links, progress log, MotoGP ingest newest first, grid ranking with an empty season; report, decision log, data-changes entry	Merged
Exchange data counts	#101, #103	Board counts span the Parquet archive plus buffer, n/a for zero, per-exchange source notes; archive counted in the background after #101 timed out the deploy smoke check	Merged; #103 not yet deployed
Sport paper portfolio	#97 draft, then #110	NASCAR and MotoGP replay trades stored as the demo accounts' Kalshi paper positions (`RACINGLINES_SPORT_PAPER=1`, labelled in-sample); buy-all never shown	Shipped in #110 and run on the VM that evening; #97 draft left open

Copilot's intermediate reports. The NASCAR parity write-up (NASCAR wrap-up (reports/2026-09-29-nascar-wrap-up/report.md, old private repo), #79) and the new-sport template (MotoGP first model (reports/2026-09-29-motogp-first-model/report.md, old private repo)) are Copilot's part, confirmed by the owner. Their figures agree with this report's where they overlap: NASCAR 14,700 links on the Mac (Kalshi 13,599, Polymarket 1,101), 7,044 fully identified outcomes, 33 NASCAR tests passing; MotoGP race-win log loss 0.0796; IndyCar blocked on its terms. Their 129 NASCAR events are the Mac database; the VM's 408 (2016-2026) came from the overnight fix, so the two counts describe different databases, not a conflict. Copilot's reports stop before any backtest, so every P&L figure here comes from the overnight run and PR #85.

The overnight VM run: what failed and why¶

The run took four starts, and every stop was a data-readiness gap, not a model bug. The root cause of the first and biggest one: the dry run checked row counts, not whether each sport had results and identified links on the VM, and NASCAR's ingest and linking had only ever been run on the Mac.

Time (UTC, 30 Sep)	Event	Cause or result	Fix
~07:00	Dry run passed; F1 golden gate +1,242.88 (target +1,243 ±400)	VM on n2-standard-2 (e2-standard-4 unavailable)	#94
to 09:25	`rl-overnight` (all): F1 broad 104/106, top 10 at 16k 24/24	Done	—
09:25	Failed at NASCAR Kalshi tape pull	13,599 NASCAR links, 0 with a race: no NASCAR results on the VM	Owner's manual fetch, ingest, sync, link (backup first); #98 preflight
12:59	`rl-overnight-sports` starts (backup 115 MB)	NASCAR 7,544 links with a race; MotoGP 202 events, 0 links with a race	—
12:59-14:45	NASCAR links, tape, replay (106 min), saves (1 min), Kalshi grid 16/16 (3 min)	Done	—
15:02	Failed at MotoGP results and link load (11 min in)	The ingest stopped the run; traceback in the VM log, not read	#100: newest seasons first, a failed load no longer stops the run
15:23	`rl-overnight-rest`: demo taker story (backup 118 MB), gate passed, record rebuilt; MotoGP tape and saves	Done	—
after 15:23	MotoGP grid ranking crashed, `KeyError 'net'`	2025 has no MotoGP race with Kalshi markets	#102
to ~16:00	`rl-finish`: MotoGP rank, OG.com buy-all (F1, NASCAR), NASCAR champion forecast; FINISH-DONE	`replay-grid/motogp/grid.md` not seen yet	Re-run ranking after deploy
~16:05	Deploy of #99, #101, #102: smoke FAIL 000 on /markets as maker; site up	#101 scanned the Parquet archive inside the request (~30 s)	#103, merged, not yet deployed

Nothing touched trading flags, migrations or vm.sh public on. Backups: racinglines-before-overnight-replay-20260930T092216Z (97 MB), …T125934Z (115 MB), racinglines-before-demo-taker-story-20260930T152320Z (118 MB), racinglines-before-nascar-ingest-link-*, all in data/backups/db/ on the VM.

Backtests: results and labels¶

One backtest passed its held-out season (F1 on Polymarket, where the incumbent held); one failed it (NASCAR on Kalshi); MotoGP has no ranking yet. All P&L is simulated at the default stake ($250 per unit of edge, $50 a market) unless stated.

F1 × Polymarket: 106-job sweep, top 10 at 16k (VM)¶

Target 2026, held out 2025, noise floor ±352. A setting replaces A only if it beats A by more than the floor in both seasons; none did.

Setting (on A's settings)	2026	vs A	2025	vs A
A's settings (row labelled "Profile A")	+2,607	—	+503	—
`reset_weight=0.5`	+3,007	+400	+525	+22
gbm, half-life 90, min edge 0.08	+3,012	+405	+16	−487
half-life 90, min edge 0.08	+2,983	+376	+284	−219
24 h volume floor $25	+2,840	+233	+725	+222

Every top row leans on the British Grand Prix: without it each 2026 total drops by 1,000 to 1,500. Decision-log entry 2026-09-30: profile A stays, no default changes.

NASCAR × Kalshi: taker settings grid (VM)¶

Ranked by the worse season. 2026 has 32 races with Kalshi markets, 2025 only 8.

Min edge	24 h floor	Sims	2026 (races up)	2025 (races up)
0.15	$50	16,000	+4,226 (19/32)	−242 (2/8)
0.10	$50	16,000	+5,419 (21/32)	−249 (2/8)
0.10	$50	4,000	+5,479 (21/32)	−278 (2/8)
0.08	$50	4,000	+5,203 (21/32)	−317 (1/8)
0.05 (default)	$50	4,000	+5,091 (21/32)	−360 (1/8)
0.10	$200	4,000	+2,721 (17/32)	−381 (1/8)

Label: 2026-specific, not robust. Prices are Kalshi last trades, which the F1 check found spiky, and the NASCAR spike-share line was not in the pasted output. Nothing is tuned; defaults stay (min edge 0.05, $50 floor).

MotoGP × Kalshi (VM)¶

Tape pulled, replay saves stored, 16 of 16 grid runs finished. No ranking: it crashed on 2025, which has no MotoGP races with Kalshi markets, so there is no held-out season at all. A Mac spot check before the run (2026, $0 volume floor, so not tradeable) made +413.62 on 93 trades.

Demo taker paper record, F1 (VM)¶

Walk-forward: defaults for 2025 rounds 1-8, then the maker's rule picks A's core settings (TW2). The gate passed, and f1 demo-history --reset --user taker rebuilt the record: 2025 +450.46 over 23 weekends, 2026 +353.01 over 15. Last five 2026 weekends: Hungary −5.44, Netherlands −29.54, Italy +22.21, Spain −5.57, Azerbaijan −93.55. Label: paper, and the rule saw 2025 before choosing, so only 2026 is out of sample.

What is not a backtest¶

Buy-all (OG.com F1 and NASCAR): proves every market is found, priced and settled. A pair loses exactly its fees by design. Per the owner, it never appears in reporting or the web app, and its output has not been read.
NASCAR champion forecast (#93): Larson 94.6% on the first Mac run vs 40-55% on the market, before rev 2's fixes; labelled NOT CALIBRATED.
Champion replays (#95, #96): built, not run.

The profile A gap: likely a strategy mix-up, not a data problem¶

The overnight report's "Profile A" row (+2,607 in 2026, +503 in 2025) is most likely A's settings run with the stage-aware early taker, not A's update taker, so it should not be compared with PR #85's +1,389 and +1,432. This is inferred from the files, not re-run.

The evidence:

The row's candidate id is early-e726f828b96a. Search candidates are named by strategy (update-, hold-, last-, early-), and every sweep job runs all four.
PR #85 measured the same thing at 16k as T8, "early, A's settings": +2,707 and +546. The VM's row is −100 and −43 from that, inside the ±700 noise floor #85 set for early.
A's update taker reproduced on the VM: the dry run's golden gate read +1,242.88 against the published +1,243 at 4k.

Run	Strategy	2026	2025
PR #85, T1 = profile A, 16k	update	+1,389	+1,432
PR #85, T8, 16k	early	+2,707	+546
Overnight VM, "Profile A" row, 16k	early (inferred from the id)	+2,607	+503
Overnight VM dry run, 4k	update	+1,242.88	not run

To close it: read the update- rows for A's settings in data/runs/search/overnight-vm/report.md on the VM. If they sit near +1,389 and +1,432, the overnight report's F1 table and decision-log entry should be relabelled as the early family. The verdict (A stays) does not change either way.

Still unexplained: the demo taker's paper record (+450.46 and +353.01) is well below #85's walk-forward story (−129 + 870 = +741 in 2025, +1,099 in 2026). The paper record goes through the signal engine's paper sizing and settlement, not the search replay, so the two may not be comparable. That has not been checked.

What this does and doesn't show (morning)¶

Shows: the new-sport path works end to end on production (probe, results, identity, links, tape, replay, saves) for NASCAR and MotoGP; F1's taker choice survived a 106-job sweep; the demo taker has a full paper record in the app; the board's data counts now reflect the exchanges' real stored history.

Doesn't show:

A tradeable edge anywhere outside F1 on Polymarket. NASCAR's profit rests on 2026 alone (2025's 8 races lose in every setting), on last-trade prices a taker may not have been able to pay.
Anything about Kalshi takers at real prices: that needs Kalshi's bid/ask candles, which today's sync does not store (re-pull from a machine with Kalshi access).
MotoGP's grid verdict, OG.com buy-all output, and the NASCAR spike share: all were written on the VM and not yet read.
The VM's exchange data counts after the overnight pulls (#101 numbers are from the cloud copy).

Gaps, and why they stay gaps:

Sport or cell	Blocker
MotoGP × OG.com	OG.com lists no MotoGP (probed 29 Sep)
NASCAR, MotoGP × Polymarket	Polymarket tag slugs unverified (#51 not run live); MotoGP champion needs sprint points, schedule, standings and classifiers (`docs/championship-markets.md`), which need a Mac probe first
IndyCar	Results source's terms forbid automated scraping (#75); season over until March 2027
Road cycling, SailGP, Le Mans	Tape-only schemas built (#50), not run on the VM; no results source or model
Downhill	No venue lists it; private book only

Open owner decisions and next steps (morning; section 7 has the current list)¶

The first three steps are the owner's and unblock everything else; none needs a trading flag or a migration.

Redeploy main so #103 clears the /markets smoke failure, then check vm.sh status.
Re-run the MotoGP ranking and read the three unread outputs (MotoGP grid.md, OG.com buy-all CSVs, NASCAR spike line), plus the update- rows that settle the profile A gap.
Look at the taker's record in the app (Signals page, nav P&L).
Mac spot check, then merge #95 and #96 in that order (handoff handoffs/2026-10-01-season-replays.md in the project files), before round 16's book opens Thu 1 Oct 20:30 PDT.
97 paper portfolio for NASCAR and MotoGP: Mac spot check and backup first (handoff handoffs/2026-09-30-sport-paper-portfolio.md).¶

Decisions waiting on the owner:

[ ] Turn on RACINGLINES_OG_VENUE=1 on the VM so the OG.com column shows in the app.
[ ] Replay race-spread setting (2.0 is tighter than NASCAR's measured ~7 places).
[ ] Kalshi bid/ask re-pull, needed before any Kalshi taker result counts.
[ ] Polymarket schema migration; freeze Kalshi maker K.
[ ] Recording timer and price-column migration on the VM (each needs a backup and sign-off).
[ ] MotoGP champion classifiers (KXMOTOGP to champion): changes stored links, needs a backup and link --apply.
[ ] NASCAR 2027 timing (recommended: decide in November, after the 8 Nov finale is recorded).
[ ] Close #58 on GitHub; empty the Mac Trash of the old MotoGP Parquet folder.