August: what shipped

[pip-robot]

The August roundup, dated 1 September and covering 1 to 28 August. Two days early against the plan on marketing#19, which said Friday, because the alternative was writing it on launch eve.

What you actually have to do: merge it on Tuesday 1 September. If anything ships between the 29th and the 31st there will be one top-up commit on this branch to re-read, and nothing else.

What is in it

Ten sections. Four of them were written and signed off across August as the features shipped and are here close to unchanged: the odds calculator, drills, the tournament card, and text size. Six are new this morning: the short version, /learn, the flop-set answer page, the bots measurement, the corrections page, and the contributor section.

The month has a spine and it is not a feature list. Half of what Pip shipped in August was Pip checking its own work, so the roundup ends on the two sections about that rather than opening on them.

Every figure in it is imported

A post about deriving numbers rather than typing them does not get to type its own. So:

• the flop counts come from config/flopSet.ts, the same registry the answer page renders
• SAMPLE_TARGET comes from lib/poker/oddsQuote.ts
• 990 is computed as C(45,2) through the module's own combinations(), not written out, because 990 is the point of the paragraph it is in
• the number of corrections is read off CORRECTIONS.length, so the sentence does not go stale the next time a row lands. It renders as seven today and will say eight on its own if #92 merges first

What it deliberately does not say

• No membership, in any form. That means the second drill kind is not in the post. It shipped on the 25th and it cannot be described without naming a thing we do not talk about yet, so it is left out rather than described badly. The drills section says the kind's name instead of counting the kinds, which is true whatever is registered.
• Nothing about improvement over time, no progress tracking, no graph, and no screenshot of the recap or the Drills section, because nobody has looked at either on a device.
• No /drills link. Drills are app: playpip.io/game/drills.
• "Exact" appears once, attached to heads-up with the board complete, which is the only shape that earns it.
• The corrections section does not mention that the corrections page corrected itself, because that lands in The account row is fixed, and the post's own headline was wrong when it published #92 and The account row is fixed, and the post's own headline was wrong when it published #92 is not merged. If it merges before the 1st, that is the natural top-up commit and it is the best line available in the section.

The one claim in it that is new rather than repeated

The bots section says the ladder gets harder as you climb and puts a number on it: entries-raised runs 32% at Friends' Garage to 60% at The Main Event. That comes off the measurement in tests/ai.test.ts, which also records that the climb is not monotone rung by rung, and the post says so. It is deliberately the weaker sentence, because the strong one has been wrong once already.

Gate

pnpm test:all green, 503 tests, pnpm build clean, biome-formatted, no em-dashes. gen-llms picks it up as the 22nd mirror.

[cloudflare-workers-and-pages]

Deploying pip-web with    Cloudflare Pages

Latest commit:	e018438
Status:	✅  Deploy successful!
Preview URL:	https://0d95cf12.pip-web-9oj.pages.dev
Branch Preview URL:	https://post-august-roundup.pip-web-9oj.pages.dev

View logs

[pip-robot]

Topped up for the end of the month, and rebased on `main` so it carries v1.18.1.

This was written on the 26th and merges on the 1st. Five PRs landed in between and two of them are things this post is about, so leaving it as-is would have published an August roundup that stops on the 26th.

Added: "Then it happened again, one street later". #98 found that every postflop gate was an absolute equity number written for a heads-up pot, so betting collapsed as the field grew. It belongs in this post because it is the same failure as the preflop one three weeks earlier: the bands shipped in that fix measure how wide the bots play before the flop, nothing measured what they did after it, and the tests written to catch exactly this were green throughout.

Added: "The library that decides who wins". The pokersolver post published on the 28th and was not in the record. It is also the third of the month's pieces about our own work rather than the product, so the short version now says three rather than two.

What I deliberately did not print, and why

No before-and-after lead rates. I went to quote them and could not make them reproduce. The repository carries three different sets of figures for the same measurement:

Source	heads-up	vs 2	vs 3
src/lib/poker/ai/policy.ts comment (pre-fix, shipped profiles)	14%	8%	4%
tests/ai.test.ts comment (pre-fix, synthetic loose)	21%	7%	7%
tests/ai.test.ts comment (post-fix, synthetic loose)	-	16%	23%

Re-running the test's own measureLeadByField on its own profile at 200 hands gives 17.0% / 14.3% / 18.5%, and Friends' Garage gives 11.9% / 12.4% / 17.1%. The two comments are measuring different profiles so they are not contradictory, but nothing on the page would have told a reader that.

More to the point, the cells that carry the finding are small: n=108 at three opponents on the synthetic profile, n=76 at Friends' Garage, n=46 at Micro Ring. strategy/positioning.md says a count and never a percentage when the sample is small, and three-opponent flops are the rarest spot in any run long enough to be worth quoting heads-up. The collapse was large enough to see through that noise. The exact percentages are not.

So the section describes the shape and the arithmetic, both of which are pinned by tests, and prints no rate at all. Raised separately with the tech side.

One thing typed rather than imported, and the file comment now says so. The 62% and 78% gates are literals inside policy.ts, not exported constants, so the prose repeats them. tests/ai.test.ts asserts the arithmetic tying them to the multiples the code uses, so moving one fails the build. Asked the tech side to export them.

Gate green: 556 tests, biome clean, knip clean, pnpm build clean.

Still: merge on Tuesday 1 September, which is its date.

[pip-robot]

Re-ran the gate on 75e8534 against today's main: 556 tests passed, build clean, 23 markdown mirrors. The body says 503 and 22 because main has moved under it, not because anything here changed.

Safe to merge as it stands. The 62%/78% swap now that #106 exports them is deliberately not stacked onto this: it is dated 1 September and should not wait on another PR.

[pip-robot]

Re-dated to 2 September. It did not merge on the 1st and a roundup that is wrong about its own date is the one error this post is least allowed to make.

Gate on e018438: format, types, lint, 556 tests, knip all green. pnpm audit fails, and it fails on main too: two high advisories in browserslist, a build-time transitive of next and shadcn, patched in 4.28.7. Nothing to do with this PR. Raised with the tech side.

Every day this sits costs one more re-date. Merging it today ends that.

[pip-robot]

The red gate here is not this branch. pnpm audit --audit-level=high picked up two new browserslist highs today, with no change in the repo, and the audit step fails for anything that re-runs. #108 pins it. Merge that first and this goes green on the next run.