Release notes

What changed in DataBaller, version by version, and why it changed.

v0.8 2026-10-09

Receipts — Every game's pre-game forecast is saved before the start and scored against the closing line, on a public page: how often our pick won against the line's, what $100 on every pick would have returned, our confidence against the line's, and the same for Scoring Shift and Sustainability across all three leagues, updated hourly with a next version shown on trial beside the current one. A model you can check is worth more than one that sounds right, and until now nothing recorded what a reader was told before a game.
A page per decision model — The decision-metrics section has a page for each model: which leagues it is live in, its test record, what is on trial and what comes next, with an "Ask DataBaller about this" link, and the knowledge base carries an evidence article for every model so a question about how one was built gets its full record. The tests, the market comparison and the versions thrown out lived in registry fields the agent could no longer see.
Follow-up questions — Three one-click questions sit under every answer — a comparison, a deeper look, a next step — and at the foot of published stories. An answer ended with nowhere obvious to go next, and a story page is where a stranger most needs to be shown what to ask.
Who is actually playing — An MLB matchup is priced with the two starting pitchers and names them, and the NBA rating is re-summed against the injury report at the moment you ask, naming who is out when that moves it. Starters are the largest single thing an MLB line prices and the rating could not see them; a two-point gap between starters moves the win probability about eight points.
NBA category leagues — Ask for a value board in your league's own categories and get every draftable player scored against the pool, with ratio categories weighed by volume so a low-usage guard cannot top assist-to-turnover on six minutes a night. The agent was rebuilding this board in SQL for every question and getting the ratios wrong. A start-or-sit question now gets a lean too; anything a sportsbook prices still gets none.
Division standing — A team's profile says where the club sits in its division and how far back, beside the league and conference rank it already carried. "2nd in the Central, 3 back" is the one number a fan quotes for a season, and the panel could not say it.
A new model and a shorter brief — Answers now come from Sonnet 5.5 under a rewritten brief: the first sentence is the answer in plain words, no first person, confidence stated once in the closing line and only for a conclusion the data cannot settle, and a substantive answer carries a chart. The first answers from the new model opened with a hedge, labelled sections with question words, and printed model versions and confidence decimals; the brief had invited most of it.
Scoring Shift held to the simple guess — Each league's scoring forecast was tested against the simplest alternative, "halfway back from this season to last": NBA scoring and MLB pitching beat it, MLB hitting and the NFL did not, so refits on what the guess leaves run on trial beside the current version on Receipts, and the switch is judged there. MLB Scoring Shift also passed its 2026 forward test in both roles. A model that merely restates reversion is not worth a page.
Misspelled names resolve — "Gunner helm" is taken to mean Gunnar Helm, and the answer says so and asks if you meant someone else. A typo used to return "we don't hold that player" and cost a free account one of its analyses.
Comparison tables mark the better side — A two-team or two-player table shows a check on the better number in each row, and knows that for turnovers the smaller number wins. A reader was left to work out, row by row, which side each figure favoured.
MLB games stop going missing — A late box score is caught the next morning, the schedule is refreshed hourly on game days, and a game the league dropped is marked cancelled. Receipts listed last night's Rays–Yankees as tonight's game with no line and left out the game being played.
Home page rebuilt — The front door opens on the ask box with three example questions, shows three real answers as screenshots, and carries a card per decision model with live figures; the sign-in page says in large type that it is free, needs no password, and keeps your conversations. A first-time visitor met a sign-in form before seeing what the product does.
Answers read like a story — An answer has no bubble: it sits left-aligned at the same inset a story uses, charts come out of the code box they were accidentally drawn in, and light mode gets a grey canvas with white surfaces. Moving from a story to an answer changed both the text's edge and its framing.
A getting-started guide — An account with no conversations sees five kinds of question to ask, four facts about what we hold, and a pointer at the box. New accounts were arriving and leaving without asking anything, and the empty list told them nothing.
Triple-A — The paid plan is called Triple-A, one step below the big leagues. "Professional" read as a sales tier.
Stories and studies cite-able by answer engines — Published stories and the study pages carry structured markup naming their teams, players and dates, each study opens on its finding, and a "How stories work" page says how an answer becomes a story. AI search engines lean on that markup to tie a page to who it covers.

v0.7 2026-09-20

Questions about games not yet played — "Who's getting exposed in Week 1", a fixture on Sunday, a team's prospects for the year: a question about what happens next is answered forward from the schedule and the power rating, and "how many wins next season" gets a range with its reasoning shown. Asked two days before the NFL opener who would get exposed in Week 1, the agent said no game had kicked off and offered last season's Week 1 instead, with the ratings and sixteen scheduled fixtures sitting unused in front of it.
The line, explained — "What are today's odds" reads the lines we hold — every book's moneyline, total and spread, and how each has moved since it opened — and sets the power rating beside them with the conversion printed, so the number can be reproduced. Six seasons of closing lines say the market is right about three times in four where it and the rating disagree, so a gap is explained as what the market prices beyond team strength, and never called an edge, a mispricing or a fault in the rating.
Playoff races, computed — A race question carries every rival's games remaining, how many are head-to-head, and the magic and elimination numbers, and a division leader has clinched nothing until its number says so. Race arithmetic done in the model's head was wrong often enough that it now arrives computed, the way games back already did.
Awards and honours — Every NBA MVP, Finals MVP, Defensive Player of the Year and Rookie of the Year back to 1952, MLB's major awards back to 2002 including Gold Gloves and Silver Sluggers, and the NFL's, on a player's profile and a team's season. Asked what Kobe Bryant meant to the Lakers in the title years, the agent could not mention his MVP or either Finals MVP because we held none. It can now, and it is told an award is a voting record rather than a measurement: "he won MVP" and "he was the best player" are two claims.
Career numbers on the career panel — A player's career view shows career averages and totals with a first-to-last span, summed by the same code that lays out one season, and says so when an early season was only partly recorded. The totals were withheld from every career because some early seasons are partial; labelling those is the right answer, not hiding the numbers from everyone.
Football gets playing time — Per-game snap counts for every NFL player from 2012 on, with snap share as a metric the agent can rank and trend. Football was the one sport with no measure of how much a player was on the field, so workload questions fell back on tackles or targets — and six tackles can come in thirty snaps or forty-five.
Basketball beyond the box score — On-court offensive, defensive and net rating, possessions, rim protection, contested shots, deflections and matchup numbers per player per game, back to 2012-13. "Is Wembanyama already the best defender in the NBA" had to name defensive rating, on/off and rim protection as unavailable; the vendor computes all of them on the tier we pay for, and nothing here read them.
The NBA power rating is built from players — Each player's per-48-minute impact on his team's margin is measured, and the team rating is the sum over the expected lineup — last game's players minus anyone the injury report marks Out, or the current roster in the offseason. A single team-season number has nowhere to put who takes the minutes, a backup's quality, a double absence or a trade, and that was where its gap to the betting line lived. Two things were tested on the way and failed their pre-registered bar, and are recorded rather than shipped: a quarterback-unit rating for the NFL, and every continuity term on top of the lineup sum.
Defense is read as categories — Basketball defense is steals, blocks, defensive rebounds and on-court margin, each with its own figure, defensive and offensive rebounds per game are rankable now, and "stocks" is quoted as the sum of two categories already shown, never as a third. A best-defender verdict rested on "both trackable defensive categories", which were blocks per game and stocks per game — one and a half metrics presented as two.
Straighter numbers underneath — A quarterback's touchdowns thrown are never counted as touchdowns scored, nine metrics that read columns no vendor fills now have a writer or say so (time of possession, sacks, caught stealing and blown saves among them), a catcher's season line shows the running game against him, and every team profile answers again after a bug had them failing since 2 September. Asked how much help Josh Allen gets, the answer summed his passing and rushing touchdowns as the share he produced himself — every passing touchdown was also a receiver's, so the share measured Buffalo's pass-versus-rush mix and no supporting cast.
Your clock, your words — Schedule start times are converted to your time zone instead of the venue's, "Kobe in 2009" means the season that ended in 2009, and "the Niners" resolves to a team. A Pacific user was handed nine start times labelled PDT and not one was right, and a nickname no vendor sends left a question with no team profile at all.
The answer opens with what changes your outlook — A role that moved — a target share that doubled, a job that changed hands — leads the answer instead of sitting four sections down, and a number from a past span stays that span's number rather than posing as the forecast. A trade answer had every figure right and the one fact that could move the valuation in a sub-bullet under a recital of model scores.
Charts that carry the finding — A chart can draw a labelled baseline and mark an event, a relationship between two measures is a scatter with the diagonal as its reference, how a whole divides is a stacked bar or a pie, and a substantive answer carries a chart whether or not you asked for one. Five teams' run differentials say which is highest; the same bars against the league average say which are good. Charts restated the table and never said where a number sat.
Stop means stop, and answers finish without you — The stop button ends the answer in the tab and on the server within seconds, and an answer keeps going if you leave the tab — switch apps on your phone, come back, and it has progressed or finished and is saved. Pressing stop used to do nothing on the server, and a question asked on a phone died the moment you switched apps, taking a two-minute answer with it.
Chat on a phone — The page scrolls like a news site so the browser's bars get out of the way, the keyboard drops after you send, and the menu and the drawer open as one motion. The thread scrolled inside a box the browser could not see, so the bars never retracted and the header stuttered.
The pages a stranger reads — A features page at /features, six study pages under /about (the rating against the betting line, what a missing star costs, keeping the core together, do touchdowns regress, record against run differential, one-run luck), every public page readable without JavaScript, and a sign-in page that says what you are signing in for and that it is free. A crawler or a reader with scripts off got an empty front door, a pricing page that priced nothing and a stories page with no stories; a study an answer cited had no page to cite.
Only a session's first answer becomes a story — A follow-up answer no longer publishes as a public story. A story built from the second answer in a conversation reads as the back half of an argument, leaning on a table it never reprints and naming people it never introduces.

v0.6 2026-09-01

"What happens next" now has an answer in all three leagues — The scoring forecast covers the NFL and MLB alongside the NBA: forward change in scrimmage yards per game for running backs and receivers, and in production per plate appearance for hitters and pitchers, each with the reasoning behind it. The forecast was basketball's alone, so a football or baseball question about where a player is headed got a read on the present and nothing about what comes next. The three models disagree with each other, which is the interesting part: NBA scoring momentum tends to persist, while NFL and MLB production above a player's own established level tends to give itself back.
Football answers whether production is real — Rushers, receivers and passers get a sustainability read built on yards per touch against the player's own multi-year level, with the excess yards a hot stretch is borrowing. "Is this breakout real" was answerable for basketball and baseball and not for football. One finding changed the model on the way: touchdown rate against a player's own history turned out to be partly skill, so the model reports a touchdown surge rather than promising it will regress.
The models that failed their tests are named — There is no quarterback scoring forecast and no baseball role-change model, and the About page says why for each. The quarterback version's evidence reversed direction between the two seasons it was tested against, and the baseball role-change question failed twice. Ask either one and the answer says the model does not exist instead of improvising a number.
Answers stop where the records do — Leaderboards and career verdicts now refuse seasons whose statistics were never fully recorded, and a partial season says how many of its games are actually covered. A per-game average computed over a season that is half-recorded is a wrong number presented with the confidence of a right one.
MLB player value arrives as real data — Wins above replacement and its components, plus the fielding block, load from the vendor, and offensive WAR gets its own honest entry rather than resolving to the total. Value questions were being answered from the pieces rather than the measure fans actually cite.

v0.5 2026-08-24

Teams get their own models — Two new team metrics cover the NBA, NFL and MLB: a power rating — each team's strength as the margin it would be expected to beat an average opponent by — and a sustainability read that measures whether a team's record matches its scoring margins, in wins. Every matchup question used to be improvised from raw standings. Now "what can we expect from Lakers vs. Sixers" is answered from two stored ratings and a measured home-court edge, and "is their record for real" has a number behind it.
Matchup forecasts with honest odds — Ask about any pairing and the answer leads with an expected margin and a win probability, built on the spot from the two teams' current ratings — nothing is precomputed, and between seasons the answer says the ratings are carried from last season. Most single games are closer to a coin flip than anyone likes to admit. When the model says 58%, the answer says 58% — not a lock of the night.
We tested the conventional wisdom, and cut what failed — The About page's Decision metrics explainer now covers both team models, including the parts we tested and threw away: roster continuity doesn't improve offseason ratings, close-game records aren't pure luck in our data, and the shooting-luck adjustment made predictions worse. These models ship only what survived testing against seasons they'd never seen. The failures are documented as publicly as the passes, because a model you can't check is just a vibe with decimals.

v0.4 2026-08-23

Every player has a read, year-round — The three DataBaller scores — role change, sustainability, and the scoring forecast — now read a player's most recent games wherever the calendar sits, straddling the offseason when they have to. Football gets its first full slate of role-change readings, and NBA answers no longer go quiet between the Finals and opening night. The models used to look only inside the current season, which meant they had the least to say exactly when "what happens next season" is the question on everyone's mind.
A page that explains the scores — About now has a Decision metrics page that walks through the three scores in plain language — what each one asks, how the number is made, and what it can and cannot tell you — with no statistics background assumed. A score you can't explain is just a vibe with decimals. These were tested against seasons of history before they shipped, and the page keeps the failures as publicly as the passes.
The scoring forecast earned its spot — The points-per-game forecast was rebuilt and re-tested against two full seasons the model had never seen before this version shipped. One refinement to the sustainability score failed its re-test and was withdrawn rather than kept. Every claim these models make is written down before it is tested, and a claim that fails is removed, not argued with. What you see is the version that passed.
Answers that answer the question — Ask about a scoring jump and the answer now leads with the scoring jump — a name, a number, a direction — with the model detail underneath instead of in place of it. An answer that is technically thorough but never gets to the point isn't an answer. Brilliant and approachable are both the job.

v0.3 2026-08-21

Plans and pricing — There is a plans page now with two cards on it — Free, and Professional at $24 a month for full access across the NFL, NBA and MLB — reachable from the front door and the account menu. A paid plan is fair use rather than an allowance: no counter, no quota, and nothing that stops you mid-question. Free stays at five analyses a week with a meter on the account page.
Fantasy questions get real answers — Projections, draft-board rankings and DraftKings salaries and slates are loaded for all three leagues, and the agent has tools that read them directly: what a player is projected for, and who is ranked where. It used to tell you flatly that we had no fantasy data. That stopped being true, and it was still saying it.
Three modes, and a bar that shows them — Analysis, Stories and Support are now three named places with a switcher across the top of every page, the lockup and version beside it, and the whole bar slides out of the way as you read down a conversation. Getting anywhere used to mean opening a "⋯" menu in the quietest corner of the sidebar, which can't tell you where you can go without being opened first.
Your question survives the sign-up — The front door now carries the real chat box under "What do you want to find out?", with example questions that fill it; type one and it is quoted back to you on the sign-in card and asked as your first question once your account is set up. A free account still means giving us an email address; what changed is that the question you arrived with is kept through the whole detour instead of being lost on the way, and the front page shows you the actual chat box rather than four bullet points about it.
Is it real, or is he running hot — The agent can now answer with a model rather than an impression: whether a scoring run is supported by the process behind it, and whether a player's underlying game has actually changed, each stated as a model's output with its version and what would prove it wrong, recomputed the morning after every game night. "Is this real" is the most-asked question in sports and the one a box score answers worst. An answer built on a model should say that it is, rather than passing as a statistic.
Rosters that know who arrived — A team page shows the players who came in, not only the ones who left, and says which club each came from; a team's standing now names the conference it is a standing within. Asked why writers favoured the Nuggets over the Lakers, the agent saw six departures and none of the arrivals, and answered as though five rotation players had been replaced by nobody.
Long conversations stop losing their tail — A conversation big enough to page through storage now reloads whole instead of quietly dropping its newest turns, and an author's stories page no longer breaks past fifty stories. The dropped turns were silent: the count matched the shortened list, so nothing looked wrong.
A faster first answer — The first turn of a session no longer waits on a cold start behind the scenes, and two waits came off the path every turn travels. The opening question of a session was taking a median five seconds before a single word appeared, which is the worst possible moment to be slow.
An About page — /about says what DataBaller is, who it is for, the five-step method behind an answer, and the principles it holds to, replacing the FAQ and the old docs page. The only written account of what this product is lived in an internal strategy document nobody outside the repo reads.
Seasons named, premises checked — An answer says which season a stat line is from, and the agent looks a season up before telling you your premise is wrong. Out of season a number with no year on it is ambiguous at best, and being corrected by something that hadn't checked is worse than not being corrected.
Better retrieval behind the explanations — Every reference document now carries the kind of document it is, so the agent can search by kind again, and the 146 metric and concept files gained the analytical phrasings people actually type; floor, ceiling and lineup correlation joined the corpus for the fantasy work. Analytical questions — "which players produce more value than their box score suggests" — matched almost nothing, because the whole corpus was phrased as what/how/why and the questions are which/who.

v0.2 2026-08-15

Answers become public stories — Standout answers now get published as public pages at their own URL, with a browsable grid at /stories filtered by sport and tag, related-story links between them, and a Published stories view where you can see the views and likes yours drew. Nothing carries a name or byline, an editorial pass decides what runs, and a takedown removes the page.
Charts inside an answer — An answer that shows a trend or compares teams now draws a line or bar chart in line with the text, with a table view underneath it and a button that copies the chart as an image. A trend you used to have to infer from a column of numbers is now visible at a glance, and pasteable straight into a group chat.
Ranking questions read a real leaderboard — "Who leads the league in rebounds" now reads a board the database already ranked, instead of the model working the order out from rows it printed. Answers occasionally called someone third when their own numbers put them second; the rank is now part of the data rather than something derived along the way.
Is his hot streak real — You can ask about a player's recent stretch and get it measured three ways at once: against his own earlier form, against his full season, and against every comparable player over the same dates, with how much he actually played alongside it. A hot fortnight and a genuine change now look different from each other instead of both reading as "he's up".
Finding what nobody asked about — The agent can scan a whole league over a window for what deviates — teams outperforming their record, players far off their own form — and when the data in front of it holds something striking you didn't ask about, it now says so in one line. "I hadn't noticed that" is the reaction this product is built for, and it needed a way to happen unprompted.
When a team plays next — "When do the Chiefs play next" now has a real answer, including in the off-season, and it says whether those games are preseason rather than calling one the opener. It's one of the most obvious questions you can ask, and the answer used to be that there was no current-season data.
The leagues you follow — The account page lets you name the leagues you follow, and a question that doesn't say which sport it's about is now read against those leagues first, falling back to whatever's in season. Asking "who's hot right now" in August used to mean baseball whether or not you follow baseball.
Rate an answer, and say why — Thumbs up and thumbs down sit under every answer, with a comment box on the thumbs down, and you can change a vote you've already cast. Telling us an answer was wrong was previously impossible, and a vote you couldn't take back isn't feedback, it's a trap.
Straighter numbers underneath — Games played now counts games a player appeared in rather than nights on the roster, NFL playoff and preseason box scores are collected instead of being reported as missing, spring training no longer counts toward the regular season, and an injury is dated from when it was first reported rather than when the feed last rewrote it. One answer told a reader a player had appeared in all 83 of his team's games when he had played 6. That number, and several like it, came from the data rather than the writing.
Saying what it doesn't know — When a metric isn't available, a game log stops at a limit, or a player has no season on file, the agent now names the gap and answers around it rather than guessing at a cause, declining the question, or handing you a menu of things it could do instead. It also won't state a number it privately doubted a moment earlier.
Football profiles read like football — An NFL player's page now lays a season out in passing, rushing, receiving, defense, kicking, punting and returns instead of one long list in database order, and shows the games he played. A kicker's line and a quarterback's line used to arrive in the same thirty-odd columns, which made both unreadable.
Every name opens — Players named anywhere in an answer are clickable now, not only the ones inside a ranking, and the profile panel opens on a phone as well as a laptop. A signed-out reader gets the panel too, with only the profile itself held back.
Nine new concepts the agent can draw on — Reference material went live for Pythagorean expectation, Game Score, whether a hot start is sustainable, how to read a league-wide scan, availability-adjusted value, and why quarterback wins is a weak stat, among others. The agent was already computing some of these and had nothing to explain them with.
A front door that says what this is — The site now leads with "Sports intelligence for the everyday baller", and a DataBaller link unfurls in a group chat with a real card instead of a grey wordmark. Signing in with Google or Apple from a shared conversation link also lands you on that conversation now, not a blank new chat.

v0.1 2026-08-08

Version numbers and a release-notes feed — A small version number now sits next to the DataBaller name on the home screen and in the sidebar, linking to a feed of what changed in each release. Every update used to be invisible unless you happened to notice something new yourself. Now you can always see what's new and why it was worth shipping.

Something you want to see in the next one? Ask for it on Discord.