From a session to what to work on.
A number tells you what a swing did. A diagnosis tells you why it keeps doing it, and what to change first. This is how PinPoint Studio gets from one to the other: the model it reasons with, how evidence builds swing by swing through a session, how it picks what you should work on, and the rules that decide what it is entitled to call true.
Session diagnostics shipped with the fourth beta in October 2026, and the sixth (1 November) moved what to work on onto the golfer's Home screen and gave the report its posture, wrist and club delivery pages. Everything on this page runs in the app today: the ledger, the patterns, the chains, the likely driver, the work-on list and the session report. The content behind it is another matter. Most of the normal ranges are starting figures, most causal links are coaching consensus rather than tested science, and 88 of the 186 conditions are still marked draft. The page says which is which as it goes. See how we earn the numbers →
A coach's knowledge, written down
Before a single ball is struck, the app already holds a model of the golf swing: what can be measured, what normal looks like, the named habits a coach would recognise, and how they cause one another. None of it is compiled into the program. It is content, stored as reviewable files a coach can read, argue with and edit in the app, and it is deliberately written ahead of what the cameras can see today. A complete model is what tells us which measurement to build next.
Seven words, one chain
Every finding the app ever makes travels the same chain, and each link has a name. You need the names to follow the rest of this page, so here they are once, in order.
- 1
A metric, then a measure
A metric is something the analysis extracts: lead wrist angle, pelvis sway, ball position. Most change continuously through the swing, so before anything can be judged you have to say which reading you mean. That decision is a measure: the value at impact, the change from address to the top, the peak between two positions. "Lead wrist at impact" and "how far the lead wrist moved from address to impact" are two measures, with two different ideas of normal. The library holds 208 of them.
- 2
A norm, then a grade
A norm says what normal looks like for one measure on one kind of shot: a centre, and a tolerance either side. Held against it, a reading gets a grade: Ideal, Good, Watch or Action, or Not measured when there was nothing to grade.
- 3
A signal, then a condition
A signal is a small rule that watches one side of one measure: "tell me when ball position is unusually far back". When it fires, its condition (a named habit in a coach's words, such as Ball too far back) becomes a finding on that swing. There are 170 signals and 186 conditions.
- 4
Causes
Conditions are joined by 639 authored links: what tends to produce what. That graph is what turns a list of complaints into an explanation, because several things you can see often trace back to one thing worth working on.
Ball position on a 7 iron, from the number to the habit and what it tends to cause.
What normal looks like, and for whom
A norm is drawn on screen as a corridor, a band most golfers' readings fall inside. It describes a population, never a person: it is "what golfers' readings look like", not your target, and nothing in it is keyed to you.
- 1
Two tolerances, not one
Each side of the centre has its own tolerance, because the two directions are rarely equally interesting. A ball a little forward is common and harmless; a ball a long way back is neither. A reading is graded by how many tolerances out it sits on its own side: inside one is Ideal, inside two is Good, inside three is Watch, and beyond that is Action. Those multiples are one policy for the whole library (Lenient, Standard or Strict), never a slider per corridor, so Ideal means the same thing everywhere. A habit that can only go one way is judged only on that side: a ball well forward of a line that watches for too far back scores nothing, rather than counting as far past a fault line.
- 2
A signal fires on a deviation
Not on merely leaving Ideal. Ideal is the middle of normal, so roughly a third of perfectly ordinary golfers sit outside it on any one measurement, and an app that spoke up every time would flag something on every swing. A signal fires at Watch or Action, past two tolerances: about one swing in 44 from a normal population, on the side it watches.
- 3
The kind of shot decides which norm
A driver is played off the lead heel and a wedge from the middle of the stance, so one "normal ball position" would be wrong for both. Norms hang on a tree of 31 shot contexts: any shot at the root, then full swing, the clubs, partial shots, bunker. The app starts at the shot's own context and walks upward until it finds a norm, so only what genuinely differs has to be written. The same tree carries styles: a bowed or cupped lead wrist at the top, which are both how tour players swing, and the two posture patterns good players use, staying level or squatting and pushing. Sometimes the fairer move is to take the club out of the reading instead. How far the ball sits from the body is now measured as how far past the toes the hands reach once the club's own length is allowed for, so one band serves wedge to driver; and the swing plane is graded as an angle to the ground, back and down, with a norm for each kind of club.
- 4
Context informs; it never silences
The app will not decide that a habit doesn't matter on your bunker shot and quietly skip it. It measures, grades and reports. A norm can be narrowed to a cohort (a segmented population, say men aged 55–64), which moves the corridor but can never flip what counts as good. 271 norms ship today.
Illustrative: an asymmetric corridor under the Standard policy.
Not everything is a swing fault
The library names 186 conditions, and it keeps them apart by what sort of thing each one is, because a coach does. Getting this wrong is how a mobility screen ends up being called your swing fault.
- 1
Seven kinds
A fault is a movement error, what a lesson is about (96). A setup position is fixed before the swing starts (25). A delivery number is what the club did at the ball: path, face, attack, low point, the downswing plane (17). Those are the scoreboard of the downswing rather than a fault in themselves. An outcome is what the ball did: slice, thin, shank (20). A capacity is what your body can and cannot do, settled by a physical test, never a camera (17). An intent is what you are trying to do, which only you can confirm (10). And there is one piece of equipment.
- 2
Seen, or only inferred
158 conditions can be seen in a swing. 28 are latent: a tight hip or a habit of mind has no signal of its own, and can only be inferred from what it explains. Separately, each condition records how it is established today: measured by a signal (157), settled by a physical screen (17), or knowable only by asking (12). The first two describe the movement; the last one is allowed to change, and does, every time a new measurement lands.
- 3
Written for the golfer
Every condition carries its consequence in plain words, and every one that can reach your work-on list must also carry a goal written as an instruction, one feel, a way to check it without the studio, which still to show you, and at least one drill. A test fails the build if any is missing. Each also lists the other names coaches use for it, so flip finds Scooping and standing up finds Early extension.
- 4
How common, in our judgement
Each condition has a prominence, from rare to ubiquitous. No study has ever counted named swing faults across golfers, so this is the library's own editorial judgement, used as a starting prior and labelled as such. Seven sit at ubiquitous: over the top, early extension, casting, an out-to-in path, an open face, the slice, and trying to lift the ball.
The shipped library, counted from the pack itself.
Causes, written in words
The 639 links between conditions form one graph: 605 say one thing tends to cause another, 28 say two readings corroborate each other, and 6 say two findings cannot both describe the same swing. The library refuses to load if the causes ever form a loop.
- 1
Strength is a word, and the word is a number
Each cause link is authored as rarely, sometimes, often, usually or always, and read back in those words, never as a percentage. Underneath, each word is a probability that the cause produces the effect: 0.10, 0.30, 0.60, 0.80, 0.95. Neither end reaches 0 or 1, because no coaching claim has earned certainty. Prominence works the same way, as the prior chance a cause is present at all: 0.05 for rare up to 0.60 for ubiquitous.
- 2
A symptom is never called the root
Ranking explains the findings by greedily choosing the causes that cover the most of them. One rule outranks the arithmetic: a condition that has a cause in the library is never presented as a root. Early extension causes loss of posture, and casting causes scooping. Handing a coach the symptom and calling it the diagnosis is the failure this whole layer exists to prevent.
- 3
What didn't happen counts too
For a cause nobody could measure, the app works out how likely it is from what it would explain: its prominence as the starting point, then each of its effects. An effect that showed raises it. An effect that was measured and absent lowers it, since a cause that "usually" produces something which never appeared is less likely to be there. That negative evidence is what stops a cause with few authored effects winning by default.
- 4
An observation beats an inference
A cause the session actually measured keeps its own measured likelihood; its effects never argue it up or down. A first version let inference overrule a measured pattern, and it read zero for something plainly on screen. Causes that can only be established by asking (trying to lift the ball, say) are offered as questions, never concluded, and never count as explaining anything.
A slice of the shipped graph: four of early extension's nineteen authored causes, and four of its effects.
Several readings, one judgement
Some habits aren't one movement. "Losing your posture" is the trunk standing up, the hips moving toward the ball, the head rising and the pelvis lifting, and no one of them has to be large. So a few conditions read a composite: several graded readings combined into one.
- 1
Two ways to combine
When the parts all push one way (posture held into impact, and to the top) they are added with weights and signs, corrected for how much the parts move together, so a compensation cancels: bend lost while height drops is not counted twice. When any departure matters in any direction (posture at address) the distance from normal is combined across all parts. Either way the result is calibrated to trip a clean swing about 2.3% of the time, the same rate a single reading trips its own line.
- 2
Too few parts is a refusal, not a zero
If the parts that could be read carry too little of the composite's weight, there is no reading, and the app says why. A composite built from three of its four parts is calibrated as a three-part reading. How strongly the parts move together was measured on three golfers' swings, not assumed.
- 3
It says which part did it
Each part's share of the result is kept, so the finding reads "loss of posture, mostly the trunk" rather than a bare verdict. The specific moves keep their own conditions too. When early extension and loss of posture are both patterns, the work-on list names the specific move and treats the family as its symptom.
- 4
Against good players, position by position
Head height, hip depth, trunk bend, both knees and pelvis height are now also read at every position from address to the follow-through, against bands drawn from published measurement of tour players where it exists. Two patterns are graded separately, staying level and squatting and pushing, because both are how good players swing. Departures like standing up or an early dip are never treated as a pattern.
Illustrative numbers: how a composite reports its parts.
A probability on every verdict
Every reading now carries an error bar. A signal that fires on 12 against a line at 10 should not fire as hard when the reading is 12 ± 3 as when it is 12 ± 0.1. So beside its yes or no, every finding carries the probability that it really fired.
- 1
Two different spreads
The norm's tolerance says how much golfers vary; it places the line. The measurement error says how wrong this reading might be; it decides how sure we are which side of the line you really were. Only the second turns a verdict into a probability: P(fire) = Φ((v − e) / σ). For 12 ± 3 against 10, that is about 0.75: a one-in-four chance the true value never crossed.
- 2
A reading may also be simply wrong
Separately from its error bar, a reading carries the chance it is on the wrong thing altogether: the shaft tracker locked onto a shadow, say. That share of the shot counts as assessing nothing. Where no error bar was stated, none is invented. The verdict stays a plain yes or no, and is marked as unquantified.
- 3
The verdict itself never changes
Fired, not fired and not measurable are computed exactly as before. A gate checks that on every row of every session in the library. The probability only adds information, and feeds the session's counts and rankings.
- 4
You see words, not the number
The session re-runs its whole picture 200 times, each time letting every borderline reading fall either way by its own odds, and counts how often each pattern and each likely cause survives. That share becomes a word: firm (at least 80% of runs), likely (at least half) or fragile. No percentage for anything inferred is ever shown.
The engine's arithmetic, drawn. The app shows words, not this number.
A shot fires findings; only a session diagnoses
That one sentence divides all the labour. A single swing is an observation, and one observation is not a habit. The hard part of this layer is the restraint not to say more than fifteen swings support, written down as rules the app has to follow.
A ledger, not a running total
After every swing, every condition the capture could speak to is checked, and the result is written down: one row per condition per swing. Nothing is summarised on the way in. Every pattern, trend, link and ranking on screen is worked out afresh from those rows.
- 1
What a row holds
Whether it fired, didn't, or couldn't be assessed; the probability it really fired; which way it went; how far the driving reading sat from normal; and which club and context it was judged against. The ledger is saved beside the session, so reopening it later rebuilds exactly the same picture.
- 2
Every measurable swing counts, and only those
A badly hit swing is never thrown out; mishits are the most diagnostic swings in the session. A reading the app doesn't believe (outside its plausible range) contributes nothing for that measure, because it is a capture fault rather than a swing fault. A swing the capture couldn't judge is left out of that condition's count altogether, so "8 of 12" always means 12 swings that could have shown it.
- 3
The first three swings count half
Warm-up swings aren't representative and would seed early patterns badly, so the first three are down-weighted rather than excluded. It matters: a habit that fired on 5 of 8 swings, including all three warm-ups, drops from a pattern back to watching once they are weighted.
- 4
The panel matures in one direction
It opens cold ("too few swings to call a pattern"), becomes forming once patterns appear, and established once two of them are linked by evidence, when the cards reorganise into causal chains. It never steps back down within a session, so you never watch the app change its mind mid-thought. At the end it freezes, and becomes the session summary.
Illustrative: the ledger as the session report's swing-by-swing grid draws it.
Recurrence is the unit of diagnosis
A habit is something you keep doing. So the question for each condition is how often it fires on the swings that could show it, judged by how much a session that size can actually support.
- 1
The cautious end of the evidence
One of two and eight of ten are both "most of the time", and they are nothing like the same evidence. So the rate is graded by the Wilson score lower bound: the lowest firing rate the swings are still consistent with, at 95% confidence. A condition becomes a pattern when that bound reaches 0.30, from at least three measurable swings. Three of three clears it (0.44). Two of three doesn't (0.21), and neither does one of two (0.10).
- 2
Watching is the outlier discard
Something that fired but hasn't earned a pattern sits in watching: seen, not yet evidence, and it never headlines. A single shank stays there forever if it never recurs. The outliers are set aside by weight of evidence, and no data is deleted to do it.
- 3
Which way, and is it consistent
Among a pattern's firings the app tracks direction. If fewer than 70% agree, the pattern is presented as inconsistency ("an inconsistent shaft plane"), which is itself something a coach wants to know, and the directional story is withheld.
- 4
Getting better, and fixed
Once a pattern has five readings, its trend is a Theil–Sen slope with a rank test of significance, immune to one wild swing by construction, reported as improving, stable or worsening and nothing finer. A pattern that hasn't fired in the last five measurable swings is shown as resolving: the panel rewards the fix rather than punishing the history.
Wilson lower bounds at 95%, computed, not illustrative.
Chains are confirmed, never discovered
186 conditions make about seventeen thousand possible pairs. Search those for links across fifteen swings and you will find nothing but coincidence. So the session never searches. It only tests the few links the library already wrote down between this session's patterns: a handful of hypotheses fixed before the first ball. The causal graph is the multiple-comparisons correction.
- 1
Four grades of link
Present together: both ends are patterns. Necessary, nowhere near sufficient. Coherent: the directions match the link, and within single swings the two fire in the swing's own order. Conditionally dependent: on at least eight paired swings, the effect shows more often when the cause did, by a one-sided exact test. That reads as a coach's sentence: "path was out-to-in on 7 of the 8 swings where the hips stalled, and on none of the 3 where they cleared". Moved together: the cause improved, and the effect followed.
- 2
Limits are states, not secrets
A cause that fires on every swing has nothing to compare against, so its links stop at Coherent, shown as "present on every swing, coupling untestable today". Two conditions read from the same measurement can't be tested against each other either, and say so. A step in the chain the cameras can't yet measure is drawn as a ghost link, never silently bridged.
- 3
No chain probability
The links aren't independent, so multiplying their strengths would be precision nobody could defend. A chain's confidence is simply its grade at each link. When two authored causes could explain the same thing and the session can't tell them apart, the more consistent chain is shown with the rival one tap away. Neither is ever claimed to be the only explanation.
Illustrative: link grades as the chain rail draws them.
The likely driver
Once the session's patterns have settled, the app asks which upstream cause would explain the most of them, and puts that at the foot of the panel. "Fix the hip clearance and the path fix comes free" is the whole value of a causal model over a list of faults.
- 1
Ranked over patterns, not swings
The explanation runs over the session's patterns, and re-runs only when their membership changes; ranking per swing would reshuffle on every ball. It waits until the pattern set has held still for three swings. An empty footer is better than a flickering one. When a session's patterns share no authored cause, it says exactly that.
- 2
It says how firm, and how it knows
The driver carries how many of your patterns it would explain, and its stability word: firm, likely or fragile. A driver nothing in the session measured, and no screen has confirmed, is labelled "inferred, not measured" under its name. That label was chosen over demoting such causes, because an unmeasured cause is often exactly what is worth checking next.
- 3
A thirty-second test instead of more kit
Many of the strongest causes in the library are physical (a tight trail hip, a stiff upper back) and no camera will ever see them. For those the app names a screen, one of 17 physical tests with a protocol and a pass criterion: "a hip test would anchor this chain; it would explain three of your patterns". Enter the result and it counts straight away, as the only ground truth the model has about a hidden cause.
- 4
Questions, kept as questions
Causes only you can confirm, such as trying to help the ball into the air, appear as questions, visually distinct, and are never counted as having explained anything. The report doesn't model intent and doesn't excuse anything: a habit can be deliberate, and the app reports what it saw either way.
Illustrative: the driver footer, with a screen recommended.
What you tell it, and what it never changes
Two optional declarations turn watching into practice. Neither one ever changes what is detected or how it is graded. Intent shapes what the app pays attention to and how the session is structured, and never the evidence.
- 1
The miss you came in with
Declare your usual bad shot at the start, as you would to a coach, and the app walks upstream from it through the graph, so the chains that could produce it are ready to be tested. When a launch monitor is connected, it checks your miss against what the ball actually did.
- 2
A focus turns the session into an experiment
A session can carry a focus from the start: setup's first step offers an open assessment, your coaching focus or one of your active work-ons, and Practise on the Home screen opens it ready chosen. Or say "working on this" about one habit mid-session, and every swing before becomes the baseline and every swing after the intervention: an experiment with a sample size of one, which is what practice is. The panel then reports the habit you're working on and what happens downstream of it ("hip clearance improving; path following"). Only under a focus can a link honestly earn moved together.
- 3
Feedback, not after every ball
Motor-learning research is consistent that feedback after every attempt can hurt retention compared with feedback that comes less often. So by default the panel surfaces a swing only when it fired one of the session's patterns, or broke a clean run; otherwise it acknowledges quietly. "Every swing" is there for a coach-led assessment. The evidence engine runs at full rate either way. This setting only changes what is shown.
- 4
No session score
Deliberately. A single number invites you to swing at the panel, and nobody can answer "why 74?". The chain is the headline. The ten-second glance between balls has one job: did that swing add a tick or a cross to my things?
Illustrative: five swings either side of a declared focus.
Three to five things, in order
Every session records its work-ons: up to five habits you keep producing and can practise your way out of. "Work-on 1" means the same habit on the panel, on the golfer's Home screen, in the charts and in the report, because all of them read one ranking function.
- 1
What may be listed
Only patterns, so never something merely watched. Only faults and setup positions, because outcomes, physical limits and intentions aren't things you drill. Delivery numbers may only fill a thin list up to three. A habit whose reading isn't trusted enough to headline is held off every list. The face-on camera can't yet tell a folding lead arm from a normal release, so chicken wing is still measured but never listed.
- 2
The order, rule by rule
Each rule is consulted only when every earlier one ties. Still happening before resolving. The root before what it causes, but only where this session's own swings support the link. Sharp findings first, such as the sequence out of order or the hips peaking after impact. Then, when the session couldn't test any link (the usual case, because a habitual golfer shows every habit on every swing), earlier in the swing first, setup before backswing before impact, which is how coaches sequence work. Last, the bigger miss. Every entry records which rule placed it, so the report can explain the order.
- 3
Across sessions: easing and cleared
The golfer's Home screen adds the session records up. Its Coaching tab leads with one focus to practise, in plain words; Your swing sorts every habit ever judged into areas of the swing (sequence, posture, arms and club and so on), each opening in place; and Matchups shows the faults that travel together. For each habit, each session says one of four things: pattern; quiet (the cautious upper bound on its rate is under a half, so 0 of 4 or 1 of 8); inconclusive; or not measured. A work-on is active while the latest session that could tell found it, easing once a later one found it quiet, and cleared after two quiet sessions running.
- 4
An area is counted, not scored
Your swing gives no 0–100 mark. Each habit in an area falls in one of four bands of its firing share, recent sessions counting for more: a habit on more than half the swings, occasional at 15–50%, clean when confidently under 15%, or not yet clear. The area shows how many of each, which way the habit count is moving, and its biggest fault. A habit with no band around its line (the sequence out of order, the hips peaking after impact) reads Action when it fired on more than half the newest session's swings, and shows no grade when clean, because how early a peak should come is a coach's call.
- 5
Absence is not improvement
Only a session that measured a habit can ease it. A habit only the down-the-line camera can see isn't marked down for sessions without that camera. It stays active, saying "not measured since 4 Jul". Older evidence counts for less, each session's swings weighted by 0.75 for every substantive session since, so fifteen of fifteen a month ago counts for less than fifteen of fifteen today. It may still be there; it may equally have gone.
Illustrative: a habitual golfer's session, where no link could be tested.
One thing to send the golfer
At the end of a session a coach can make a branded PDF report and send it straight from their own mail account. Page one alone tells the golfer three things: what went well, what we found, and what to do first. The pages behind it are for anyone who wants to know why: a page per work-on, then posture, wrist and club delivery pages, and a launch monitor appendix with every shot and the session's figures with their spread.
- 1
Going well means well inside
A strength is something the golfer did comfortably: the median reading across the session sits inside the Ideal band. A posture held right at the edge of its range on every swing isn't called a strength. An outcome counts only when both ends of its range were clean.
- 2
A page per work-on
What we saw, as a count and a run of dots against the target range, one per swing, so every claim can be checked against the page. Then why it matters, what came with it, the feel, a drill in full with what to do if it doesn't work, and still frames from the golfer's own swings.
- 3
What might be behind it, all of it
Every cause the library authors for the habit, sorted into three groups: spotted in this session, could be (with how it could be checked), and ruled out. When the evidence for a cause is thin, the causes we couldn't check are the most useful thing the report can give: other things for the golfer and coach to consider.
- 4
Seen together is not cause
A link between two habits is described as "fits the coaching model", or "followed it in the right order on 7 swings", and never as cause. The single exception is a link the session's own swings support, and even then it reads only "goes away when the first one does". At the back, once, the report says what we couldn't check, in plain words. Those words, like every goal, feel and drill in the report, are part of the model, and a coach can edit them in the app.
Illustrative: the shape of the report's first page.
What the app is entitled to call true
A diagnosis is a chain of claims, and each link in it is a different kind of claim. A reading is a measurement. A norm is a statement about golfers. A causal link is coaching knowledge. A likely driver is an inference. The app keeps those apart everywhere, because the easiest way to mislead a golfer is to let a strong claim lend its authority to a weak one. These are the rules it follows.
Measured, estimated, inferred, asked
Every statement the app makes sits on one rung of this ladder, and it says which. A claim never travels up a rung by being repeated, summed or ranked.
- 1
Measured
An instrument read it directly: a pelvis sensor's turn, a launch monitor's club path. Where a launch monitor measures something the cameras also estimate, the monitor is read first, and the camera's estimate is the fallback, so club path is judged on the monitor's number when there is one.
- 2
Estimated
Worked out from what a camera can see, such as rotation from how narrow the shoulders look as they turn away. That's a real reading and a weaker one, and it is labelled estimated, with its error bar. If better kit is added later, the same reading becomes a measurement with nothing else changing.
- 3
Inferred
Reasoned through the graph from what was seen: the likely driver when nothing measured it. It is labelled "inferred, not measured", it ranks below an observation of the same thing, and it is never the headline of a report.
- 4
Asked
Knowable only from you: what you were trying to do. It is offered as a question and never concluded. And one rung stands apart from the rest: a screen, a physical test you actually ran, is the only ground truth the model has about a hidden cause.
Every claim says which rung it stands on.
Rules about readings
Not measured is never fine
the single most important branch in the engineIf no camera saw it, no sensor was worn, or the swing couldn't be split into positions, the answer is "we could not assess this", never "clean". The two are kept apart on every screen and in every count. Silence is not a clean bill of health.
Three states, never merged
gap · finding · faultA capture gap (nothing arrived), a swing finding (something arrived and was poor), and a capture fault (something arrived that the instrument can't produce). A smash factor of 1.62 is a mis-tracked ball. Grading it in either direction would launder a hardware problem into a confident diagnosis.
A reading that fails validation is held
trunkRise · 22 October 2026Before a measure may diagnose anything, it has to agree with an independent reading of the same thing. Trunk rise, the face-on view of standing up, was tested against the down-the-line forward bend on two-camera swings. It ranked swings with a correlation of 0.12, 0.15 and 0.23 on three golfers, against a bar of 0.7 for each. Swing-to-swing noise was bigger than the change it had to see.
A new rule is checked on real golfers
lead knee folding going backA new fault is run over the library before it ships, and its firing table is the evidence. The lead knee folding going back fired on 21 of 23 swings for the golfer it was written for, and on none of 127 and none of 28 for two golfers who don't do it. No other condition's result changed.
Uncertainty never changes a verdict
the verdicts-identical gateThe probability that a finding really fired is added beside the verdict, never substituted for it. A gate re-runs the library with and without it and requires every swing's fired, clean and not-measurable to match. A borderline reading may count as less than a whole firing, but only a borderline reading can.
No number is invented
no fabricated precisionWhere a reading has no stated error, none is made up. Where a session has fifteen swings, recurrence is a count ("8 of 12 measurable swings"), never a percentage. Readings are shown in whole units or the model's own decimals, and anything inferred is shown as a word. A count is something you can check against the dots.
Rules about knowledge
The library grades its own evidence
supported · indirect · practiceEvery condition and every link records whether anyone has tested it, and when the search was made. Supported: a study measured the thing itself (6 conditions, 6 links). Indirect: a study measured the mechanism underneath, and the named claim is our reading of it (34 conditions, 47 links). Practice: searched for, coaching consensus found, no peer-reviewed test of the named claim (142 conditions, 569 links).
Norms say where they came from
literature · measured · heuristicOf 271 norms, 22 are taken from a published results table: tour-player trunk and knee angles at each position, and head height among them. 3 were measured from our own swings. The other 246 are authored starting figures, marked as such, with no sample behind them. They are expected to move as measured swings accumulate.
Population, never person
a norm is never keyed to a golferA norm describes golfers, not you. Widening a corridor because your own swings fall outside it doesn't improve your swing. It only removes the app's ability to tell you about it. Your history is used for ranking and presentation, never for firing or grading.
Words that can be checked
calibration against screensThe posterior that ranks likely drivers is built from authored words, so the app checks how much those words matter. It moves every strength and prominence one step either way and names any link whose single word decides a top cause. Across sessions it compares what screens found against what the words predicted, and flags links that disagree.
The library checks itself
the health listThe app audits its own knowledge and reports, in plain words, anything that can't work as written. That covers a condition watching a measure with no norm (it can never fire), a shot context that nothing grades, a corridor that puts nearly every stored swing in one band, mismatched units, and one condition watching both ends of a range.
The model leads the cameras
142 of 186 can fire todayThe library describes more than can be captured, on purpose. A complete model is something a coach can review, and it is what decides which measurement gets built next. What can't yet be seen is reported as coverage (one quiet line saying how many of the library's conditions this capture could speak to), never as the absence of a fault.
Rules about conclusions
Facts, not judgements
what a session may assertA session asserts that readings deviated, that the deviations recurred, and that they are consistent with an authored chain. It never asserts that it has proved a cause, and the kind of shot never flips whether something counts as good. Everything statistical is used as a gate for ordering things, never as a hypothesis test with a verdict.
Report, never soften
materiality weights ranking onlyA finding that matters less on this shot is still recorded, listed and counted. It just adds nothing to the ranking. Not a fraction, because a fraction would be a number nobody could defend. Nothing is hidden because it is inconvenient or unlikely to matter.
The golfer's own swings are the evidence
every claim checkableEvery claim on a work-on page can be checked against the dots on that page: one per swing, against the target range. If the report says "9 of 10 swings", the reader can count nine. A claim that cannot be traced to swings the golfer can see doesn't make the page.
Measured before inferred, and the unmeasured named
never droppedA cause the cameras couldn't see is never a headline. It isn't dropped either. When the evidence is thin, the causes nobody could check are often the most useful thing to hand a coach: other things to consider, each with how it could be checked.
Every word on the panel, in plain terms
The app speaks in a small, fixed vocabulary, and every word in it has an exact rule behind it. These are the ladders, from weakest to strongest, with the rule each word stands for. All of the thresholds are settings the developers can tune against the corpus, not constants of nature.
A single reading
Ideal · Good · Watch · Action
gradeWithin one, two, three or more than three tolerances of the centre, on the reading's own side, under the Standard policy. Lenient and Strict move all four together.
Fired · clean · not measurable
one swing, one conditionWhat the ledger records for each condition on each swing, with the probability it really fired beside it. A condition that doesn't apply to the shot is left out entirely, and not marked clean.
A session
This shot · watching · pattern
tierOne observation; fired at least once but not yet evidence; and a habit, when the Wilson lower bound on its rate reaches 0.30 over at least three measurable swings.
Improving · stable · worsening · resolving
trendA Theil–Sen slope with a rank test at the 5% level, from five readings up. Resolving means a pattern that hasn't fired in its last five measurable swings.
Cold · forming · established · closed
the panel's stageToo few swings; at least one pattern; at least two patterns joined by a link graded coherent or better; and frozen as the session summary. One way only.
Firm · likely · fragile
stabilityHow often a pattern or a likely driver survives 200 re-runs of the session with every borderline reading allowed to fall either way: at least 80%, at least half, or less.
A link between two habits
Present together · coherent
link grade, lowerBoth ends are patterns. Then: their directions match the link, and within single swings they fire in the swing's own order.
Conditionally dependent · moved together
link grade, upperThe effect fires more often when the cause did, by a one-sided exact test at 5% over at least eight paired swings. Then: under a declared focus, the cause improved and the effect followed, five swings either side.
Authored knowledge
Rarely · sometimes · often · usually · always
link strengthHow often the cause produces the effect: 0.10, 0.30, 0.60, 0.80 and 0.95. The shipped graph uses the middle three: 97 sometimes, 383 often, 159 usually.
Rare · uncommon · occasional · common · ubiquitous
prominenceThe prior chance a cause is present at all: 0.05, 0.10, 0.20, 0.35 and 0.60. Editorial judgement, re-seatable once the swing library can count prevalence.
Across sessions
Pattern · quiet · inconclusive · not measured
what one session says about a work-onA pattern; at least three measurable swings with the cautious upper bound on its rate under a half; looked at but neither; and never looked at.
Active · easing · cleared
a work-on's statusA pattern in the latest session that could tell; quiet there after being a pattern before; and quiet in two such sessions running, when it leaves the list. A later pattern makes it active again.
Habit · occasional · clean · not yet clear
how Your swing counts an areaMore than half the swings; 15–50%; confidently under 15%; and too little to say. Each habit's share weights recent sessions more, and the area shows the four counts in place of a score.
Its own side
one-sided gradingA habit that can only go one way is graded against its own edge of the range only. A cupped wrist far out on the cupped side does not also read Watch on bowed at the top.
Diagnosed, not declared.
Every norm, condition and causal link on this page is content you can open in the app, read, and change, with the evidence behind it named row by row. It's still a beta: the sixth shipped on 1 November 2026, with 1.0 targeted for late in the year, and the numbers are being seated on measured swings as they accumulate.