Your track record

The panels above say what the Track record page counts and what it cannot know. This long read goes underneath them: what the texts this app was built from say about being right, exactly how each of the three tracks is counted, how to read every state the page can show, one record run through the real engine, and three limits that only appear when the code is executed. The page itself reads nothing from a chart, so most of what follows is about counting. But the things being counted - an insight from the Kundali, a phase reading from Timing, a verdict from the Decision engine - are all derived from a chart, and that is where the classical material and the birth time come back in.

Two kinds of number appear below. Chart figures come from the shared sample person used on this site's calculator pages - born 14 August 1994, 06:42 IST, Ahmedabad - evaluated on 5 October 2026. Track-record figures come from handing the accuracy engine logs written for this page: the rows are constructed, and the arithmetic on them is the engine's own output. Nobody's real record is shown, and no accuracy figure for the app as a whole appears, because the code does not compute one: every count is over one account's own rows.

What the source texts say about being right

Brihat Parashara Hora Shastra does not promise that its rules come true. It says who can make them come true. Ch. 27, on evaluating strengths, closes with two verses (Sl. 39-40) on who is eligible to issue fruitful predictions: the words of someone skilled in mathematics, trained in grammar, versed in justice, intelligent, knowledgeable about geography, space and time, in command of their senses, logical in estimation and well disposed to Jyotish will be truthful. The closing chapter repeats the condition (Ch. 97 Sl. 5-6): only a person with adequate knowledge of time and of the positions of the grahas and nakshatras can understand the Hora Shastra, and only one with complete knowledge of it who is also truthful can make correct favourable or unfavourable predictions. In Parashara's framing, correctness belongs to a qualified practitioner applying the whole text. A miss is the practitioner's, never the text's.

The horary tradition is more confident still. Daivajna Vallabha opens (Ch. 1 Sl. 2) by saying that the astrologer who announces results after weighing the ten avasthas of the planets and the results of the houses "never goes wrong". The same chapter (Sl. 11) tells the astrologer to study the syllables of the question and the omens around it before speaking, and Ch. 2 Sl. 1 says that when good and bad omens appear together the result is to be judged by their relative strength. Reliability is promised on condition of method, and the method includes judgement calls. Nothing in these verses describes going back afterwards to see what happened. How much of that text this app computes is set out on the Prashna page.

Saravali shows why authority alone cannot settle the matter. Kalyana Varma says at the outset (Ch. 1 Sl. 2-4) that he is gathering the useful essence of earlier teaching and leaving other portions out - an editor's judgement, stated openly. He then records disagreement where he finds it. On which planets count as mutual karakas he gives Acharya Hari's view, Chanakya's view, and a third that he says the great sages do not accept (Ch. 6 Sl. 1-4). On planetary rays he reports two schools he calls authoritative - one giving the exalted Sun 10 rays, the Moon 9, Jupiter 7, Venus 8 and the others 5, the other giving every planet 7 - and notes that most lean to the second (Ch. 36 Sl. 2-4). BPHS does the same in places: Ch. 32 Sl. 1-2 reports two positions held by unnamed others on whether Rahu belongs among the Chara Karakas. Where two schools give different answers for the same chart, both cannot match the same life, and no text can tell you which one does.

The classical rules are also layered in a way that makes a miss hard to pin down. BPHS Ch. 47 Sl. 5-6 gives one test for a period - favourable if its lord is exalted, in its own or a friendly sign or in the Lagna, unfavourable if it is debilitated, in an enemy's sign or in the 6th, 8th or 12th. Saravali Ch. 41 Sl. 17-18 then adds that if the Moon is badly placed relative to the dasha lord when the period begins, even a planet exalted at birth gives inauspicious results; Sl. 19 adds that a friendly sign with an enemy's navamsa gives mixed results, and that a weak planet's dasha comes to nothing. BPHS Ch. 27 Sl. 37-38 says whatever is stated for a house comes to pass through the strongest graha. Each condition is reasonable alone. Together they mean a reading that did not match can always be explained by a condition that was not weighed. The track record is built to remove exactly that escape: the engine's rule set is fixed in code before you ask, so when it misses, the miss is counted. What a Mahadasha is has its own article.

The modern literature moves toward checking, but not toward counting. Advanced Predictive Astrology (Vol-II) is built from case charts. On p. 131 its author states two three-planet combinations he says he has always observed and asks readers to verify them. On p. 308 the book says that by watching the slow planets many events can be forecast with 98% accuracy, provided the astrologer knows at least two dasha systems - with no count of forecasts behind the figure - and adds that a single dasha system cannot time events unless it is checked against the transits. That last sentence matters later on this page, because the Decision engine reads one dasha system at one level and no transit at all.

The numerology books come closest to a test. The Power of Numbers by Numero describes one (pp. 66-68): a woman born in 1900 is asked what happened to her in 1910, 1921 and 1934 - each year reached by adding the previous year's digits to it - and also in 1929, a year outside the series chosen at random to see whether she would report something for any year she was asked about. She named an event for each of the three series years and nothing for 1929, and the author invites readers to try the same on their own lives. It is one person, recalling events afterwards and judging their importance herself; but it has a control question, which is more than most. Numerology Made Easy warns in the other direction (p. 79): do not fall into living by your numbers, by the stars or by anyone's words, and do not accept what others say without checking it against your own experience. The same book faults a rival letter-to-number system for lacking demonstrated proof (p. 24), then on its last page (p. 194) calls a companion volume the demonstrated proof of its own. The Numerology page says which of these systems the app computes.

Three things "accurate" can mean

The three are independent, and only the third needs you, which is why it is the only one that starts empty. A chart can be computed correctly from a correctly cited rule and still say something that does not fit your life. Equally, a reading can feel right off a chart cast for the wrong hour. Methodology and Sources cover the first two senses, and why apps disagree on your nakshatra shows how two correct computations can still differ. Everything below is about the third.

Where the three kinds of row come from

Insight reactions

On the Kundali screen's Intelligence tab, the insight stream shows short statements, each one a rule over a measured condition in your chart. The rule table (insightStream.js) holds 28 rules across 25 categories - thinking style, decision bias, delay tendency, internal conflict, failure zone, what drains, what energises and so on - and a rule that cannot evaluate does not fire. Up to four pinned insights lead and five more rotate on a seed taken from the date, so the stream is the same all day and different tomorrow. Each insight prints the figures it fired on and a confidence derived from three things: how much of its scale the measurement covered, how far past its threshold the condition is, and whether a second, independently computed fact agrees. Under each one are four buttons: Strongly relate, Relate, Does not land, Confused by this.

A tap stores one row: the insight's id, its category, its kind (micro, deep or situational), your reaction, the exact wording you saw, the figures it cited and the engine's confidence at that moment. The wording is stored because rules get edited, and a row holding only an id could later claim you agreed with a sentence you never read. There is one row per insight per profile. Tapping again replaces the earlier reaction; it does not add a second.

Phase ratings

On the Timing page, the Phase Intelligence panel describes the period you are in and asks "Does this match your current life?" with three answers: Strongly accurate, Somewhat, Not accurate. This is a separate scale from the insight buttons on purpose - one asks whether a statement lands, the other whether a period reading is accurate. The row is keyed to the Mahadasha-Antardasha pair, not the date, so rating the same reading twice replaces the first rating, and a new row becomes possible only when the Antardasha changes. The text you rated and the Phase Intensity score at that moment are stored with it.

Decision outcomes

On the Decision page, "Log this decision" sends only your question and its category. The server recomputes the verdict from your chart and stores it beside the question; a verdict asserted by the client is never accepted, because a stored verdict that was not actually computed would poison the statistics that read it back. The verdict is two terms added together (decide() in decisionEngine.js): the quality of the running Mahadasha lord, and the Shad Bala of the ruler of the house the category belongs to - 10th for career, 2nd for money, 7th for love, 6th for health, 11th for business and the 1st for a general question - as a share of 40. A total of 65 or more is YES, 35 to 64 is WAIT, below 35 is NO.

Later, "Record what happened" asks two things. How did it go? Went well, Mixed or Went badly. Did you follow the verdict? Yes, I followed it, or No, I went against it. For a WAIT or a NO, following means you held off. Both answers are your own report. Neither is inferred, and a field you leave blank stays blank rather than defaulting to something that would read as a fact later. You can see how strong each house ruler is on the Shad Bala calculator and which period is running on the Vimshottari Dasha calculator.

How the counting works

Those three sets of rows are the only inputs to buildAccuracyRecord() in accuracyEngine.js. What it does with them, in order:

1. 1.One threshold governs everything: MIN_SAMPLE, which is 5. A rate is a rounded whole percentage, and it is null - not zero - whenever its own count is under five.
2. 2.Insights: the landed rate is the Strongly relate and Relate rows divided by all your reaction rows. Does not land and Confused by this both sit in the denominator. The engine counts the confused rows separately too, but the page does not print that count.
3. 3.Categories: the same division inside each category, each needing its own five rows, listed largest first.
4. 4.Phase readings: Strongly accurate plus Somewhat, divided by all your ratings. The note beneath says how many were Strongly accurate.
5. 5.Decisions: keep only rows that carry a verdict, then only those with an outcome recorded. The engine calls these resolved, and the track has enough when there are five.
6. 6.Split the resolved rows by your answer to "Did you follow the verdict?". A row with an outcome but no answer to that question is resolved and sits on neither side.
7. 7.On each side, the rate is the Went well rows divided by all rows on that side, shown only when that side holds five. Mixed is in the denominator and not the numerator, so it lowers the rate exactly as Went badly does.
8. 8.The two sides are declared comparable only when both hold five. Until then a copper-coloured line states how many each side has and why they are not being compared.
9. 9.If no track has reached five, one card above the three panels says so. The panels still render, each with its count against five.

The browser does none of this. Every rate, count and sentence is a field of one server response, and the page is forbidden by its own header comment from averaging the tracks, from drawing a withheld rate as 0% or an empty bar, and from softening a low figure. A withheld rate is drawn as a dash with the engine's note underneath.

Why five? The engine's comment gives the reason as arithmetic: a rate off two entries moves about thirty points on the third. The general rule is that with n entries recorded, one more can move the rate by up to 100 divided by n + 1 points. Five is where that worst case first drops under twenty. It is a floor against the most misleading numbers, not a promise of a stable one.

One record, worked through the engine

The sample chart has Leo rising and the Moon in Vishakha. On 5 October 2026 it is in the Mercury Mahadasha (18 January 2024 to 17 January 2041) and the Ketu Antardasha (16 June 2026 to 13 June 2027), so the phase reading open for rating is Mercury-Ketu. Mercury sits in Cancer, a neutral sign, but in the 12th house, so the engine reads the Mahadasha as challenging and the first term of every verdict is 0. The background to that period is in the Mercury Mahadasha article. The six verdicts the engine would store today:

The insight stream generates 17 insights for this chart and shows nine on the day: four pinned and five rotating. The highest-confidence one reads off Saturn's index of 90 and Rahu's of 52 and says the life builds slowly and then moves all at once (confidence 91). The lowest shown says the chart is 25 points more supportive of career than of wealth (confidence 40, the condition having cleared its threshold by nothing at all). Fifteen categories fire; the largest, internal conflict, holds three insights. Visiting daily from 5 October, all 17 have appeared at least once after fourteen days.

Now four constructed logs, each handed to buildAccuracyRecord() as written. Stage A is a new account with nothing logged. Every rate is null and each panel reads its count: 0 of the 5 reactions needed, 0 of the 5 phase ratings needed, 0 of the 5 recorded outcomes needed.

Stage B is a first week. The reader reacts to all nine insights on screen - two Strongly relate, four Relate, two Does not land, one Confused by this - rates the Mercury-Ketu reading Somewhat, and logs three decisions, recording an outcome for one. The insight track clears its floor at once and returns 67%, with the note "You marked 6 of 9 insights as landing." All nine category rows read "1 so far - too few to rate". The two decisions with nothing recorded are simply absent from the decision track, which reads 1 of 5.

Stage C is the following summer, after the Antardasha has changed on 13 June 2027 and a second phase rating has become possible. All 17 insights have a reaction: eleven landed, two confused, four did not land, for 65%. Internal conflict leads the category list with three rows, two of them landed, and still no rate. Fourteen decisions are logged: six followed (four went well, one mixed, one badly), three not followed (two went well, one badly), two with an outcome but no answer about following, and three with no outcome. The engine returns 11 resolved of 14; followed, 67% on six; not followed, a dash on three. The copper line says the comparison needs five decisions on both sides and that this record has 6 followed and 3 not followed.

Stage D is a run that went badly, constructed to see what the engine does with it. Six reactions, one landed: 17%. Ten decisions, all resolved: followed, one of five went well, 20%; not followed, four of five went well, 80%. The comparison line reads "Both sides have enough entries to compare." Nothing else is added, because the engine holds no softening sentence to add.

How to read each state of the page

• A dash. Not enough rows to quote a rate. It is not a low score, and the note under it says how many of the five exist.
• A landed percentage. The share of the insights you reacted to that you marked Relate or Strongly relate. It says nothing about insights you saw and did not react to.
• A category row with a count, such as "3 so far - too few to rate". That category holds fewer than five reactions. For the reason given in the next section, expect every category row to look like this.
• An accurate percentage on phase readings. The share of the periods you rated that you called Strongly accurate or Somewhat. Because Somewhat counts, read the "of them strongly" figure in the note as the stricter number.
• Two decision rates, one with a dash. One side has five resolved decisions and the other does not. The copper line tells you the two counts. Do not compare a number with a dash.
• Two decision rates and "Both sides have enough entries to compare." The only state in which the two numbers are offered as a comparison. The page still does not say which is better, and it does not test whether the gap could be chance.

Three limits that only show when the code is run

A category cannot currently reach five

Reactions are stored one per insight, and each category has only a few rules behind it. Reading the rule table, the most distinct insights one category can produce for one chart is four - internal conflict, from up to three measured contradictions plus one rule about two closely matched forces - and for timing-related insights, four spread over years as the phase type changes. To check this against real output, the stream was generated for 400 charts at 21 dates each, two years apart from 2026 to 2066. On the first of those dates a chart produced between 4 and 23 insights, 11.5 on average, and across all the dates no chart reached five distinct insights in any category. So on a chart whose birth details do not change, with the rule table as it stands on 5 October 2026, every category row will show a count and none will show a percentage. The percentage branch exists in the code and will begin to appear only if more rules are written.

The phase track takes years

One rating per Mahadasha-Antardasha pair means five ratings need five different Antardashas. BPHS Ch. 51 Sl. 1 sizes an Antardasha as the product of the two lords' years divided by 120, which gives lengths from about three and a half months (Sun-Sun) to three years and four months (Venus-Venus). Depending on where in the 120-year cycle you stand, being present in five different Antardashas takes from about fourteen months at the very fastest to more than eleven years. For the sample person, rating Mercury-Ketu today, the fifth distinct reading - Mercury-Mars - does not begin until 19 July 2032. Only the reading on screen can be rated, so earlier periods cannot be filled in afterwards. For finished Mahadashas the Timing page offers a retrospective instead, which sets what the chart said beside what you logged and deliberately refuses to score the match. Your own dates are on the Timing page.

The verdict for a category does not move within a Mahadasha

Both terms of decide() are read from the birth chart: the natal condition of the Mahadasha lord, and the natal Shad Bala of a house ruler. Neither changes from day to day. So for the whole of a Mahadasha, every decision you log in a given category receives the same verdict and the same number. As long as the chart and the engine stay as they are, the sample person will be told NO at 11 for every career question until January 2041. The decision track is therefore not testing many separate forecasts. It is testing up to six standing verdicts, over and over, against whatever you happened to ask about. When the standing verdict is NO or WAIT, "followed" means "did not act", and the outcome you record is the outcome of not acting - which is harder to judge than the outcome of something you did.

What a good or a bad number cannot tell you

• It cannot tell you whether astrology works. It counts your reports about one engine's output on one chart. The Decision page's own trust panel says so in terms: this is a correlation, not proof that the guidance caused the outcome.
• It cannot correct for mood. A period reading rated on a hard day is rated by someone having a hard day, and "Went well" is whatever you meant by it when you tapped it. The engine behind the Timing retrospective refuses to turn logged events into a match percentage for this very reason: people log unevenly and log more when things are hard.
• It cannot weigh confidence. Each reaction row stores the engine's confidence at the time, each rating the Phase Intensity score, and each decision the verdict's number, but the accuracy engine uses none of them. A NO at 11 and a WAIT at 35 count the same; so do an insight at 91 and one at 40.
• It cannot separate verdict types. YES, WAIT and NO are pooled on this page. The decision profile computed for the Decision page is where follow-through is broken down by verdict.
• It cannot tell a weak method from a thin one. Advanced Predictive Astrology (Vol-II, p. 308) says a single dasha system needs checking against transits before it can time events. The Decision engine uses the Mahadasha lord alone, by design - its header explains that a third, transit term would have meant inventing a favourability rule the app does not otherwise use. A low decision rate may be reporting that simplification.

Birth time: the record is only as good as the chart it rated

This page reads no chart, but every row on it is a judgement about something a chart produced. If the birth time is off, you have been rating readings of a slightly different person. The sample chart shows how much moves. Its Lagna is only 2.3 degrees into Leo, so it is close to a sign boundary - the kind of case the article on birth-time sensitivity describes, and one you can test for your own chart on the Lagna calculator.

Three things follow. Four minutes moves the Antardasha boundary by seventeen days and none of the six verdicts. Fifteen minutes earlier crosses into Cancer rising, every house changes its ruler, and two of the six verdicts flip. An hour either way leaves you rating a different Antardasha altogether - Mercury-Mercury or Mercury-Venus instead of Mercury-Ketu - so a "Not accurate" could be a correct rating of the wrong period.

With no birth time at all, the Decision engine declines: it returns WAIT with a confidence of 0 and a sentence asking for a birth time. If that is logged, the row carries a verdict, and the accuracy engine's filter - any row with a verdict - does not tell a refusal from a reading. The Decision page hides the word WAIT on such rows; the Track record page counts them. If your birth time is missing, read what still works without one before treating the decision track as evidence. And if you later correct your birth details, stored rows are not recomputed: old verdicts stay as they were logged, and an insight that still fires keeps showing your earlier reaction, because the mark is matched by the insight's id. React again to replace it.

The same decisions, counted four ways

The app holds more than one figure built from decision rows, each with its own definition. They are not interchangeable, and seeing them side by side explains why two screens can show different numbers for the same history. All four were run on Stage C's fourteen decisions.

The 67% in every row is not one number. Three rows reach it as 4 of 6 and the profile reaches it as 6 of 9; they only happen to round the same. The Track record figure is the only one that reports the not-followed side as a rate of its own, and the only one that counts a decision with an outcome but no followed answer as resolved. The follow-through line needs three closed loops, which is why it can appear on the Decision page before the Track record shows anything.

Where the track record connects to the rest of the app

The track record is a reader of other features' rows and nothing reads it back. Reactions and phase ratings are loaded by the Track record page, by the two screens that show your earlier mark beside the item, and by account deletion - by no engine that produces a reading. Marking an insight as not landing does not remove it, reorder the stream or change what AI Chat or the daily page says. The reasoning is in the code: every insight is a rule over a measurement, so a "does not land" is you disagreeing with a measurement, and an app that hid the readings you disliked would end up telling you a more agreeable story than your chart supports.

What each track rates is explained elsewhere on this site. The measurements under the insights - graha and house strength, and the bindus that feed them - are on the Shad Bala and Ashtakavarga calculators and in the Kundali long read. The periods under the phase readings can be dated without an account on the Vimshottari Dasha page. The verdicts under the decision track are set out on the Decisions page, and the transits they leave out can be seen on the Gochar calculator. When a chart reading and a number reading pull apart, this article and the Convergence page describe how the app handles it; the decision verdict itself comes from astrology alone.

The rows are personal data. Each reaction carries the sentence you reacted to and each decision the question you typed. Deleting your account removes all three collections along with everything else, as the privacy page describes. Short of that, an entry can be changed but not individually removed.

Keeping a record worth reading

1. 1.React to every insight on screen, including the ones that miss. The rate is taken over what you reacted to, so skipping the misses is the one thing that reliably inflates it.
2. 2.Rate a phase reading once you have lived in it for a while. A later rating replaces the earlier one, so there is no cost to revising.
3. 3.Log a decision before you act, not after. A decision logged in hindsight is one whose outcome you already knew.
4. 4.When you record an outcome, answer both questions. An outcome without the followed answer counts as resolved and feeds neither side of the comparison.
5. 5.Close the decisions that went badly as readily as the ones that went well. An open row is excluded, not failed, so leaving the bad ones open flatters the rate.

Questions people ask about the track record

Is this an accuracy score for the app?

No. It is three separate counts over rows that you created on your own account, and the engine refuses to average them into one number. The code computes no figure across users, so there is no app-wide accuracy rate to quote, here or anywhere else on the site.

Why does my page show dashes instead of percentages?

A dash means that track, or that side of the decision comparison, holds fewer than five entries. It is deliberately not drawn as 0% or as an empty bar, because "not enough to say" is a different thing from a low score. The note under each dash gives the count.

Why do my category rows never show a rate?

Each category needs five reactions of its own, reactions are stored one per insight, and with the rule table as it stands no category produces more than four distinct insights for one chart. A sweep of 400 generated charts over forty years of dates found none that reached five. Until more rules are written, expect a count in every category row.

How long before the phase track shows a number?

It needs ratings on five different Antardashas, and only the one on screen can be rated. That takes anywhere from about fourteen months to more than eleven years depending on which periods you are passing through. On the sample chart it is July 2032. The lengths follow from the Vimshottari proportions described in Antardasha and Pratyantar, explained.

If I mark an insight as not landing, will I stop seeing it?

No. Reactions are recorded and change nothing about which insights are shown or in what order. The screen marks the insight with your reaction and leaves it exactly where it was.

How are "Confused by this" and a "mixed" outcome counted?

Both sit in the denominator and not the numerator, so each lowers its rate the way a plain miss does. Confused is stored as its own value, and the code comments say it is meant as a signal about wording rather than about the measurement. The trust score computed for the Decision page reports mixed outcomes as their own figure.

Can I change or delete an entry?

You can change one: a new reaction or rating replaces the old, and recording an outcome again overwrites it. There is no control that removes a single reaction, rating or logged decision, and a recorded outcome cannot be set back to blank. Deleting the account removes all of them.

Do reactions I make on a family member's profile count here?

No. Those are stored under that profile, and the Track record page asks only for your own default scope. Phase ratings and decisions are always made on your own profile.

My followed rate is high. Does that mean I should always follow the verdict?

It means that, on the decisions you chose to log and close, the ones you followed mostly went well by your own account. It does not show the verdict caused that, and until the other side also holds five there is nothing to compare it with. Within one Mahadasha the verdict for a category is the same every time, so read the rate alongside what the verdict actually is - the Decisions page shows it with its two terms.

I corrected my birth time. What happens to my record?

Nothing is recomputed. Decisions keep the verdicts they were logged with, ratings keep the text you rated, and reactions keep the wording you saw. New entries are made against the corrected chart, so for a while the record mixes two charts. The page does not mark which rows belong to which.