Privacy
Offer Filter privacy
Beta privacy policy, effective 19 November 2026 · for Offer Filter 0.6.52
Offer Filter does not require an Offer Filter account and has no ads or mandatory analytics. It does not sell or rent your data. Offer reading and decisions happen on your phone. Update requests, feedback you send, the optional anonymous diagnostics, the anonymous daily count (on unless you turn it off), place-name lookups and links you choose to open can send information to the services described under What leaves the phone. The developer's public project identity is Dillxn. Offer Filter is beta software, and this text changes with it (see Changes).
What it reads on the phone
- When you enable Time-aware scoring (off by default, from 0.6.46), existing retained offer timestamps and numeric facts are compared on this phone by local time of day. Only eligible offers from the last 28 days of the existing last-200 history are used; this does not lengthen retention or create a second offer history. The device's current time zone interprets those timestamps, including older ones; the app does not recover an earlier time zone or infer a market from location. The switch is saved, and a decision's numeric scoring snapshot records its multiplier, nearby and overall sample counts, number of nearby dates, and local reference minute. Working estimates live in memory and are recomputed after restart. Clear history removes their evidence and the snapshots. No new screen content, location fix, identity or network request is collected for this feature. Existing optional diagnostic reports can include the decision's score and whether time-aware scoring is on, under the same reporting choices described below.
- Dasher's screen, through Android's accessibility service, while Dasher is on screen or in its half of a split screen: each offer's pay, miles, time, stops and displayed total item count, and the other text Dasher shows, which can include store names, customer names and addresses. With no rule set, it reads nothing of Dasher's screen; Android's list of windows alone places its tab. While auto-decline is paused (from 0.6.35; before, it read nothing then either), it reads Dasher's screen as it does when on, to show each offer's score and what auto-decline would do, and taps nothing; turning Offer Filter off in Android's Accessibility settings stops all reading of Dasher's screen. Map views it recognizes inside Dasher's screens are skipped, not read, except while it checks a failed decline or an automatic Accept, or when a map's own label shows a sign of an offer; whether Dasher's maps are recognized depends on how Dasher draws them.
- On Dasher's decline question ("Are you sure you want to decline this offer?"), whether the decline is the app's or yours and whether Autopilot is on or off: the acceptance-rate percentage Dasher shows, and whether Dasher says declining that offer does not lower your acceptance rate. See Autopilot and your acceptance rate below.
- Dasher's notifications, through Android's notification access. Android shows the app every notification; it ignores every app but Dasher. To decide whether a card of its own is needed, it also checks how Android ranks Dasher's offer notification: whether it can pop up or sound, whether it sounded, and whether Do Not Disturb lets it through. It remembers which of Dasher's notification channels its offers came on and whether that channel alerts by itself. Its logs note that channel's sound, vibration and importance settings, whether the notification asks to fill the screen, and the names (never the values) of the notification's fields.
- Dasher's short error messages (toasts), through accessibility, only to recognize a failed decline. Their wording is not stored or sent; recovery logs keep a fixed error category and attempt counts.
- Peek, on by default after the current notice is accepted, can briefly bring Dasher forward to read a fresh background offer, or one that arrived while the phone was locked, right after you unlock it (within 40 seconds of its notification, while that notification is still up). To return you afterward, it checks Android's window information and the app identifier of the app previously in front. It does not read that app's text. To tell which app a window belongs to, it asks Android once for that window's top element and uses only its app identifier. It also checks screen lock, keyboard and window state, call/audio mode, and Android's available microphone/camera-in-use indicators; it does not record audio or images. If Dasher opens without showing the offer's details, or is already on screen still showing its wait for offers 5 seconds after an offer's notification, Peek (or, after you tap one of the app's offer cards, the screen reader) may send Dasher's own offer-notification tap once for that offer, an intent Dasher itself made, or, when Dasher shows its Finding offers screen, tap Dasher's ⚙ button once instead (after a fresh read, never within 2 seconds of your touch); when you tap a card saying Dasher didn't show the offer, it opens Dasher's own offer screen first, and Dasher's launcher if nothing of Dasher's comes up.
- When you tap one of the app's offer cards, it checks Android's window information and the identifier of the app you tapped it from, only to offer a Back to map button when that app is a navigation app (Google Maps, Google Maps Go or Waze). After an update installed while Offer Filter's screen was open, it checks which app is in front the same way (window information and that app's identifier, never its text), to reopen Offer Filter only over its own screen or the home screen.
- When a finger lands on the screen, as a time only. While it declines an offer (and for a few seconds after), while it checks an offer before an automatic Accept, while it waits to decline an offer with Decline after 10 seconds on (from 0.6.35), and before and during a peek, an invisible one-pixel watch notes when a finger lands, so your touch hands the offer back, keeps a waiting offer yours or keeps Peek from opening Dasher; the touch still goes where you touched. It never learns a touch's position or what was touched; Offer Filter's own screens note when they are touched, so in split screen a touch on Offer Filter's half is not taken for one on Dasher's. Touch times stay in memory; the logs say only that a touch came, how soon after the app's own tap, and whether it was on Offer Filter's half.
- For the screen hold during a dash (see Your choices): whether the phone is on and unlocked, whether a call is under way, the battery level, whether it is charging and whether battery saver is on. It looks at no other app's windows or words for it.
- Approximate location, only if you allow it, for the offer map. Auto theme can use the same existing, fresh approximate location while the offer map is enabled, to calculate sunrise and sunset on the phone. Theme selection never requests a new location fix or precise location.
- Android's own record of how the app last stopped (Android 11 and later), to notice a crash or a stop for not responding.
- Android's own settings for Offer Filter itself, to show setup's page, the homepage's update line and Settings' rows: how Offer Filter was installed (Android 13 and later: from a downloaded or local file or an app store, and the name of the installing app); Android's restricted-settings record for Offer Filter, where Android lets the app read it; whether Accessibility and notification access are on for Offer Filter and connected; whether its notifications are allowed; whether it may install its own updates; and whether Android limits its background battery use (the Restricted battery setting, or the restricted standby bucket on Android 11 and later). Apart from how Android ranks and sets up Dasher's offer notifications (above) and, for reports, Dasher's version and how its launcher screen starts, it reads no other app's settings.
Before you accept the current first-run notice, it reads none of this and sends no feedback, diagnostics or daily count, with three exceptions that read no words: the channel a new notification of Dasher's came on, to post one reminder that the app is paused; after an update, which app is in front, to reopen Offer Filter only over its own screen or the home screen; and whether Android lets it install its own updates. It checks for updates only once you accept the current notice (from 0.6.18), and accepting it runs the first check.
Optional offer timing study
The offer timing study is part of Share anonymous diagnostics (from 0.6.34; before, Settings offered it as a separate choice, “Include future offer timing study in diagnostics I choose to share”). It records nothing while that switch is off. Turning the switch on starts a fresh study; turning it off, or Clear history, ends it and removes its data from this phone and from unsent reports. The update to 0.6.34 forgets the separate choice and removes what it kept. Accepting the notice never turns the switch on.
While it is on, new observations are processed locally to compare an estimate frozen before an outcome with subsequently observed waiting and numeric offer quality. This study uses a separate instance of the existing historical-mean estimator, trained only on observations made since the switch was turned on (and reset after a process, session or policy change); it is not a claim to capture DoorDash's algorithm or every homepage estimate. It stores no old waiting-history export. Records contain method/schema versions, the estimate or an unavailable marker, actual evidence counts and exposure, a last-complete-wait baseline when available, normalized filter score/bar, a local sequential policy epoch, UTC hour without date, numeric pay/miles/minutes/stops/items, fixed requested-action and observed-outcome categories, touch-takeover evidence and explicit unknown/censor reasons. It does not compute or invent prediction confidence. Hidden, locked, paused, disconnected and missing time is not counted as observed waiting; unfinished waits remain censored. Requested taps are not proof of acceptance or decline. Time, location, demand and unseen offers confound this observational comparison.
The study retains at most 32 result rows and up to 200 new numeric training samples, for at most 24 hours during use; training and live timers are not restored after restart. Its report section is at most 10,000 characters and ten recent rows, with omission counts, inside existing report limits. No customer/merchant text, address, coordinates, raw nodes/screenshots, account/device identifier, acceptance rate, Autopilot goal/plan, configured minimum amounts or free text is included. The feedback dialog's Preview shows the study section with the rest of what goes. Study data goes only with the diagnostics the switch sends: with feedback you send and in the capped summaries after each dash. It uses the same feedback recipient, queue limits and stated 90-day server retention below; this candidate does not establish that server retention has been verified. Existing automatic-summary exclusions remain unchanged.
Uncheck and save to stop new study collection/sharing and discard local study data and study sections in unsent reports. Other feedback and the automatic-report choice remain unchanged. An emptied study-only report part carries a fixed omission notice so the ordinary report can finish under its original submission token and part count. Clear history also removes study data and invalidates prepared study sections. A report already sent or a request already in flight cannot be recalled by these controls. No driving/filter/acceptance behavior changes. Review before sharing; numeric combinations can still be identifying in context.
What it keeps on the phone
- Your rules (pay per stop, per mile, per hour and max stops) and when your current minimums took effect, your separate auto-accept choice (off by default), and Autopilot's settings: whether it is on, your goal (pay first, the default from 0.6.18, or an acceptance rate of 70% or 50%) and whether you picked it yourself in its goal question, its current bar and whether you have answered that question (until 0.6.21, also whether Let my minimums grow was on; that choice is removed from the phone with the growth it switched). What else Autopilot keeps is under Autopilot and your acceptance rate below.
- Your separate scoring-mode and acceptance-learning choices, and whether you picked the mode yourself: area mode is on by default (from 0.6.27, and set once then for earlier installs); learning is on by default (from 0.6.1, and turned on once then for earlier installs). Learned evidence starts empty: deleted old lessons and historical decisions are not reconstructed. While filtering and learning are on, positively confirmed manual standalone acceptances can store the best accepted pay per stop (that offer's pay and its stops, the pickup included) and independent exact best pay-per-mile and pay-per-minute numeric pairs, and a new best that asks more than a saved minimum applies as a learned minimum alongside it; the saved minimum stays as you set it unless you adopt the learned ones. Automatic accepts, declines, add-ons and ambiguous outcomes do not train it. Unknown facts do not teach their axes; any known implausible figure rejects the lesson. The app also keeps a bounded local identity record to prevent the same acceptance teaching twice and a generation marker to reject stale observations. Learned bests have no age expiry; pausing or switching learning off retains them. Confirmed Reset clears active bests and Undo. Earlier decision snapshots keep their original numeric saved/learned requirements and model for provenance until Clear history, eviction by newer decisions or uninstall. Clear history and uninstall remove both active learning and decision history. Adopt keeps the saved minima before and after that adoption and its identity/time for conditional Undo until a later change, Reset or Clear history invalidates it. When a lesson raises the learned minimums, the app keeps the learned bests of that lesson, with its identity, for the homepage note until you close it, adopt them, learning changes, Reset or Clear history. Adopted saved minima remain settings after those clears.
- The latest 200 decisions: when, the pay, miles, minutes, stops and total items read, a numeric local snapshot of the saved, learned and independent Autopilot requirements used, whether the offer declared items or shopping, the result, what the app did, the recorded scoring model and score (strict pay fitness or area fitness; from 0.6.52 an add-on's line also keeps the pay, miles, minutes, stops and items of your route before it and of the whole route after accepting it, and its score is worked out from those each time, not stored), the bar it was judged at and whether Autopilot set that bar, whether Dasher said declining it does not lower your acceptance rate, what was seen of it afterward (accepted, declined by hand, or not counted and why), and a few masked lines read from the offer. This history has no age-based expiry: entries remain until newer decisions replace them or you clear history or uninstall. From 0.6.48, when it is full, the oldest offer whose pay was never read and that decided nothing (not accepted, not declined by the app or by you, with no acceptance-rate outcome), at least 30 minutes older than the newest, is replaced first, so offers the app could not read do not push the readable ones out; only when there is none does the oldest decision go. Aggregate outcome counts are kept separately until you clear history or uninstall.
- Your skip list (from 0.6.10): the names of the stores you add from an offer's details, until you remove them in Settings → Skipped stores or uninstall; Clear history leaves it, as it leaves your rules. The stores read on recent offers, for their details' Skip buttons, are kept only in the app's memory, never written to the phone's storage or put in a report. The offers declined for the skip list keep the reason "on your skip list" in the decision history, never the store's name, and the logs note only that a store was added or removed and how many are on the list.
- Your cost per mile (from 0.6.10): a whole number of cents, off unless you set it in Autopilot's details, kept until you change it or uninstall; Clear history leaves it. Autopilot's log notes when you change it.
- "Same store" notes (from 0.6.9): when an offer from one store (two orders from it included) comes from the store of an offer declined in the last 30 minutes, both offers' lines in the decision history note how many minutes apart they came and the other offer's pay. The store's name is read from Dasher's screen and kept only in the app's memory for 30 minutes, never written to the phone's storage or put in a report; the lines keep only the minutes and pay.
- Up to 200 numeric wait-estimate records in a rolling 24-hour window, pruned when used: the observation time, positively observed waiting duration and any observed offer's pay, miles, minutes, stops and item count/applicability, used for the wait estimate and by Autopilot to estimate how often offers arrive. These separate records contain no merchant names, screen text, addresses or coordinates and are never included in reports or uploaded; Autopilot's log lines carry only the offer rate (offers an hour) it works out from them. Only visible, unlocked, eligible waiting is measured; rejected-offer handling counts only after a later recognized wait or standalone offer confirms it continued. Hidden, unknown, delivery, locked, paused and stopped-service time is excluded. A restart does not resume an old waiting timer. The estimate rechecks these numeric offers against your current rules and stays unavailable until enough readable observations exist; old decision timestamps alone are not treated as waiting time.
- Two rolling logs of screen text, one for offers and the app's status and one for other dash screens. Other screens are captured only during a dash or within ten minutes of a readable offer, and only when positively recognized as offer, confirmation, idle, delivery or navigation screens. Unrecognized/partial screens retain only a generic text-not-kept note. Each holds about what one report carries, with a 24-hour retention window. Older entries are pruned during app use and when the logs are read; files can remain while the app is stopped or a storage operation fails. Screens recognized as payment, account or earnings pages are discarded; at most one line a minute records that a screen was not kept. The logs also hold lines in fixed words, counts and times: how much reading cost Dasher (once a minute), changes of the screen hold, a dash's end, setup steps, Autopilot's plans and bar changes, and what Peek did, including Dasher's own screens after it opens (short class names, at most six per launch), what a peek cost Dasher (counts of reads, nodes and slow fetches), the shapes of money labels when pay could not be read (counts only, never amounts), how much of the screen Dasher had when it was resized, what happened to Dasher's own notification tap and to an offer that arrived while the phone was locked, and what became of the daily count (sent, refused with the service's status code, or waiting to retry) and when its switch changed, never its code.
- A few latest states, in fixed words, for reports: Dasher's version and how its launcher screen starts, the settings of Dasher's offer notification channel and the names of its notification's fields, whether the touch watch could be put up, that Dasher reported an error after a decline, the last update result and what last became of the daily count (sent, refused with the service's status code, or waiting to retry). At most 12 are kept, each replaced when it changes, until Clear history; reports carry those from the last day.
- The app's last status line, with its date and time: what it last did or found, which can include the last offer's figures (pay, miles, minutes, stops and item count) and the pay it required. The next status replaces it; Clear history does not remove it.
- The offer map, if it is on and location is allowed: up to 300 historical squares of about 2 km where offers came in, each with its number of offers, their total and best pay, the pay and miles of those whose miles were read, and when it last had one, with no age-based expiry; also the last offer's pay, miles, minutes and stops and when it came, so one offer seen twice counts once, and a count of offers that came with no location. The separate place-name cache holds at most 300 rounded positions, which can include where you opened the app (the place the homepage names), also without age-based expiry; older names are removed as others are used. Clear history or uninstalling removes all of it.
- While you are dashing: whether a dash is on, when it started and was last seen (also on the phone's since-boot clock with its boot number, so a changed clock cannot release an update held for a dash early: an update the app finds by itself waits for the dash to end, or for 8 hours with nothing of a dash seen, Dasher not in front and the screen off), when it ended and whether Dasher showed it paused; and the current route's pay, miles, time, stops and observed item count/applicability, until Dasher shows the wait for offers, the dash's end or its home screen, or you tap Forget on the homepage. A route is never used more than three hours after it was observed; it is deleted the next time the app looks at it after that. For the screen hold, when something of the dash was last seen is also kept in memory only.
- One temporary record that prevents automatic taps on a possibly taken-over or automatically accepted offer after a restart: numeric pay, miles, minutes, stops, observed item count/applicability, countdown, and device uptime/boot number. It can suppress taps for at most two minutes; the app discards it when it next checks an expired record, observes the offer ended, or you clear history. It never restores permission to tap. One scanner-error category, without the error message or screen text, is kept until the next service connection or Clear history.
- A separate temporary automatic-Accept request record keeps only numeric offer facts and device uptime/boot number, for up to two minutes. It only marks an automatic acceptance as automatic in your history; it never proves acceptance or restores tap authority. Clear history removes it.
- The state of updates: the version last advertised; when updates were last checked; a verified update still waiting (its version name) and why it waits (a dash, Android's confirmation, or an installation Android blocked); the newest version an "allow updates" notice was shown for; and the downloaded update itself, until it installs, a newer one replaces it or it is no longer needed.
- Setup progress on this phone: which of Android's pages setup opened, and why (to turn a switch on, or off and on); which switch you were turning on when you left for Android's settings or App info (acted on only within ten minutes); whether Android blocked a switch tried from setup, and whether restricted settings were found allowed; whether setup's first steps are done, and whether Android was already asked for alerts, for location from setup, and for location all the time. Also the one-time notes already read: Peek's introduction, and the version whose "What is new" note was shown. Uninstalling removes it.
- Feedback waiting to send. A submission is saved on the phone before it is sent, so no signal or a restart loses nothing: your words, its category, and any masked diagnostics or offer report that go with it. At most 5 submissions wait, each at most 7 days; one still unsent after 7 days is discarded unsent. Each is deleted from the phone once the feedback service has accepted it, or refused it for good.
- The last 10 references of accepted submissions: the date, the short reference and whether it was feedback, an offer report or diagnostics after a dash, so you can refer to them. Newer ones replace older ones; uninstalling removes them.
- For the anonymous daily count: whether Send an anonymous daily count is on; the last UTC day whose count is done, with when it was done (also on the phone's since-boot clock with its boot number, so setting the phone's date forward or back cannot send a second count the same day); that day's random code until its count is done (only to retry it, and deleted the next time the app looks once its UTC day is over); and whether the feedback service has ever accepted a count from this install. Numbers and a yes or no; turning the switch off discards the day's code, and uninstalling removes all of it.
- After the app stopped unexpectedly (a crash, or Android found it not responding), a short summary of each stop in the last 24 hours: the kind of stop, the minute (UTC), whether the app was in front, and where in the app's code it happened (the error's type and the top of its stack, or the top of the main thread for a stop for not responding). Never an error's message, a trace's other text or anything read from a screen. It is kept for 24 hours; Clear history removes it.
- With Share anonymous diagnostics on: counts of the current dash's problems by fixed category, and the masked read lines of up to six recent problems, until that dash's summary is queued. Turning the option off or Clear history removes them.
- Your Day, Night, System or Auto theme choice. Auto is the default for new choices; it calculates the local sun cycle using only an existing permitted approximate location, without a network request or a separate saved location. When none is available it uses a disclosed local-clock fallback: day from 6 am to 6 pm. Existing explicit day/night choices are preserved. Also where you moved the filter tab over Dasher, and whether one-time hints were shown.
- While it has turned media or alarm sound down during a decline, the level to put back, so the sound is restored even after a crash; it goes once the sound is back.
- Your other switches (such as Peek, Decline after 10 seconds, the offer map and the sound turn-down), which notice you accepted and when, and small flags about what it has already done (one-time cleanups, the paused reminder), as numbers and fixed words.
- During Peek, the previous app's identifier and launcher component stay only in memory. They are cleared when Peek ends without a return, when a return fails, or after the return check (up to 1.5 seconds after returning), with one exception. When Peek leaves Dasher up, or a card's tap opens Dasher, from a navigation app, that app's identifier and launcher component stay in memory for the Back to map button until it has shown for 12 seconds, the offer did not end within 90 seconds, an offer is accepted, a delivery starts, the phone locks or the screen goes off, another app comes in front, you tap it, auto-decline is paused, or screen reading stops. A peek held for the unlock keeps them at most 60 seconds. An offer kept for the unlock holds no app at all: only its notification's key and post time, the store name it shows and its card, in memory until the next unlock (looked at then only if its notification is still up and at most 40 seconds old), or until you tap one of the app's cards, a newer offer takes its place or screen reading stops. The previous app's identity is never saved or sent; Peek and Back to map logs describe only the kind of previous app, such as a navigation app or the home screen. Its screen text is never kept.
Item counts on shopping offers are kept only as offer figures; no rule uses them. An unread declared count stays unknown; no quantity is inferred from product names, unique-item counts, stops or orders. These numeric item facts can appear in the reports and diagnostics you choose to send, like the other offer figures.
The historical 0.5.0 migration deleted the old learned minimums and the held hand-decline record from the phone the first time it ran: what older versions' adaptive minimum learned from offers you accepted or declined (their pay and rates, and when learning was on or reset), the record of offers you declined by hand, and the retired rules (per stop, per item, the hotspot distance, the minimums percentage and score by area), after building per stop and the minimums percentage into your minimums. Until you close the one-time note about this change, a summary of it (your old and new minimums, as numbers) stays with your rules. Decisions an older version recorded keep the reasons and notes they were recorded with, which can name an amount its adaptive minimum had learned (such as "must match best accepted $0.45/min"), until newer decisions replace them or you clear history; like the rest of the history, they can go in a report you share or send. Log lines it wrote about what it learned age out with the logs within 24 hours.
Decision and confirmation lines take priority over repeated scan timings within the existing log limits. Reports include numeric counts of discarded log lines, so missing evidence is visible. Atomic history writes keep a recovery copy only while replacing the same bounded history.
Screen text is masked before it is kept: customer names, your own name, street addresses, city, state and ZIP lines, apartment numbers, phone numbers, email addresses, delivery instructions, card numbers and their security codes/expiry/PIN, and streets in navigation become [name], [address], [phone], [email], [instructions], [card] and [street]. Store names, offer figures, buttons and shopping items stay. Offers are judged from what is on screen; only what is kept is masked. Click diagnostics retain the control shape and action category, never its text labels.
Masking is pattern-based and can miss unfamiliar wording. It is not a guarantee of anonymity or removal of every personal detail. Review what you send with Preview first. What you type in feedback, including notes about a selected offer, is not masked; do not include customer, payment or account details. Diagnostics and offer reports waiting to send are masked again with the current rules immediately before sending.
Versions before 0.5.0 could connect to GitHub and file problem reports and diagnostics there, as issues in the project's GitHub repository (https://github.com/Dillxn/dasher-offer-filter). That repository is public, so what was filed there could be read by anyone; earlier versions of this text wrongly called it private. This version sends nothing to GitHub. The first time it runs, it deletes any stored GitHub connection, its history, the old report queue and its scheduled jobs, once. A one-time privacy cleanup also clears older diagnostic logs if it has not already completed on this installation. Neither cleanup can remove copies that were already sent.
Android backup is turned off for this app, so none of this goes into your phone's cloud backup.
Autopilot and your acceptance rate
Autopilot is on by default (from 0.6.2, and turned on once then for earlier installs) until you turn it off. It uses your decision history, the wait-estimate records above and the acceptance rate Dasher shows to move its shared bar between 10% and 150% (from 0.6.32; between 50% and 150% before). From 0.6.29 it also follows each new positively confirmed personal standalone acceptance while on, whether or not acceptance learning is on: each readable active pay-per-stop, pay-per-mile and pay-per-hour working minimum moves independently up or down to that offer's value. From 0.6.31 it also keeps that offer's total pay, and in area mode a quarter of each Autopilot score weighs an offer's total pay against it. This happens between offers, at the screen reader's safe point. The latest acceptance replaces earlier accepted levels and starts the bar's next-24-offers hold again; earlier accepted scores do not override a newer lower acceptance. From 0.6.37 it also reads the decision history for offers you declined yourself since that acceptance: each holds the bar at least one point above its score under your current rules for the 24 counted offers after it. Nothing new is kept for this; the decline is the history's record of what was seen of the offer afterward. Later bar changes move these working minimums together. Turning Autopilot off uses exactly your saved and enabled learned minimums again; turning it on with the same goal restores its retained working adjustments. Autopilot never changes saved minimums or the learner's independent bests. Its earlier growth of saved minimums ended in 0.6.21.
- What is kept. Only the latest acceptance-rate reading: a whole percent, when Dasher showed it, and the offer it was shown for as a numeric fingerprint (pay in cents, miles, minutes, stops and, for shopping offers, the item count). It is kept whether Autopilot is on or off. A reading older than 7 days is never used or sent: it is deleted the next time the app looks at it (when you open the app, Autopilot plans or a report is built), and until then stays on the phone unused. A percentage counts only when exactly one stands beside acceptance-rate wording on Dasher's decline question; no text of the question is kept with it.
- When Dasher says declining an offer does not lower your acceptance rate, that offer's line in the decision history is marked so, and Autopilot leaves that offer out of its acceptance-rate count (unless Dasher says so of most of your recent declines, which Autopilot then takes as Dasher's general wording).
- Autopilot also keeps a note of its last change (time, old and new bar, a fixed reason), from 0.6.40 its last 40 changes of the bar for the chart's bar line (each its time, old and new bar, a fixed reason and whether a hold ended), when the bar last rose, a pending reason for its next change (a fixed word, such as that you turned it on) when it last noted that Dasher's "does not lower" wording looked general, and when you last changed your goal or rules (an accepted offer holds its bar only after that; before 0.6.19, turning it on or off counted too), and, while it is on, its working figures: whether it is recovering toward your goal, a correction of at most 10 points and, for that correction, the acceptance rate it carried forward from Dasher's reading at its last checkpoint (replaced after 25 more counted offers, and dropped when you change the goal or turn Autopilot off or on). All of this is numbers and fixed words, never screen text.
- The offers you accept yourself while Autopilot is on (from 0.6.19): for each, its pay in cents, miles, minutes, stops and item count, and when you accepted it, so an accept still counts after newer offers have pushed it out of the decision history. These legacy records supply the accepted-score hold only before a new independent working adjustment is active; its newest acceptance and anchor bar then replace their highest-score hold. Each old score can hold the legacy bar until 24 offers have come after it (from 0.6.21; before, for 7 days). At most 100, each for 7 days and only since you last changed your goal or rules; that change removes them. They never leave the phone, not even in a report. Numbers only.
- Independent working minimums (from 0.6.29): at most three exact numeric rate pairs, the latest acceptance's time and the bar it set (from 0.6.40 that offer's own Autopilot score as it showed, when known; otherwise, and before, the bar at that acceptance), and from 0.6.31 that acceptance's total pay as one more exact number (none for working minimums saved before 0.6.31). Unread or disabled measures keep their prior value, without becoming active. A confirmed update waiting for the safe point also keeps one pending copy with numeric consent and policy markers so a stale update cannot apply later; the live observation permit is memory only. These rates have no age expiry. A newer eligible acceptance replaces the observed active rates; changing your goal, minimums, max stops, mode or learning, Clear history or uninstall clears the working adjustments. Pausing or turning Autopilot off retains committed rates, the prior bar and an already confirmed pending update, without collecting new adjustments while off. The latest 200 decision snapshots retain the exact adjustments used for their original decision until normal eviction, Clear history or uninstall. From 0.6.44, with the current notice accepted, Autopilot uses the latest eligible confirmed personal acceptance in that retained history, its original recorded score, offer facts and numeric rules to rebuild its current working minimums and correct its bar at the next safe point. This includes eligible records already on the phone at the update and repairs after a restart; it changes no original decision snapshot and does not replay acceptance learning. Records from before a goal or rules reset, offers observed with Autopilot off or auto-decline paused, automatic acceptances, add-ons, skipped stores, misread offers and unconfirmed outcomes cannot supply that adjustment. The history must retain the original score and numeric rules; the older accepted-offer records described above are not enough. If the newest otherwise eligible acceptance lacks that record, no older acceptance replaces it. During the acceptance's original next-24-counted-offers hold, the bar stays at that score unless a newer personal decline holds it higher. Rebuilding does not start that hold again. Clear history removes the evidence, so it cannot rebuild cleared adjustments. No new observations are collected for this repair. The stored exact rates, the total pay and pending copies stay local and are not included in reports, automatic summaries or the timing study. Reports can identify a recorded score with the fixed category Autopilot, without exporting the working rates.
- If Autopilot raised your minimums before 0.6.21, the app keeps a note of that last growth: when, by how much, how many offers on how many days it rested on, your minimums and Autopilot's bar before and after, and whether its note on the homepage (with Undo) is still waiting. Undo removes it; tapping OK keeps it for Autopilot's details. Numbers only.
- The 24-hour logs note each new reading, and Autopilot's plans with the acceptance rate and the offer rate (offers an hour) they count with, as numbers, and the Undo of such a growth (your minimums before and after); as for other recognized screens, they can also hold the decline question's masked words with its percentage.
- Where it goes. Your acceptance rate leaves the phone only in a report you share (Export diagnostics), in masked diagnostics sent with feedback (with Share anonymous diagnostics on, or in the report offered after a stop), and in Feedback about a selected offer, which sends Autopilot's state and the acceptance rate it counts with. It is never in the summary after each dash: that summary carries Autopilot's bar and why it changed (never your acceptance rate), so a reason such as "acceptance rate below your goal" can show it was below your goal, never what it was. The fingerprint kept with a reading never leaves the phone. The mark that declining an offer does not lower your acceptance rate is part of that offer's history line and goes with it in Export diagnostics, the diagnostics sent with feedback and the summary after each dash; feedback about a selected offer leaves it out.
- Clear history removes the reading, Autopilot's working figures, the accepted offers it keeps, its independent working minimums and any pending adjustment, its last-change note and its kept changes, the note of its last growth (and with it the Undo) and its other notes above. Your rules and Autopilot's settings stay.
The acceptance learner and independent Autopilot minimums add no new recipient or automatic-sharing choice. Existing diagnostics can include scoring mode, learning on/off and the recorded score kind, alongside observed offer figures and per-decision required pay, and fixed-word lines saying whether each of your accepts taught the learner and, if not, why (no amounts). Raw learned-best snapshots and adoption receipts are not sent. The offer timing study continues to exclude minimum amounts and learner evidence.
What leaves the phone, and when
Feedback in Settings is the single reporting screen. Under About, choose General feedback or a recent offer. Send feedback on an offer's ticket opens the same screen with that offer selected; setup and stop notices also open it. The preview, Send action, submission status and last ten references are shared. Each selection keeps the scope below. Export diagnostics is an optional diagnostic export inside this screen and does not submit the feedback draft.
- Update checks. The app asks its public update server (dash-offer-filter-build.onrender.com, hosted by Render) for the latest version and downloads it. No account is required. These requests carry ordinary internet connection metadata such as your IP address and the app's own request information, but no offer history.
- General feedback in Feedback, only when you tap Send: the category you pick, what you type (at most 4,000 characters), and the app's version. With Share anonymous diagnostics on (from 0.6.34; before, each submission chose with Attach masked diagnostics), it also sends the masked diagnostic report with the offer timing study's section: the Android version and phone maker, the app's readiness, settings and last status line, the offer map's counts (never a position), your current rules and Autopilot's state, including the latest acceptance-rate reading, the decision history, the state of updates, the latest states, the bounded masked logs described above and any recent stop summary. Long diagnostics go in at most 4 parts; Preview shows exactly what will be sent. With the switch off, it sends only your words, the category and the app's version, and the dialog says which.
- Feedback about a selected offer, only when you tap Send: what went wrong (Misread, Wrong decline, Wrong accept or Other), your optional note, and that offer's report: its figures, decision and outcome, the bar it was judged at, its masked read lines (with every word that is not offer vocabulary reduced to its shape), your current rules and Autopilot's state (on or off, goal, bar, mode and last change, and the acceptance rate it counts with: Dasher's latest, carried forward, or its own estimate), the app's and Android's versions (not the phone's maker) and the minute it was decided. With Autopilot's state go whether it is recovering toward your goal, its correction and, for Dasher's own reading, how many minutes ago Dasher showed it. The dialog says what is sent; tapping Send is your choice to send it.
- Share anonymous diagnostics, only if you turn it on in Settings or with the switch at the end of the notice, "Share anonymous diagnostics (helps fix it for your phone)" (from 0.6.18; from 0.6.20 it shows your current choice, and accepting the notice leaves the option as that switch is left). From 0.6.34 this one switch covers the summary after each dash, the diagnostics with feedback you send (above) and the offer timing study, which were three choices before. It covers more than any of them did, so the update to 0.6.34 turns it off for everyone, whatever was chosen before, and no update turns it on. After each dash ends (Dasher shows the dash ended, you end it in Dasher with its End dash, or nothing of the dash was seen for 30 minutes), it sends one masked summary of at most 30,000 characters: the app and Android version and phone maker; how long the dash lasted (to the nearest 5 minutes) and how it ended; your switches and which rules are set, and whether screen reading, background offers and alerts were ready; Autopilot's bar and why it changed (never your acceptance rate); counts of offers by result, action and outcome; unreadable offers; declines still showing, by stage; decline-error recoveries; automatic accepts not sent, and why; the types of screen-read and notification errors with the place in the app's code where they happened; how Dasher's window was laid out; what Peek did, by fixed category; that dash's decisions, with times counted from the dash's start; masked log lines only around those problems (from 0.6.48 copied, on the phone, with that dash's counts once each problem's moment has passed, so they are not lost as the log rotates; they go when the summary is sent or discarded, or you clear history); any recent stop summary; and the offer timing study's rows for that dash. Never coordinates, place names, or an install or device identifier. At most one a dash and three a day. Turning it off discards summaries not yet sent and the timing study's data.
- After a crash or a stop for not responding: with Share anonymous diagnostics on, the stop's summary goes with the next summary after a dash. Otherwise the homepage shows "Offer Filter stopped unexpectedly last time" with Send feedback, which opens the same Feedback screen with the stop selected and Bug chosen; that report sends the masked diagnostics, the stop's summary among them, for it only, and nothing is sent unless you tap Send.
- Where feedback goes. All of the above go only to the project's own feedback service, a Supabase Edge Function reached through Cloudflare. Every submission carries a fresh random token, used only to tie its parts together and store each once; it is new for each submission and not tied to your phone. The service stores each part as sent, with its kind, category, app version, whether it came from the app or the website, and when it arrived: no account, name, email, IP address, user agent, install or device identifier. To limit how often one network can send, it counts recent submissions under a one-way value made with a secret key from the network's address and the date, so the value changes every day; the address itself is not stored. Rate-limit counts older than two days are deleted as new submissions arrive. Cloudflare and Supabase necessarily receive ordinary connection metadata, such as your IP address, while handling a request and may keep infrastructure logs under their own policies. The developer can see the request logs Supabase keeps for the project, which can record each request's IP address and the app or browser that sent it, for as long as Supabase keeps them. So this is not a guarantee of anonymity.
- Submissions are kept on the feedback service for 90 days, then deleted. The developer reads them, and may use AI systems (Anthropic's Claude or OpenAI's ChatGPT/Codex) to investigate problems with them. A submission does not guarantee review or a fix.
- Waiting and retrying. If the phone is offline, the service is busy or too many submissions came from your network recently, a submission waits on the phone (see above) and Android sends it later, once a network is available; the app also sends what waits when it next opens. Nothing is sent unless the current notice is accepted, and an after-dash summary only while Share anonymous diagnostics is still on; both are checked again before every part.
- Send an anonymous daily count, on unless you turn it off in Settings (on by default from 0.6.3, for new installs and installs updating alike). Once a UTC day at most, and only after you accept the current notice, the app sends the same feedback service one count, so the developer can see how many people use the app. It carries only: that it is the app's daily count; the app's version name and number; whether it would be this install's first count (it says so until the service has accepted one, so the totals can tell new installs from ones counted before; reinstalling, or clearing the app's data, starts over); and a random code made new for each UTC day. The code is sent again only to retry that day's count, so a retry is not counted twice; the next day's count has a new one, so no two days' counts can be tied together by it. Nothing else goes with it: no offers, rules, history, logs or message, and no account, device, Android version, phone maker, language, time zone or install or device identifier. If the phone is offline or the service is unavailable, the count is tried again later that day with the same code; if the service refuses it (its earlier version refuses every count), it is not sent again until the next UTC day. Turning the switch off stops it and discards that day's code; if that day's count was already tried, turning it back on sends nothing more until the next UTC day, since the first try may have arrived.
- The Download link. The Download link on offerfilter.org's install page, and the one on the update server's own page, go first to the same feedback service, which counts one download and at once sends your browser on to the app's file on the update server. The count says only that the app was downloaded. The service reads your browser's description (its user agent) and whether the request was made ahead of a tap only to leave out link previews, search robots and prefetches, and never keeps either. The page sets no cookie and stores nothing in your browser for it, and the download goes ahead whether or not the count is stored. Like any web request it carries your browser's ordinary request information (such as your IP address, the browser and the page's address), which Cloudflare and Supabase receive as described above.
- How counts are kept. The feedback service keeps counts as daily totals: for each UTC day, how many daily counts came for each app version and how many of those were an install's first, and how many downloads came. A total carries no IP address, user agent, code, network value or identifier, so it cannot tell one phone or person from another. Totals are deleted after 400 days. Each counted code is kept with its day and nothing else, for at most 2 days, only so a retried count is counted once. To count one network's downloads once a day, and to stop counting more than 50 daily counts from one network in a day (many phones can share one address), the service keeps, under the same day-keyed one-way network value the feedback rate limit uses, whether that network's download was counted today and how many daily counts came from it today, and deletes this once the day is over; no table records which network sent which code. It counts at most 20,000 a day in all. Cloudflare and Supabase receive each count's ordinary connection metadata, as described under Where feedback goes.
- Export diagnostics inside Feedback, only when you tap it, hands a masked report to the app you choose: readiness, settings, recent decisions and masked logs, with the latest acceptance-rate reading. Older decisions can be omitted to fit Android's sharing limit. From there, that app handles the copy under its own privacy practices.
- Place names. To name the squares on the map and the place you are in, the app asks Android's own place lookup (Google's servers on most phones) about positions rounded to about half a kilometre, the phone's own approximate position among them, once while a position remains cached. An evicted or cleared position may be looked up again. Nothing about your offers goes with them.
- Navigation, only when you tap it. Opening an offer area hands that historical area's center coordinates to your chosen Maps or Waze app, or to Google Maps in a browser. Gas and gas-price choices hand a search phrase to the map provider. A Back to map tap opens your navigation app the way its own icon does. The receiving app or website handles the request under its own privacy practices; these actions do not upload your offer history or establish your arrival.
- Help, only when you tap it in Settings: your browser opens the install and setup page on offerfilter.org, and that website receives the ordinary information any web request carries.
- The Jesus Loves You signature at the end of Settings, only when you tap it: your browser opens jesuslovesyou.xyz, and that website receives the ordinary information any web request carries. The app sends nothing with the tap.
- Tips. The Cash App, Venmo and PayPal links open only when you tap them; the app sends nothing for them and counts nothing.
- Web and email addresses in these texts open in your browser or email app only when you tap them.
Render, Cloudflare, Supabase, Google, Anthropic and OpenAI, and the email services that carry a message you send to the privacy contact below, handle what reaches them under their own terms and privacy policies. Copies you deliberately share through another app are controlled by that app and recipient.
What it never collects
The app does not request your DoorDash password, contacts, photos, messages, precise location or payment details. Accessibility can expose unrelated text when Dasher displays it; recognized payment, account and earnings screens are discarded, and the masking described above is applied before diagnostic text is saved.
Your choices
- Turn off Peek at background offers in Settings to stop automatic temporary opens of Dasher. Peek requires an unlocked, quiet phone (an offer that arrived while it was locked is looked at only just after you unlock, within 40 seconds of it) and skips when its checks find typing, a call, microphone/camera use, split or floating windows, a pinned or unrecognized app, or another conflicting action. Its checks depend on what Android exposes; they are not a safety guarantee.
- During an active dash, with auto-decline on and a rule set, the screen is kept from timing out whatever app is in front, but only while something of the dash was seen in the last 15 minutes or Dasher shows it, and never with battery saver on or the battery at 20% or less and not charging. It never wakes or unlocks the phone, and the power button still turns the screen off; pausing auto-decline ends the hold. Peek pauses while your phone is locked.
- The scoring mode (Area by default, or Strict) is the word in the radar's middle on the homepage; acceptance learning is the Learn button on the homepage (a long press shows what it learned). Turning learning off retains bests; confirmed Reset or Clear history clears them. Neither turns on Share anonymous diagnostics or its timing study.
- Autopilot is on by default; turning it off from its chip on the homepage uses your saved and enabled learned minimums again. It retains committed working adjustments for when you turn it back on, and records no new ones while off. The acceptance-rate reading is kept whether or not Autopilot is on; Clear history removes it and the working adjustments. Autopilot never writes the saved minimums or learned bests; enabled acceptance learning can still raise its own bests independently.
- Clear history (Settings) removes the decisions, both logs and the latest states, the observed-wait records, the offer map and cached place names, the temporary restart/error records above, stop summaries, the current dash's problem counts and after-dash summaries not yet sent, plus Autopilot's acceptance-rate reading, its working figures, its last-change note, the note of its last growth and its other notes, restored learned bests, the learner deduplication record and adoption Undo. User-adopted saved minima remain settings. Lookups and wait-record writes already in progress cannot restore cleared history. Feedback and offer reports you sent and that still wait to go stay until they are sent or 7 days pass. Your rules, Autopilot's settings and the app's last status line (which can show the last offer's figures until the next status replaces it) remain.
- Feedback and offer reports leave only after you tap Send. With Share anonymous diagnostics off, general feedback sends only your words, the category and the app's version; a selected offer or the report offered after a stop keeps its disclosed scope above.
- Share anonymous diagnostics is off unless you turn it on, and you can turn it off at any time: summaries not yet sent and the timing study's data go with it.
- Send an anonymous daily count is on unless you turn it off in Settings; off, the app sends no count. Counts already made stay in the service's daily totals, which cannot tell which phone sent them, until those are deleted after 400 days.
- Uninstalling removes the app's local data, including submissions still waiting and the kept references. It does not remove submissions the feedback service already received, the daily totals counts went into, or copies you deliberately shared through another app.
Help and privacy contact
Use Feedback in Settings, or the feedback form on offerfilter.org, for non-sensitive feedback without creating an account. Do not type customer, payment, account or other sensitive details. If you need to refer to an earlier submission, keep the short reference shown after it is accepted; Settings' feedback dialog lists the last ten.
For privacy, data-deletion and security requests only, email privacy@offerfilter.org. Unlike feedback, email is not anonymous: the developer sees your address and uses it only to answer you. Your message passes through your own email provider and the email services that deliver mail for offerfilter.org to the developer. Feedback carries no name or account, so to ask about a submission you sent, include its reference; without it the developer may be unable to tell which anonymous submission is yours. Send everything else as anonymous feedback.
The project's public issue tracker is suitable only for public, non-sensitive development discussion; posts there are public. Do not attach diagnostic reports, customer details, payment or account information, addresses, screenshots containing personal information, or tokens.
Children
Offer Filter is not for anyone under 18.
Changes
When this text changes in substance, the app shows its notice again before it reads anything more or sends feedback or its daily count.
This website
The sections above describe the app. This one covers offerfilter.org itself.
- Hosting. The site is a set of static pages hosted by GitHub Pages. Like any web host, GitHub receives the usual request information (your IP address, browser details and the page you asked for) and may keep it in server logs under GitHub's own privacy statement. The site has no accounts and no sign-in.
- No cookies, no analytics. The site sets no cookies and runs no analytics, ads or tracking. Its pages, fonts, images and film come from offerfilter.org itself. The home page remembers three small settings in your browser's local storage, and they never leave your device: the sky you chose with the sky button (Day, Night, System or Auto), whether the film's captions are on, and whether your device draws the page's 3D graphics in software, so the next visit starts lighter. Clearing this site's data in your browser removes them.
- Feedback form. Nothing is sent until you tap Send. Then the site sends the type you chose and the message you typed, marked as coming from the website, to the Offer Filter feedback service: a Supabase Edge Function, reached through Supabase's network (which includes Cloudflare). The service turns your connection's IP address into a rotating, one-way rate-limit value and does not store the IP address with your feedback. Supabase and Cloudflare handle ordinary connection data under their own policies, and the developer can see the request logs Supabase keeps for the project, which can record each request's IP address and browser details, for as long as Supabase keeps them. Feedback is deleted after 90 days.
- Who reads feedback. Feedback may be reviewed by the developer, with the help of AI assistants such as Anthropic's Claude or OpenAI's ChatGPT/Codex, to investigate problems. There is no reply address, so we can't answer you directly. Please leave out customer names, addresses, payment, account or password details.
- Download. The download button goes first to the Offer Filter feedback service, which counts the download as described above and sends your browser on to the app's file on its update server on Render (dash-offer-filter-build.onrender.com); both receive ordinary connection information. The install page reads the current version number from offerfilter.org.
- Tips and other links. The tip links open Cash App or Venmo; the site sends them nothing and counts nothing. Source-code links open GitHub, and the emblem at the foot of the home page opens jesuslovesyou.xyz. Those sites handle your visit under their own policies.
- Contact. For privacy, data deletion or security requests, email privacy@offerfilter.org. Unlike feedback, email isn't anonymous: the developer sees your address and uses it only to answer you. To ask about feedback you sent, include its reference: feedback has no name or account, so the reference is how it can be found. Please send everything else as anonymous feedback.