{
 "../meeting_en/recordings/meeting_part1.m4a:en": {
  "status": 200,
  "secs": 25.0,
  "text": "Hey everybody, good to see actual faces instead of little loading circles today. Jackson, that plant behind you has gotten enormous since our last sync. Either you finally watered it or you borrowed a better plant for the camera. Before we get serious, does anybody have a small personal victory from the last two weeks? Mine is that my coffee maker stopped making the noise of a departing helicopter. The plant is real, and the secret is that I moved it away from the heater. Apparently plants prefer not to live inside a hairdryer. My personal victory is making it through an entire grocery trip without checking an error alert. I did check the dashboard in the parking lot afterward, but I am choosing to count that as progress rather than inventing a stricter scoring system. I successfully explained my internship to my grandmother without saying machine learning, which felt like an achievement. I told her I help a computer understand people when they talk. She immediately asked why her television understands everyone except her. So apparently I have acquired a second evaluation project, and the stakeholder has strong opinions about the remote control and absolutely no patience for a discovery interview. That is a very good stakeholder, honestly. My weekend achievement was assembling a bookshelf without having a mysterious handful of extra screws afterward. Caitlin, I also tried the little bakery you recommended. You were right about the bread, but you undersold how long the line gets. I had enough time to mentally redesign their ordering system before I reached the counter, which probably means I need a hobby. I deliberately left out the line because I wanted you to go, that is probably not a product principle we should carry into voice flow. My win is less exciting. I finally cleared the stack of unopened mail next to my door. Half of it was offers to help me organize my life, which felt a little aggressive coming from the actual source of the mess. All right, excellent evidence that we are functioning adults in at least a few settings. I mean, the bar is not high, but we cleared it. Today is our bi-weekly voice flow sync, and I want decisions rather than a guided tour through every ticket. We will cover launch readiness, deployment costs, a short investor check-in, then design and AFIS evaluation work before we lock owners and dates. Saoirse is traveling and listening from her phone, so we will take infrastructure before the break. Kaveh will join partway through for a quick conversation about the business metrics. After that, we will take five minutes, then come back for the homepage, settings, and name recognition. I mean, those last topics are connected more closely than they look, so please keep the customer experience in view. One housekeeping thing before Jackson starts. The launch document is the version titled controlled rollout, not the older general availability draft. I put the link in the meeting notes. Priya has the same version at the partner company, and she knows we are deciding the next cohort today. If somebody opens the old document and sees a promise of unlimited usage, please close it before your blood pressure changes. On launch status, the core dictation path is healthy. Intake accepts the audio, creates the job, and hands it to the AI builder space transcription API. We then apply our formatting rules and return the draft to VoiceFlow. The failure rate is down substantially from the previous sync. What remains embarrassing is proper names. Ordinary sentences look clean, and then somebody introduces themselves, and the transcript invents a completely different person. That matches the intake interviews. People are willing to forgive a missing comma, but they notice immediately when their own name is wrong. Jackson became Jackson in three sessions, and Sochitel got three different spellings in one conversation. Apha had an even stranger substitution. The larger issue is that users interpret those mistakes as evidence that we do not understand their workplace, even when the rest of the dictation is useful. Let us separate availability from quality before those become one giant red status dot. Jackson, are we saying the system is reliable enough for the next cohort, but not accurate enough to invite everybody? I mean, that sounds manageable if we are honest about the boundary. I do not want us holding a small, useful release hostage to a standard we have not actually written down. Yes, that is the distinction. Reliability is good enough for a controlled expansion. Name handling is not good enough for an unqualified launch claim. We also fixed the duplicate job problem. If a browser retries during a network interruption, intake now reuses the existing job identifier. That means we no longer pay for two transcriptions just because someone walks between rooms and their wireless connection briefly becomes a philosophical concept. Can I check what counts as a name failure in the launch document? If someone says Caitlin and we output Caitlin, that sounds identical but is still wrong for the person. If someone says Cave and we output coffee, that is wrong in a different way. My clips distinguish spelling errors from substitutions, and I would rather not collapse them before we know which problem the next cohort actually experiences. Keep that distinction, Aoife. For the launch gate, either one counts as a failed exact name match, but the breakdown belongs beside the headline number. We should also record whether a user can correct it once and have that correction stick. A spelling mistake that takes one correction is different from a mistake that returns every morning, like an extremely persistent visitor who has learned nothing from previous conversations. The correction flow currently makes that second problem worse. Someone fixes a name in the transcript, but nothing tells them whether VoiceFlow has learned anything. The interface implies confidence without explaining scope. I can change the confirmation so it says the correction applies to this transcript, unless we actually save a preferred spelling. We should not let a small animation quietly make a promise the backend cannot keep. Exactly. Right now, editing text changes that document only. We have a workspace vocabulary table behind a feature flag, but we are not consistently sending those entries through to the AI builder space transcription API. Priya confirmed the supported hint format yesterday. I need to wire that into intake and add a size limit because sending somebody's entire employee directory with every short recording would be a bad default. That sounds like a practical lever. Jackson, please do not call it learning in the interface if we are really passing a vocabulary hint. I mean, the distinction matters when a user asks why a correction worked in one workspace and disappeared in another. Caitlin, can we make the next cohort explicitly test that boundary rather than just collecting another general satisfaction score that tells us everybody has feelings? Yes, I would like ten additional workspaces with at least two regular dictation users in each. We can ask each workspace for a small optional list of frequently spoken names, explain where that list is stored, and compare their first session with their third. Priya can recruit four of those workspaces from her partner network. The remaining six should come from our existing waitlist, so the cohort is not entirely friendly colleagues. Please include at least one workspace where people use names from several language backgrounds in the same recording. A clean English sentence containing Sochital, Sarsha, and Kaveh is a useful case, not an unusual edge case we should discard. I also want a few speakers who say their own name naturally, instead of reading it very carefully because they know a model is being tested and want to help it succeed. There is another intake issue that is smaller but worth saying out loud. We preserve audio until the job finishes, then follow the workspace retention setting. We were not showing that setting clearly during onboarding. Nothing is secretly keeping recordings forever, but the explanation is buried. I want the launch checklist to include a test that the displayed policy matches what the deletion worker actually does for that workspace. I can make that explanation visible without turning onboarding into a consent maze. A short summary beside the recording controls would help, with a link to the full setting. Jackson, I will need plain language for the difference between raw audio and the transcript. People assume deleting one deletes the other. If that is not always true, the screen has to say so before they make the choice. For launch messaging, I am proposing fast dictation with editable transcripts, not perfect transcription, and definitely not automatic understanding of your entire company. We can mention custom vocabulary only once the hint path is working and measured. Priya asked whether AI builder space needs to appear in the onboarding copy. My instinct is to put the provider information in the privacy explanation rather than crowding the first screen with infrastructure names. That works as long as the disclosure is easy to find and specific enough to be meaningful. We are not hiding a provider, and we are not making people read an architecture diagram before they dictate a sentence. Jackson, what would actually stop the cohort expansion this week? Give us the short list of things that would make you say no, even if everybody else were excited. A repeatable loss of audio before processing completes would stop it. So would cross-workspace access to a transcript, obviously, or a billing path that submits the same recording repeatedly. Those checks are passing right now. A slower response under a burst would make me reduce the cohort or add a queue limit, not automatically cancel. Incorrect names are a quality gate, but I need IFA's measured baseline before I propose a responsible threshold. Then the launch status is limited expansion pending the safety checks and the evaluation readout, with no public announcement yet. I will update the document in exactly those terms. I also want one shared support label for name problems. Right now they are scattered across accuracy, onboarding, and something called weird output, which is honest, but not particularly useful when we try to count the impact across customers. Good. I mean, weird output is a mood, not a taxonomy. Let us give support categories that map to decisions. Wrong spelling, wrong person or phrase, correction not retained, and vocabulary hint apparently ignored. IFA, you can tell us later if that split is technically sensible. Caitlin, make sure Priya gets the same labels, so partner reports do not arrive as a separate pile of anecdotes. That split works, with one caution: apparently ignored is an observation, not proof that the provider ignored the hint. The audio might be ambiguous, or the hint might have been malformed before it reached the provider. I can label the user visible symptom and attach the request metadata separately. That should let Jackson investigate without forcing support to make an inference they cannot actually verify from a screenshot. I can add a safe trace field that records whether vocabulary was attached, how many entries it contained, and which format version we used. It will not contain the actual names. That should help us distinguish a missing hint from an ineffective hint while keeping the logs boring in the best possible way. I will coordinate the field names with Priya so her team can send us comparable reports. Sarsha's on the phone. Sarsha, go ahead. Before you start, we can hear some station noise, but your voice is coming through clearly. Give us the bill change first, then what caused it, then the move you are proposing. After that, Jackson can ask about the numbers, and we will decide whether Thursday is a real migration window or just an optimistic calendar decoration that makes everybody feel better. Thanks, Aish. Was that minimum instance in the service configuration, or did the health check keep waking it? I ask because I remember changing the readiness route when we added the re-ranker. If the platform is following a setting we gave it, I want us to fix this setting. If we are accidentally generating activity, moving the same behavior to another machine will just make the waste less visible. That is a helpful distinction. I mean, we should not solve an availability requirement by accidentally applying it to everything with a network port. Sarsa, tell us which services moves, what stays where it is, and the actual day you want to do it. I also want the fallback stated clearly enough that Caitlin can explain it to a customer without borrowing Jackson's entire vocabulary for the afternoon. That is a helpful distinction. I mean, we should not solve an availability requirement by accidentally applying it to everything with a network port. Sarsa, tell us which services moves, what stays where it is, and the actual day you want to do it. I also want the fallback stated clearly enough that Caitlin can explain it to a customer without borrowing Jackson's entire vocabulary for the afternoon. We are moving the small text model that scores vocabulary candidates after transcription. If the new pass misbehaves, a feature flag routes requests back to the existing COIN service, which stays available during the observation window. That switch does not require a new client release. That last point matters. I can explain a backend move with an immediate fallback, but I do not want the cohort hearing that we changed the whole dictation engine. Seriously, will there be any visible maintenance period? Priya is scheduling onboarding sessions for Thursday morning, and I would rather move those now than discover that our careful internal wording failed to mention an actual interruption for her customers. There should be no planned interruptions because the flag changes the rerank for destination while the old service remains healthy. I want to start at 10 in the morning Pacific, after Freya's earliest onboarding session, and send only internal workstations first. The box has enough memory for the current model and our measured concurrency, but that is not permission to send every queue job at once. I will cap each light reranker request, watch the queue, and expand gradually. If the queue grows continuously for 10 minutes, we roll back rather than improvise. Can we get the performance numbers behind that confidence? Specifically, I want the KV cache hit rate, what workload produced it, and whether the prefill time includes waiting in the queue. We have had a few dashboard numbers that looked wonderful because the timer started after all the painful waiting was over. I would prefer an uglier number that reflects what the person dictating actually experiences on their screen. On the scraping replay, the keying cassette rate was 68%, with prefix tracking enabled, premier with 21% in the earlier configuration. Medium prefill times fell from 410 milliseconds to 170 milliseconds. Those prefill measurements exclude key weights, so I am not presenting them as end-to-end latency. Medium key weight was 30 milliseconds, and the 96th percentile was 140 milliseconds at our replay load. The replay used standardized prompt structures and synthetic vocabulary lists, not production recordings or custom releases. I want to understand whether this 68% result survives the next cohort. If the replay accidentally sends the same short list over and over, the KV cache could look much better than it will when Priya brings in several workspaces with different names and people start editing those lists between sessions. Good. Let us carry the conservative number forward. I mean, a cache hit rate is only useful if we remember what was hitting the cache. Sasha, name the people you need, because travel should not turn this into you debugging from a train platform alone. Jackson, I am assuming you can own the request routing, but I want you to confirm before I volunteer your entire Thursday morning. I need Jackson for the routing plan, timeout handling, and a tested rollback scenario. I need Ava for that adversarial replay and a quick accuracy comparison so the faster path does not quietly change races. I need Priya to confirm her session schedule and send one representative sanitized request case. Caitlin can coordinate that introduction. I will own deployment, dashboards, and the cost comparison. Lupito does not need to redesign anything for this move. If the users use a new screen because of it, we have probably taken an unnecessary detour. Confirmed. I can pair with you on the flag Wednesday and be online Thursday for the rollout. I would also like to test a dead connection to the box, not just a clean error response. A hanging request is how a supposedly optional re-ranker turns into a transcription outage. We should return the original transcript when the re-ranker times out, then record that degradation without retrying forever in the background. I will contact Priya right after this and get the session schedule into Sarah's rollout note. For the customer-facing description, I will say, transcripts remain available if the optional vocabulary pass is temporarily unavailable. Jackson, please confirm that wording against the actual timeout behavior. I would rather promise a slightly less polished transcript than imply the whole product is down when only the extra name ranking pass has been skipped. That is the plan. I mean, the headline is not that we discovered a cheaper machine. It is that we now understand the idle service, the fallback, and the real cost boundary. Seriously, stay for anything you can, but do not miss your train on our behalf. We will repeat your owners and dates at the end, and nobody should treat a dropped phone connection as approval for a risky change. Kaveh just joined, so let me bring him in briefly. Kaveh, welcome. We have covered launch readiness and a targeted infrastructure move, and we are about to get into product quality. You asked for a few minutes on retention and what evidence would matter for the next round. We have about four minutes now, so give us the practical version. Thanks. The practical version is that I care less about a big registration spike than about whether people return after their first useful transcript. Retention should tell us that voice flow becomes part of a workflow. Please separate users who tried one demo from people who actually had another reason to dictate and chose you again that week. That is how I'm framing the next cohort. We will track a first successful dictation, then a return on a different day, and then recurring use across the following weeks. We are not planning to count opening an onboarding email as product activity, even if that would make the chart look friendlier. Good. Daily actives are useful if the activity means something consistent. I would define an active user around a completed dictation, with a reasonable minimum so an accidental button tap does not count. Then show daily actives alongside weekly actives. Some valuable users dictate twice a week, and a daily chart alone can make that healthy behavior look disappointing. That fits the jobs we are hearing about. I mean, a person dictating meeting notes on Tuesdays and Thursdays is not necessarily less successful than somebody sending five tiny messages every day. We will make the event definition explicit and segment by workflow instead of quietly rewarding whichever behavior produces the largest number. For the next funding round, the question will be whether usage repeats, whether those users can become paying customers, and whether serving them has sensible economics. A low infrastructure bill is nice, but it is not a substitute for retention. Likewise, strong retention does not excuse a cost curve that gets worse with every additional minute of audio. I would also keep the cohorts visible. If the newest cohort retains better after the name fixes, that is a useful story about learning and execution. If all the improvement comes from mixing new users with a few very loyal testers, it is harder to interpret. Small samples are fine at this stage. Hidden denominators are not. Understood. I will include the cohort sizes, the activation definition, and the usage window on the same page as the charts. For costs, we will separate transcription provider charges from the vocabulary pass and shared infrastructure. Sasha already warned us not to describe an owned machine as free, which seems especially relevant to this discussion. Exactly, and when you talk to users who stopped, ask what interrupted the habit. Was it accuracy, speed, trust, or simply no recurring meet? A short, honest explanation from five people can be more actionable than a broad satisfaction average. I would be especially interested in whether repeated name corrections are enough to make someone abandon an otherwise useful product. I need to drop now, but this is the right level of focus. Send me the cohort readout once it has actual observations, not a prettier forecast. I am happy to review the metric definitions beforehand if that saves rework. Thanks, everyone, and Saoirse, safe travels. I will leave you to the product decisions rather than taking the rest of your meeting. Let's take a five-minute break now, and when we return, we will start with Xochit's design refresh before moving into IFA's evaluation results and our final assignments. Alright, welcome back everybody. Satchel, please take us through the homepage button first, then the color mismatch with AI Builder Space, and finally the settings page. The homepage has three things that look like the primary action, and only one of them starts dictation. The large button says create, the microphone tile says capture, and the small button in the header says new session. That is too much vocabulary for the same basic intent. I want one main button labeled start dictation, with the microphone icon supporting the words rather than replacing them entirely. Does start dictation request microphone access immediately, or does it open the recording screen first? Right now, the browser permission prompt appears after the second click. I would love to",
  "duration": 1801.57,
  "segments": [
   [
    2.3,
    1801.39,
    "Heyeverybody,goodtoseeactualfacesinsteadoflittleloadingcirclestoday.Jackson,thatplantbehindyouhasgottenenormoussinceourlastsync.Eitheryoufinallywatereditoryouborrowedabetterplantforthecamera.Beforewegetserious,doesanybodyhaveasmallpersonalvictoryfromthelasttwoweeks?Mineisthatmycoffeemakerstoppedmakingthenoiseofadepartinghelicopter.Theplantisreal,andthesecretisthatImoveditawayfromtheheater.Apparentlyplantsprefernottoliveinsideahairdryer.Mypersonalvictoryismakingitthroughanentiregrocerytripwithoutcheckinganerroralert.Ididcheckthedashboardintheparkinglotafterward,butIamchoosingtocountthatasprogressratherthaninventingastricterscoringsystem.Isuccessfullyexplainedmyinternshiptomygrandmotherwithoutsayingmachinelearning,whichfeltlikeanachievement.ItoldherIhelpacomputerunderstandpeoplewhentheytalk.Sheimmediatelyaskedwhyhertelevisionunderstandseveryoneexcepther.SoapparentlyIhaveacquiredasecondevaluationproject,andthestakeholderhasstrongopinionsabouttheremotecontrolandabsolutelynopatienceforadiscoveryinterview.Thatisaverygoodstakeholder,honestly.Myweekendachievementwasassemblingabookshelfwithouthavingamysterioushandfulofextrascrewsafterward.Caitlin,Ialsotriedthelittlebakeryyourecommended.Youwererightaboutthebread,butyouundersoldhowlongthelinegets.IhadenoughtimetomentallyredesigntheirorderingsystembeforeIreachedthecounter,whichprobablymeansIneedahobby.IdeliberatelyleftoutthelinebecauseIwantedyoutogo,thatisprobablynotaproductprincipleweshouldcarryintovoiceflow.Mywinislessexciting.Ifinallyclearedthestackofunopenedmailnexttomydoor.Halfofitwasofferstohelpmeorganizemylife,whichfeltalittleaggressivecomingfromtheactualsourceofthemess.Allright,excellentevidencethatwearefunctioningadultsinatleastafewsettings.Imean,thebarisnothigh,butweclearedit.Todayisourbi-weeklyvoiceflowsync,andIwantdecisionsratherthanaguidedtourthrougheveryticket.Wewillcoverlaunchreadiness,deploymentcosts,ashortinvestorcheck-in,thendesignandAFISevaluationworkbeforewelockownersanddates.Saoirseistravelingandlisteningfromherphone,sowewilltakeinfrastructurebeforethebreak.Kavehwilljoinpartwaythroughforaquickconversationaboutthebusinessmetrics.Afterthat,wewilltakefiveminutes,thencomebackforthehomepage,settings,andnamerecognition.Imean,thoselasttopicsareconnectedmorecloselythantheylook,sopleasekeepthecustomerexperienceinview.OnehousekeepingthingbeforeJacksonstarts.Thelaunchdocumentistheversiontitledcontrolledrollout,nottheoldergeneralavailabilitydraft.Iputthelinkinthemeetingnotes.Priyahasthesameversionatthepartnercompany,andsheknowswearedecidingthenextcohorttoday.Ifsomebodyopenstheolddocumentandseesapromiseofunlimitedusage,pleasecloseitbeforeyourbloodpressurechanges.Onlaunchstatus,thecoredictationpathishealthy.Intakeacceptstheaudio,createsthejob,andhandsittotheAIbuilderspacetranscriptionAPI.WethenapplyourformattingrulesandreturnthedrafttoVoiceFlow.Thefailurerateisdownsubstantiallyfromtheprevioussync.Whatremainsembarrassingispropernames.Ordinarysentenceslookclean,andthensomebodyintroducesthemselves,andthetranscriptinventsacompletelydifferentperson.Thatmatchestheintakeinterviews.Peoplearewillingtoforgiveamissingcomma,buttheynoticeimmediatelywhentheirownnameiswrong.JacksonbecameJacksoninthreesessions,andSochitelgotthreedifferentspellingsinoneconversation.Aphahadanevenstrangersubstitution.Thelargerissueisthatusersinterpretthosemistakesasevidencethatwedonotunderstandtheirworkplace,evenwhentherestofthedictationisuseful.Letusseparateavailabilityfromqualitybeforethosebecomeonegiantredstatusdot.Jackson,arewesayingthesystemisreliableenoughforthenextcohort,butnotaccurateenoughtoinviteeverybody?Imean,thatsoundsmanageableifwearehonestabouttheboundary.Idonotwantusholdingasmall,usefulreleasehostagetoastandardwehavenotactuallywrittendown.Yes,thatisthedistinction.Reliabilityisgoodenoughforacontrolledexpansion.Namehandlingisnotgoodenoughforanunqualifiedlaunchclaim.Wealsofixedtheduplicatejobproblem.Ifabrowserretriesduringanetworkinterruption,intakenowreusestheexistingjobidentifier.Thatmeanswenolongerpayfortwotranscriptionsjustbecausesomeonewalksbetweenroomsandtheirwirelessconnectionbrieflybecomesaphilosophicalconcept.CanIcheckwhatcountsasanamefailureinthelaunchdocument?IfsomeonesaysCaitlinandweoutputCaitlin,thatsoundsidenticalbutisstillwrongfortheperson.IfsomeonesaysCaveandweoutputcoffee,thatiswronginadifferentway.Myclipsdistinguishspellingerrorsfromsubstitutions,andIwouldrathernotcollapsethembeforeweknowwhichproblemthenextcohortactuallyexperiences.Keepthatdistinction,Aoife.Forthelaunchgate,eitheronecountsasafailedexactnamematch,butthebreakdownbelongsbesidetheheadlinenumber.Weshouldalsorecordwhetherausercancorrectitonceandhavethatcorrectionstick.Aspellingmistakethattakesonecorrectionisdifferentfromamistakethatreturnseverymorning,likeanextremelypersistentvisitorwhohaslearnednothingfrompreviousconversations.Thecorrectionflowcurrentlymakesthatsecondproblemworse.Someonefixesanameinthetranscript,butnothingtellsthemwhetherVoiceFlowhaslearnedanything.Theinterfaceimpliesconfidencewithoutexplainingscope.Icanchangetheconfirmationsoitsaysthecorrectionappliestothistranscript,unlessweactuallysaveapreferredspelling.Weshouldnotletasmallanimationquietlymakeapromisethebackendcannotkeep.Exactly.Rightnow,editingtextchangesthatdocumentonly.Wehaveaworkspacevocabularytablebehindafeatureflag,butwearenotconsistentlysendingthoseentriesthroughtotheAIbuilderspacetranscriptionAPI.Priyaconfirmedthesupportedhintformatyesterday.Ineedtowirethatintointakeandaddasizelimitbecausesendingsomebody'sentireemployeedirectorywitheveryshortrecordingwouldbeabaddefault.Thatsoundslikeapracticallever.Jackson,pleasedonotcallitlearningintheinterfaceifwearereallypassingavocabularyhint.Imean,thedistinctionmatterswhenauseraskswhyacorrectionworkedinoneworkspaceanddisappearedinanother.Caitlin,canwemakethenextcohortexplicitlytestthatboundaryratherthanjustcollectinganothergeneralsatisfactionscorethattellsuseverybodyhasfeelings?Yes,Iwouldliketenadditionalworkspaceswithatleasttworegulardictationusersineach.Wecanaskeachworkspaceforasmalloptionallistoffrequentlyspokennames,explainwherethatlistisstored,andcomparetheirfirstsessionwiththeirthird.Priyacanrecruitfourofthoseworkspacesfromherpartnernetwork.Theremainingsixshouldcomefromourexistingwaitlist,sothecohortisnotentirelyfriendlycolleagues.Pleaseincludeatleastoneworkspacewherepeopleusenamesfromseverallanguagebackgroundsinthesamerecording.AcleanEnglishsentencecontainingSochital,Sarsha,andKavehisausefulcase,notanunusualedgecaseweshoulddiscard.Ialsowantafewspeakerswhosaytheirownnamenaturally,insteadofreadingitverycarefullybecausetheyknowamodelisbeingtestedandwanttohelpitsucceed.Thereisanotherintakeissuethatissmallerbutworthsayingoutloud.Wepreserveaudiountilthejobfinishes,thenfollowtheworkspaceretentionsetting.Wewerenotshowingthatsettingclearlyduringonboarding.Nothingissecretlykeepingrecordingsforever,buttheexplanationisburied.Iwantthelaunchchecklisttoincludeatestthatthedisplayedpolicymatcheswhatthedeletionworkeractuallydoesforthatworkspace.Icanmakethatexplanationvisiblewithoutturningonboardingintoaconsentmaze.Ashortsummarybesidetherecordingcontrolswouldhelp,withalinktothefullsetting.Jackson,Iwillneedplainlanguageforthedifferencebetweenrawaudioandthetranscript.Peopleassumedeletingonedeletestheother.Ifthatisnotalwaystrue,thescreenhastosaysobeforetheymakethechoice.Forlaunchmessaging,Iamproposingfastdictationwitheditabletranscripts,notperfecttranscription,anddefinitelynotautomaticunderstandingofyourentirecompany.Wecanmentioncustomvocabularyonlyoncethehintpathisworkingandmeasured.PriyaaskedwhetherAIbuilderspaceneedstoappearintheonboardingcopy.Myinstinctistoputtheproviderinformationintheprivacyexplanationratherthancrowdingthefirstscreenwithinfrastructurenames.Thatworksaslongasthedisclosureiseasytofindandspecificenoughtobemeaningful.Wearenothidingaprovider,andwearenotmakingpeoplereadanarchitecturediagrambeforetheydictateasentence.Jackson,whatwouldactuallystopthecohortexpansionthisweek?Giveustheshortlistofthingsthatwouldmakeyousayno,evenifeverybodyelsewereexcited.Arepeatablelossofaudiobeforeprocessingcompleteswouldstopit.Sowouldcross-workspaceaccesstoatranscript,obviously,orabillingpaththatsubmitsthesamerecordingrepeatedly.Thosechecksarepassingrightnow.Aslowerresponseunderaburstwouldmakemereducethecohortoraddaqueuelimit,notautomaticallycancel.Incorrectnamesareaqualitygate,butIneedIFA'smeasuredbaselinebeforeIproposearesponsiblethreshold.Thenthelaunchstatusislimitedexpansionpendingthesafetychecksandtheevaluationreadout,withnopublicannouncementyet.Iwillupdatethedocumentinexactlythoseterms.Ialsowantonesharedsupportlabelfornameproblems.Rightnowtheyarescatteredacrossaccuracy,onboarding,andsomethingcalledweirdoutput,whichishonest,butnotparticularlyusefulwhenwetrytocounttheimpactacrosscustomers.Good.Imean,weirdoutputisamood,notataxonomy.Letusgivesupportcategoriesthatmaptodecisions.Wrongspelling,wrongpersonorphrase,correctionnotretained,andvocabularyhintapparentlyignored.IFA,youcantelluslaterifthatsplitistechnicallysensible.Caitlin,makesurePriyagetsthesamelabels,sopartnerreportsdonotarriveasaseparatepileofanecdotes.Thatsplitworks,withonecaution:apparentlyignoredisanobservation,notproofthattheproviderignoredthehint.Theaudiomightbeambiguous,orthehintmighthavebeenmalformedbeforeitreachedtheprovider.Icanlabeltheuservisiblesymptomandattachtherequestmetadataseparately.ThatshouldletJacksoninvestigatewithoutforcingsupporttomakeaninferencetheycannotactuallyverifyfromascreenshot.Icanaddasafetracefieldthatrecordswhethervocabularywasattached,howmanyentriesitcontained,andwhichformatversionweused.Itwillnotcontaintheactualnames.Thatshouldhelpusdistinguishamissinghintfromanineffectivehintwhilekeepingthelogsboringinthebestpossibleway.IwillcoordinatethefieldnameswithPriyasoherteamcansenduscomparablereports.Sarsha'sonthephone.Sarsha,goahead.Beforeyoustart,wecanhearsomestationnoise,butyourvoiceiscomingthroughclearly.Giveusthebillchangefirst,thenwhatcausedit,thenthemoveyouareproposing.Afterthat,Jacksoncanaskaboutthenumbers,andwewilldecidewhetherThursdayisarealmigrationwindoworjustanoptimisticcalendardecorationthatmakeseverybodyfeelbetter.Thanks,Aish.Wasthatminimuminstanceintheserviceconfiguration,ordidthehealthcheckkeepwakingit?IaskbecauseIrememberchangingthereadinessroutewhenweaddedthere-ranker.Iftheplatformisfollowingasettingwegaveit,Iwantustofixthissetting.Ifweareaccidentallygeneratingactivity,movingthesamebehaviortoanothermachinewilljustmakethewastelessvisible.Thatisahelpfuldistinction.Imean,weshouldnotsolveanavailabilityrequirementbyaccidentallyapplyingittoeverythingwithanetworkport.Sarsa,telluswhichservicesmoves,whatstayswhereitis,andtheactualdayyouwanttodoit.IalsowantthefallbackstatedclearlyenoughthatCaitlincanexplainittoacustomerwithoutborrowingJackson'sentirevocabularyfortheafternoon.Thatisahelpfuldistinction.Imean,weshouldnotsolveanavailabilityrequirementbyaccidentallyapplyingittoeverythingwithanetworkport.Sarsa,telluswhichservicesmoves,whatstayswhereitis,andtheactualdayyouwanttodoit.IalsowantthefallbackstatedclearlyenoughthatCaitlincanexplainittoacustomerwithoutborrowingJackson'sentirevocabularyfortheafternoon.Wearemovingthesmalltextmodelthatscoresvocabularycandidatesaftertranscription.Ifthenewpassmisbehaves,afeatureflagroutesrequestsbacktotheexistingCOINservice,whichstaysavailableduringtheobservationwindow.Thatswitchdoesnotrequireanewclientrelease.Thatlastpointmatters.Icanexplainabackendmovewithanimmediatefallback,butIdonotwantthecohorthearingthatwechangedthewholedictationengine.Seriously,willtherebeanyvisiblemaintenanceperiod?PriyaisschedulingonboardingsessionsforThursdaymorning,andIwouldrathermovethosenowthandiscoverthatourcarefulinternalwordingfailedtomentionanactualinterruptionforhercustomers.Thereshouldbenoplannedinterruptionsbecausetheflagchangesthererankfordestinationwhiletheoldserviceremainshealthy.Iwanttostartat10inthemorningPacific,afterFreya'searliestonboardingsession,andsendonlyinternalworkstationsfirst.Theboxhasenoughmemoryforthecurrentmodelandourmeasuredconcurrency,butthatisnotpermissiontosendeveryqueuejobatonce.Iwillcapeachlightrerankerrequest,watchthequeue,andexpandgradually.Ifthequeuegrowscontinuouslyfor10minutes,werollbackratherthanimprovise.Canwegettheperformancenumbersbehindthatconfidence?Specifically,IwanttheKVcachehitrate,whatworkloadproducedit,andwhethertheprefilltimeincludeswaitinginthequeue.Wehavehadafewdashboardnumbersthatlookedwonderfulbecausethetimerstartedafterallthepainfulwaitingwasover.Iwouldpreferanugliernumberthatreflectswhatthepersondictatingactuallyexperiencesontheirscreen.Onthescrapingreplay,thekeyingcassetteratewas68%,withprefixtrackingenabled,premierwith21%intheearlierconfiguration.Mediumprefilltimesfellfrom410millisecondsto170milliseconds.Thoseprefillmeasurementsexcludekeyweights,soIamnotpresentingthemasend-to-endlatency.Mediumkeyweightwas30milliseconds,andthe96thpercentilewas140millisecondsatourreplayload.Thereplayusedstandardizedpromptstructuresandsyntheticvocabularylists,notproductionrecordingsorcustomreleases.Iwanttounderstandwhetherthis68%resultsurvivesthenextcohort.Ifthereplayaccidentallysendsthesameshortlistoverandover,theKVcachecouldlookmuchbetterthanitwillwhenPriyabringsinseveralworkspaceswithdifferentnamesandpeoplestarteditingthoselistsbetweensessions.Good.Letuscarrytheconservativenumberforward.Imean,acachehitrateisonlyusefulifwerememberwhatwashittingthecache.Sasha,namethepeopleyouneed,becausetravelshouldnotturnthisintoyoudebuggingfromatrainplatformalone.Jackson,Iamassumingyoucanowntherequestrouting,butIwantyoutoconfirmbeforeIvolunteeryourentireThursdaymorning.IneedJacksonfortheroutingplan,timeouthandling,andatestedrollbackscenario.IneedAvaforthatadversarialreplayandaquickaccuracycomparisonsothefasterpathdoesnotquietlychangeraces.IneedPriyatoconfirmhersessionscheduleandsendonerepresentativesanitizedrequestcase.Caitlincancoordinatethatintroduction.Iwillowndeployment,dashboards,andthecostcomparison.Lupitodoesnotneedtoredesignanythingforthismove.Iftheusersuseanewscreenbecauseofit,wehaveprobablytakenanunnecessarydetour.Confirmed.IcanpairwithyouontheflagWednesdayandbeonlineThursdayfortherollout.Iwouldalsoliketotestadeadconnectiontothebox,notjustacleanerrorresponse.Ahangingrequestishowasupposedlyoptionalre-rankerturnsintoatranscriptionoutage.Weshouldreturntheoriginaltranscriptwhenthere-rankertimesout,thenrecordthatdegradationwithoutretryingforeverinthebackground.IwillcontactPriyarightafterthisandgetthesessionscheduleintoSarah'srolloutnote.Forthecustomer-facingdescription,Iwillsay,transcriptsremainavailableiftheoptionalvocabularypassistemporarilyunavailable.Jackson,pleaseconfirmthatwordingagainsttheactualtimeoutbehavior.Iwouldratherpromiseaslightlylesspolishedtranscriptthanimplythewholeproductisdownwhenonlytheextranamerankingpasshasbeenskipped.Thatistheplan.Imean,theheadlineisnotthatwediscoveredacheapermachine.Itisthatwenowunderstandtheidleservice,thefallback,andtherealcostboundary.Seriously,stayforanythingyoucan,butdonotmissyourtrainonourbehalf.Wewillrepeatyourownersanddatesattheend,andnobodyshouldtreatadroppedphoneconnectionasapprovalforariskychange.Kavehjustjoined,soletmebringhiminbriefly.Kaveh,welcome.Wehavecoveredlaunchreadinessandatargetedinfrastructuremove,andweareabouttogetintoproductquality.Youaskedforafewminutesonretentionandwhatevidencewouldmatterforthenextround.Wehaveaboutfourminutesnow,sogiveusthepracticalversion.Thanks.ThepracticalversionisthatIcarelessaboutabigregistrationspikethanaboutwhetherpeoplereturnaftertheirfirstusefultranscript.Retentionshouldtellusthatvoiceflowbecomespartofaworkflow.Pleaseseparateuserswhotriedonedemofrompeoplewhoactuallyhadanotherreasontodictateandchoseyouagainthatweek.ThatishowI'mframingthenextcohort.Wewilltrackafirstsuccessfuldictation,thenareturnonadifferentday,andthenrecurringuseacrossthefollowingweeks.Wearenotplanningtocountopeninganonboardingemailasproductactivity,evenifthatwouldmakethechartlookfriendlier.Good.Dailyactivesareusefuliftheactivitymeanssomethingconsistent.Iwoulddefineanactiveuseraroundacompleteddictation,withareasonableminimumsoanaccidentalbuttontapdoesnotcount.Thenshowdailyactivesalongsideweeklyactives.Somevaluableusersdictatetwiceaweek,andadailychartalonecanmakethathealthybehaviorlookdisappointing.Thatfitsthejobswearehearingabout.Imean,apersondictatingmeetingnotesonTuesdaysandThursdaysisnotnecessarilylesssuccessfulthansomebodysendingfivetinymessageseveryday.Wewillmaketheeventdefinitionexplicitandsegmentbyworkflowinsteadofquietlyrewardingwhicheverbehaviorproducesthelargestnumber.Forthenextfundinground,thequestionwillbewhetherusagerepeats,whetherthoseuserscanbecomepayingcustomers,andwhetherservingthemhassensibleeconomics.Alowinfrastructurebillisnice,butitisnotasubstituteforretention.Likewise,strongretentiondoesnotexcuseacostcurvethatgetsworsewitheveryadditionalminuteofaudio.Iwouldalsokeepthecohortsvisible.Ifthenewestcohortretainsbetterafterthenamefixes,thatisausefulstoryaboutlearningandexecution.Ifalltheimprovementcomesfrommixingnewuserswithafewveryloyaltesters,itishardertointerpret.Smallsamplesarefineatthisstage.Hiddendenominatorsarenot.Understood.Iwillincludethecohortsizes,theactivationdefinition,andtheusagewindowonthesamepageasthecharts.Forcosts,wewillseparatetranscriptionproviderchargesfromthevocabularypassandsharedinfrastructure.Sashaalreadywarnedusnottodescribeanownedmachineasfree,whichseemsespeciallyrelevanttothisdiscussion.Exactly,andwhenyoutalktouserswhostopped,askwhatinterruptedthehabit.Wasitaccuracy,speed,trust,orsimplynorecurringmeet?Ashort,honestexplanationfromfivepeoplecanbemoreactionablethanabroadsatisfactionaverage.Iwouldbeespeciallyinterestedinwhetherrepeatednamecorrectionsareenoughtomakesomeoneabandonanotherwiseusefulproduct.Ineedtodropnow,butthisistherightleveloffocus.Sendmethecohortreadoutonceithasactualobservations,notaprettierforecast.Iamhappytoreviewthemetricdefinitionsbeforehandifthatsavesrework.Thanks,everyone,andSaoirse,safetravels.Iwillleaveyoutotheproductdecisionsratherthantakingtherestofyourmeeting.Let'stakeafive-minutebreaknow,andwhenwereturn,wewillstartwithXochit'sdesignrefreshbeforemovingintoIFA'sevaluationresultsandourfinalassignments.Alright,welcomebackeverybody.Satchel,pleasetakeusthroughthehomepagebuttonfirst,thenthecolormismatchwithAIBuilderSpace,andfinallythesettingspage.Thehomepagehasthreethingsthatlookliketheprimaryaction,andonlyoneofthemstartsdictation.Thelargebuttonsayscreate,themicrophonetilesayscapture,andthesmallbuttonintheheadersaysnewsession.Thatistoomuchvocabularyforthesamebasicintent.Iwantonemainbuttonlabeledstartdictation,withthemicrophoneiconsupportingthewordsratherthanreplacingthementirely.Doesstartdictationrequestmicrophoneaccessimmediately,ordoesitopentherecordingscreenfirst?Rightnow,thebrowserpermissionpromptappearsafterthesecondclick.Iwouldloveto"
   ]
  ]
 }
}