See setup progress separately from operating status
Setup now has four clear steps. Once a business operates, its progress ring stays complete while the status explains what it is doing or waiting for.
Release details
Available to: Running production workspaces verified on 18 November, on releases installed during 17 November Pacific Time.
The sidebar, Home cards and business view share the same status and supporting facts, including open menus, waiting reasons and retry countdowns.
Repeated retries that reach the same data wait are spaced farther apart, up to seven days, unless the robot supplied a specific retry time.
Explore the trade’s timeline in the Market tab
A timeline under “How this business makes money” shows the sequence of buying, lead times, selling, write-offs and payout. You can inspect phases and zoom across time scales.
Release details
Available to: Running production workspaces verified on 18 November, on releases installed during 17 November Pacific Time.
The timeline distinguishes a deliberate hold for a future price or demand from time imposed by delivery, transfer or another process. Trade classification follows the reason for waiting, rather than duration alone.
Robots record the timeline in the scheme and use it when assembling menus. Until a recorded timeline is available, the card shows a draft derived from the description.
Known outcomes no longer need to be left blank
Builder’s data checks and robot instructions now accept fully labelled histories. Record every outcome the evidence can establish, including declined deals; leave only genuinely unknown outcomes blank.
Release details
Available to: Running production workspaces verified on 18 November, on releases installed during 17 November Pacific Time.
Warnings based only on having all outcomes labelled, or on where unknowns are clustered, are retired. Stored checks hide those warnings when read; existing research files refresh on their next normal update.
Current default P34 models support fully labelled histories; legacy model exceptions retain their own requirements. Unknown outcomes are not replaced with invented zeros.
Carry a business change through the affected stages
Your connected agent can submit changes to venues, logistics, data or menus. Builder returns to the earliest affected stage and tracks the work in a Market tab card.
Release details
Available to: Running production workspaces covered by the 16 November rollout and verified release descendants.
By default, a change waits for ongoing work or an open question. Each affected stage reports what it changed before the business returns to operating; new venue access is requested from the member.
The card explains progress and waiting reasons, and offers Cancel. A queued description change does not silently overwrite edits made while it waited.
Check whether the history supports quantity decisions
New outcome warnings flag histories where larger quantities never pay more, winning opportunities are rare, or outcomes appear only in a particular slice of the data.
Release details
Available to: Running production workspaces covered by the 16 November rollout and verified release descendants.
The rare-opportunity warning applies when fewer than 0.5% of at least 100 labelled deals win at their best labelled quantity. These audit warnings do not block fitting by themselves.
Readiness checks expose how many menus have outcomes, fixed history windows and assumptions behind the labels. Robots are asked to record all units sold, and a quantity-identification refusal is treated as a data issue rather than an infrastructure failure.
Repeated deeper-step refusals and delays get bounded holds
Builder records how each P34 step ends and can temporarily hold a deeper step that repeatedly refuses the same history or runs out of menu time, while earlier steps continue.
Release details
Available to: Running production workspaces covered by the 16 November rollout and verified release descendants.
While operating, three consecutive data refusals can trigger an investigation and a step hold with a weekly retry. Five consecutive menu-time cuts can hold a step with a daily retry. The member can use “Run again” to lift a hold.
A queue wait can count toward the time-cut threshold. Timing holds are explained as unfinished analysis, rather than a verdict on the history; adding more history alone does not lift a timing or quantity-identification hold.
Fit timing follows the size of the history
Builder estimates full-fit time from measurements at a comparable history size, and can prepare a context ahead or skip automatic menus whose fits are unlikely to finish in time.
Release details
Available to: Running production workspaces included in the 16 November rollout and verified follow-up fixes.
A much larger history no longer relies only on timings from smaller fits. Waiting and skipped-fit messages explain the constraint, with a daily probe available to check again.
Contexts prepared ahead use the model that subsequent robot fits request, allowing the next eligible menu to reuse the prepared context. Existing access and fit-allowance rules still apply.
More registered fitting models appear in Settings
Global and per-market fitting settings now use P34’s complete authenticated model catalog, making registered variants visible in the model selector.
Release details
Available to: Running production workspaces included in the 15 November Pacific Time rollout, verified on 16 November.
The model selector and progressive-fit routing use the same catalog, so newly listed progressive variants stay on a compatible fitting route.
Existing default resolution, saved model preferences and API access rules are preserved. Older P34 servers retain the shorter discovery-list fallback.
A fit allowance that refills continuously
Full automatic fits now use an allowance that refills throughout the day. Waiting messages show the UTC time another fit can go, rather than suggesting a midnight reset.
Release details
Available to: Running production workspaces covered by the 13 November rollout, verified on 14 November.
The default rate is 10 full fits per business per day, with room for up to 30 at once. Saved rates still apply. Reusing a kept context, practice fits and P34 grounding or model-fit failures do not consume this allowance; other plan limits and applicable charges still apply.
“Fit now” uses the allowance but is not blocked by it. Automatic work resumes once no full fit needs to wait, and the status names whichever limit still holds the business. A separate guard bounds repeated full-fit attempts.
Username and password stay together in Vault
The username starts a new row with the password beside it on wider screens. On phones, the fields stack in the same order.
Release details
Available to: Running production workspaces covered by the 13 November rollout, verified on 14 November.
The URL moves to its own row at medium widths, keeping the login fields together without leaving a gap.
Keyboard tab order follows the visual order.
See when the browser is starting or reconnecting
The Browser tab shows loading and reconnect status while preparing the live view, and keeps the indicator until a usable page frame is ready.
Release details
Available to: Running production workspaces included in the 13 November rollout. Suspended workspaces were not started for it.
Interrupted connections show automatic retry status. Paused and deferred sessions explain why they are waiting without showing false activity; deferred sessions do not start a browser.
The indicator respects reduced-motion preferences and announces status to screen readers. Leaving the tab cancels its outstanding frame request.
Large history profiles recover from an outdated memory allowance
When a growing history exceeds a profile’s previous memory allowance, Builder can retry with more memory in the same pass instead of waiting for the next scheduled pass.
Release details
Available to: Production workspaces verified on 13 November; this is the verification date.
Retries are bounded by host capacity, the operation’s memory ceiling and a maximum of five runs. Timeouts and other kinds of stops do not trigger this retry.
A lighter history-file reader also reduces memory overhead. A pipeline question raised before a profile succeeds is not automatically closed.
Choose passwords and keys for each business
Vault’s “Usable by” field now lets you select several businesses. Each business also has a Passwords & Keys card in Settings, with switches for the entries its robots may use.
Release details
Available to: Production workspaces verified on 13 November; this is the verification date.
“Robots may also ask for it as” lets each business use its own entry under a name a robot expects. The card warns when a business may use several logins for one site.
Requests for another business’s login offer an explicit sharing choice and default to saving a separate entry. Changing the chosen login does not itself sign the browser out of its previous account.
Audit findings stay visible when a robot changes its claims
A robot’s statements, method renames or missing comparison data no longer silently dismiss serious outcome findings. Builder explains what changed and preserves the member’s decision for review.
Release details
Available to: Production workspaces covered by the 13 November rollout and verified release descendants.
When relabelling affects outcomes you chose to leave blank, Builder notifies you and asks again instead of automatically extending the list to other method names. Until you answer, relabelling can change which outcomes are sent to P34; fits continue.
History comparisons require enough labelled deals on both sides. A review waiting for the daily review allowance can resume when that allowance becomes available.
See when a limit or budget pause starts and ends
Daily-limit and token-budget pauses, and their resumptions, now appear in the business console with a reason. Status timing is refreshed when the status actually changes.
Release details
Available to: Production workspaces verified on 13 November; this is the verification date.
A new pause no longer inherits the age of the previous status in checks that use that timing.
Support diagnostics also record changes of business, robot, fit, menu and portfolio state to help investigate interrupted work.
Partner-market calculations respect their budget and result timing
Builder better handles partner-market calculations that produce a portfolio only when they finish, while respecting the calculation budget and menu expiry.
Release details
Available to: Production workspaces verified on 12 November.
When P34 explicitly requires a different endpoint, Builder uses a compatible model and records which model was used.
If a stopped calculation has issued an answer that is not yet readable, Builder waits for a bounded interval instead of repeatedly stopping it or immediately treating the answer as cancelled.
Retries wait for the daily fit allowance
A retry now waits until the business’s daily fit limit permits another calculation. A retry blocked by that limit returns to its waiting state instead of getting stuck as running.
Release details
Available to: Production workspaces verified on 12 November; this is the verification date for the documented rollout.
A retry is not scheduled when the allowance would free only after the menu expires.
A review check can recover a retry left without a submission, with safeguards for cases that have already resumed or been resolved.
Clearer verdicts when a deeper model declines the history
Builder uses the result of each calculation step to distinguish a model’s data refusal from a calculation failure, avoiding unnecessary failure cases and retries.
Release details
Available to: Production workspaces verified on 12 November; this is the verification date for the documented rollout.
When every step that could run has answered and the remaining steps declined the history, the menu receives a clear data verdict.
A genuine failure in the current calculation remains an error, even if an older model summary would suggest a data refusal.
Unread daily ideas stay in place
Opening a daily idea frees its slot for a future suggestion while the other unread ideas remain pinned. A full unread list does not trigger another daily search.
Release details
Available to: Production workspaces included in the 11 November rollout.
If you open one of three unread ideas, the next due daily search can save one replacement without displacing the other two.
The existing once-per-day policy and 12-hour gap remain in place.
Review the quality of outcomes used for model fitting
Builder audits historical outcomes at deal level and exposes its findings alongside data and history, helping you review the evidence P34 learns from.
Release details
Available to: Production workspaces covered by the 11 November rollout and verified release descendants.
Audit findings are warnings; they do not automatically block a fit, menu, or trade. Significant findings can trigger a review within the configured limits.
Members can choose how outcomes are included in future fits. Any proposed test trade still requires the member’s approval.
Review descriptions and proposals in the Market tab
The Market tab now brings together your business description, proposed changes, the description editor, and parameters.
Release details
Available to: Production workspaces included in the 10 November Pacific Time rollout.
A proposed description has its own Apply or Keep question. Your answer acts on that exact proposal, and the question and proposal card stay in sync.
Older Description-tab links open the editor in the Market tab. Only the member can decide whether to apply the proposed description.
Model preparation is no longer treated as a failure
Builder distinguishes a model that is still being prepared from a failed model, and resumes scoring when the required models are ready.
Release details
Available to: Production workspaces included in the 10 November rollout.
A model that is still being built no longer causes an incorrect failure escalation or repeated no-selection warnings. Waiting remains bounded by the menu and model context.
When a model cannot handle the size of the selected history, Builder explains the limit and suggests using fewer recent menus or fewer columns, instead of retrying with more data.
Recover a fit after an interrupted upload
If an app restart interrupts an upload, Builder can reconcile the fit already received by P34 and clear the resulting hold.
Release details
Available to: Production workspaces included in the 10 November rollout.
Recovery checks the original submission before sending anything again, and keeps the result attached to the menu it was submitted for.
If a resend is needed, the member sees a clear “Send the P34 fit again?” choice. An upload that is still running is not replaced.
Clearer results when some models cannot use the history
Builder now explains when available models scored a menu while other models declined its historical data, instead of treating that outcome as a general P34 failure.
Release details
Available to: Production workspaces included in the 10 November rollout.
This avoids unnecessary failure escalations and retry cycles when the models that could run have answered the fit.
A missing answer caused by a timeout or an unexplained model failure remains a separate issue; it is not reclassified as a data verdict.
Outcome bars count deals, not quantity options
Outcome summaries now count each deal once, so several quantity options for one deal no longer make its outcome coverage look incomplete.
Release details
Available to: Production release tracks described in the 10 November release notes.
The Data tab, historical-menu explorer, and relevant Advanced Analytics charts use one outcome class per deal, choosing the most trusted outcome among its options.
Older profiles refresh on their normal background schedule. Row-level checks and the “Rows with a known outcome” tile retain their existing meaning.
Less clutter in every histogram
Log-scale controls now sit inside the chart. They fade into the background until you hover or focus them.
Release details
Available to: Production workspaces included in the 9 November rollout.
Controls remain usable by keyboard, and your scale choice stays in place when the chart redraws.
This change adjusts the controls’ presentation; histogram data and calculations are unchanged.
See the difference between losses and gains
Histogram bins now separate negative and positive values exactly at zero. Range labels and individual log-scale controls make distributions easier to read.
Release details
Available to: Production workspaces included in the 9 November rollout.
Tiny nonzero boundaries keep their precision in tooltips. Finite values remain counted, including values folded into marked tail bins.
Each histogram remembers your scale choice when redrawn. Highly uneven distributions can start with logarithmic bar heights automatically.
Charts that stay readable with uneven data
Adaptive histograms handle sparse values and outliers. Profit labels fit on smaller screens, and zero percentages have a clearly marked baseline.
Release details
Available to: Production workspaces included in the 9 November rollout.
These changes cover the historical-menu explorer and shared Data / Advanced Analytics plots.
Outliers are retained in marked end bins. Tooltips and CSV exports retain the original values.
Explore profitability across historical menus
Historical-menu profitability statistics and outcome histograms help you inspect how results change over time.
Release details
Available to: Production workspaces included in the 9 November rollout.
Historical profitability summaries are available in the menu explorer and shared Overview / Data views.
Older profiles refresh on their existing background schedule. Statistics that are still pending remain unknown until that refresh completes.
Investigate an individual item
Item investigations bring a focused research workflow to individual menu items in Business Builder.
Release details
Available to: Member workspaces included in the 8 November rollout.
The workflow was installed across the member workspaces covered by the release.
Release checks confirmed installation, routes, and served assets. Model execution was tested with mocks; a paid investigation was not run during rollout verification.
A faster path to Item history
A dedicated interactive lane lets smaller Item history requests run alongside background work when capacity permits.
Pilot details
Available to: A limited pilot workspace. A broader release is not recorded in this update.
Larger analyses and exports continue through the regular queue. Availability still depends on capacity.
The pilot remains separate from the general release track in the reviewed release records.