Skip to content

AI uses for records

Hotel overbooking recovery dataset: testing agents on an oversold night

By SourceX Editorial · Reviewed by Noah Loul ·

9 sections

Short answer

Build the evaluation from one oversold night: freeze the inventory, arrivals, rate conditions and channel terms as they stood before the decision, list every constraint that limits who can be moved, record the decision and its approver, and grade the agent on the resulting reservation, folio and channel records, not its message. A hotel overbooking recovery dataset pairs each such case with a near-miss night where the right answer differs.

Agent run

Hotel overbooking recovery

This page covers hotel overbooking recovery dataset: testing agents on an oversold night.Illustrative

Key takeaways

  • An oversold night tests judgment under constraints; a routine reservation tests a single write.
  • Freeze inventory, arrivals, blocks, channel origin and rate conditions at a stated time before the decision.
  • Name every binding constraint with its source, including system rules that make some moves impossible.
  • Grade the reservation, folio and channel end state against an expected state, not the agent's message.
  • Pair each recovery case with a near-miss night where the correct answer is to move no one or to take a different path.

Why is an overbooked night a harder agent test than a routine reservation?

Because a routine reservation is one write against free inventory, while an oversold night forces the agent to choose which confirmed guest moves, where, under which rate conditions and on whose authority, and then to leave every linked record consistent. That is the difference between checking a form fill and checking judgment under constraints.

Overbooking is usually a configured state, not an accident. Mews documents that a property can allow it by allocating more inventory than physically exists, for example an adjustment of minus 10 when there are 9 units, and its Connector API exposes service overbooking limits, each with a limit value and the time units it covers (the operation is marked as beta). Oracle OPERA Cloud describes the same idea through sell limits: a positive sell limit adds rooms to sell on top of the physical rooms, set at the house, room class or room type level.

When the night arrives and every guest shows up, the distribution contract adds pressure. Booking.com tells partners that bookings are confirmed immediately and must be honored, and that when no other room is free the property is responsible for finding accommodation of the same or a better standard and covering the cost, with transport worth arranging as well. An agent that only knows how to create a booking has nothing to offer here. One that can read the arrivals, the rate rules and the channel obligations, and pick the least harmful move, is doing real work.

What does the agent see at the start of the oversold evening?

It sees a snapshot: inventory by room type, the arrivals list with each reservation's state, stay-overs and expected departures, rooms out of order, group blocks, channel origins and the rate conditions attached to each booking. Freeze that snapshot at a stated time and give it to the agent as the starting state, because a case built from today's tables will quietly include check-ins and cancellations that had not happened yet.

Reservation state needs exact definitions. Mews uses Inquired, Confirmed, Started (checked in), Processed (checked out), Canceled, Optional (held by the property but not confirmed by the guest) and Requested (confirmed by the guest but not the property, effectively a waitlist). A packet that collapses Optional and Confirmed into one bucket will grade an agent as wrong for relocating a guest who never held a firm room.

What does the agent see at the start of the oversold evening?
Starting-state elementWhat to freezeTypical sourceWhy the agent needs it
Inventory positionPhysical rooms, out-of-order rooms and sell limit by room type for the nightPMS availability and sell-limit settingsShows how many rooms short the night really is
ArrivalsEvery arrival with state, room type, rate, length of stay and arrival time if knownPMS reservations listDefines the pool of candidates to relocate
Stay-overs and departuresIn-house guests, due-outs and late checkout requestsPMS in-house and departures viewsA late departure can create or erase the shortfall
Group and block allocationsBlock pickup, release dates and contracted roomsGroup or availability block recordsBlock rooms carry contract terms individual bookings do not
Channel originOrigin plus the channel's own confirmation numberReservation origin and channel number fieldsChannel terms can dictate who pays for relocation
Rate conditionsPrepaid, deposit taken, cancellation terms and loyalty tierRate plan, folio and profileLimits which reservations can be moved or changed
Nearby supplyPartner or sister properties checked, with time of checkPhone log, hub availability viewThe relocation option set must be real at decision time

What should the consented recovery case packet with decision constraints include?

The packet holds one oversold night as a replayable task: the frozen starting state, the evidence references, the constraints that bind the decision, the decision that was taken, the outcome check and at least one counterexample. Every field below names its source system and a provenance label, so a reviewer can tell recorded fact from reconstruction. A hospitality team can copy the table as its packet schema and adjust field names to its own PMS.

At a glance

What should the consented recovery case packet with decision constraints include

What should the consented recovery case packet with decision constraints include?
Packet fieldContentsSource of truthProvenance label
case_id and versionStable case identifier, packet version and the snapshot timestampPacket builderrecorded
consent_recordApproval reference from the property owner, the privacy review outcome and any distribution-agreement checkOwner approval, privacy logrecorded
starting_stateThe frozen inventory, arrivals, blocks, channels and rate conditions from the previous sectionPMS snapshotrecorded
shortfallRooms short by room type, with the calculation shownDerived from starting_statederived
candidate_setReservations eligible to move and the reason each other reservation is excludedDerived plus policyderived
constraintsEach binding rule with its source: rate terms, channel terms, move restrictions, approval limitsRate plans, contracts, PMS rules, role settingsrecorded or cited
decisionWhich reservations moved, where, at whose cost and who approved itReservation notes, move log, manager logrecorded
expected_stateThe reservation, folio and channel records as they should look once the recovery is doneReviewer, from the decisionderived
outcome_checkFields compared against expected_state and the guest-facing facts that must be communicatedReviewerderived
counterexampleA near-identical night where the right answer differs, with the one changed fact namedReal or labeled illustrativerecorded or illustrative

Which dependent booking, purchase-time condition or approval limits the proposed change?

The binding limits are the ones attached to the reservation at the moment of purchase and the permissions of the person acting: what the guest prepaid, which channel sold the room, whether the booking sits in a group block, and whether the user's role allows an overbook or a move. Write each one into the packet as a named constraint with its source, so a grader can say which rule an agent broke rather than only that it failed.

System rules belong in the packet because they make some choices impossible, not merely unwise. OPERA Cloud documents that canceled or no-show reservations cannot be moved to another property, nor can reservations that have incurred financial charges such as no-show charges or payments such as deposits; a completed move issues a new confirmation number and the move needs a reason. In Mews, the cancel operation succeeds only for cancellable reservations, requires notes describing the reason, and charges cancellation fees according to rate conditions only when the caller sets the PostCancellationFee flag, which defaults to false.

Which dependent booking, purchase-time condition or approval limits the proposed change?
ConstraintWhere it is recordedWhat it rules outWhat the agent must produce instead
Deposit or prepayment takenFolio and payment recordsA plain cross-property move in systems that block itA cancel-and-rebook path with the refund or transfer documented
Channel bookingOrigin and channel confirmation numberTreating the guest as a direct booking with direct termsRelocation that follows the channel's terms and a matching channel update
Group block membershipBlock or availability block linkMoving a block room without checking contracted pickupA choice that keeps the block commitment or records the exception
Multi-room or linked staysReservation group or shared bookerMoving one room of a family or party aloneA move of the whole party or an explicit note of the split
Role permissionPMS role and task settingsAn overbook or move by a user who lacks the rightAn escalation to an authorized approver
Cancellation chargesRate conditions and cancel request flagsCharging a guest the property is relocatingA cancel with charges suppressed and the reason noted

Illustrative case: a convention night two rooms short

Everything in this case is illustrative: the counts and timings are invented to show the packet working and describe no real property. A 180-room hotel holds a convention block of 60 rooms on a Thursday. At 5 p.m. the snapshot shows 3 rooms out of order after a leak, 2 more arrivals than sellable rooms in the standard king type, and no free upgrades. Arrivals in that type include 9 direct bookings, 7 channel bookings and 14 block rooms.

Linked evidence narrows the candidate set. Four direct bookings are prepaid with deposits, which rules out a simple property move in the PMS; 2 channel bookings arrive after 10 p.m.; 1 direct booking is the second room of a family traveling together. The sister property 1.5 miles away confirmed 2 standard rooms at 5:20 p.m. by phone, logged with the name of the person who confirmed.

The recorded decision relocated 2 single-night direct bookings without deposits, arriving late and not linked to other rooms, to the sister property, with transport arranged, cancellation charges suppressed and a note citing the leak and the overbook. The verified outcome is the state, not the message: both original reservations canceled with the reason recorded and no cancellation charge posted, both replacement bookings present at the sister property, the block untouched and the family's two rooms still together.

The counterexample changes one fact. On an otherwise identical night, the only late, unlinked arrivals hold deposits. The correct answer is no longer a property move; it is a cancel-and-rebook with the deposit refunded or transferred and documented, or an escalation if the agent's role cannot do that. An agent that repeats the first night's actions on the second fails, and the packet shows exactly why.

Which reservation, folio and channel records prove the recovery worked?

The proof is the end state of the reservation, folio and channel records compared with an expected state, plus the facts the guest had to be told. A polite confirmation message or a status field reading Canceled proves very little on its own: a reservation can be canceled with a charge the guest should never have paid, or the replacement booking may not exist.

At a glance

Which reservation, folio and channel records prove the recovery worked

The tau-bench reference code shows a strict version of this approach. Its reward function replays the ground-truth actions against a freshly loaded database, hashes the resulting data and compares it with the hash of the database the agent left behind; any mismatch sets the reward to zero. When a task also lists required outputs, each must appear in the agent's replies or the reward is zero as well. A hotel packet can borrow the pattern without the hash: list the fields that must match and the facts that must be said.

  • Original reservation: state Canceled, reason note present, no cancellation charge posted.
  • Replacement booking: exists at the receiving property for the same dates and party size, with its new confirmation number recorded in the original's notes.
  • Folio: deposits refunded or transferred with a matching payment record; no orphan charges.
  • Channel: the channel booking reflects the change, so the guest does not arrive at a hotel that still expects them.
  • Block and linked rooms: pickup counts and party links unchanged unless the decision says otherwise.
  • Guest-facing facts: new property, address, transport arrangement and who bears the cost appear in the agent's message.

Which near-miss nights keep the evaluation honest?

Nights where moving someone is the wrong answer. A packet set built only from true shortfalls teaches an agent that every tight night ends in a relocation, so pair each recovery case with nights that looked oversold at 5 p.m. and resolved without any move.

Good near misses come straight from the same property's history: a night when two out-of-order rooms returned to service by 7 p.m., a night when expected no-shows and early departures closed the gap, a night when an upgrade into an unsold suite category solved it, and a night when the only movable guests had deposits or were part of a group. Each needs the same frozen snapshot and the same outcome check, with expected_state showing no relocation.

Settle three things before any packet is shared: that the property owner approves the use, that guest identifiers are removed from every linked table and note, and that channel and franchise agreements allow booking records to be used this way. These are separate questions; removing names does not create a right to license the records, and permission to access the PMS is not permission to license what is in it.

Guest identifiers hide in more places than the profile. Reservation notes, folio descriptions, phone logs with the sister property and transport bookings all carry names, phone numbers or loyalty numbers. Replace them with stable pseudonymous keys so the family's two rooms still link, and drop payment card data entirely. Channel confirmation numbers are identifiers too; keep a hashed key if the packet needs to show the channel link.

Have counsel read the distribution and franchise agreements and the guest privacy notice against the planned use, and record the outcome in consent_record. The hospitality operations reviewer then checks that the constraints in each packet match how the property really worked on that date, including rate plans and role settings that may since have changed.

Licensing a hotel's oversold-night history through SourceX

Oversold nights are among the most informative records a hotel keeps, because they capture decisions under real constraints rather than routine bookings. SourceX turns histories like these into licensed AI datasets for the hotels that hold them, taking care of rights review, contracting, preparation, delivery and payment.

A hotel group can start with a short company assessment and no data upload. The group approves the scope and terms and signs first; only after that may SourceX use, sublicense and distribute the de-identified dataset, and only as the agreement allows.

Frequently asked questions

What is a hotel overbooking recovery dataset?

It is a set of oversold nights, each packaged as a replayable task: the inventory and arrivals as they stood before the decision, the rules that bound it, the relocation or other action taken and the records that show how it ended. It is used to train or evaluate agents on hotel operations, not to forecast demand.

Why not evaluate a hotel agent on routine bookings?

A routine booking has one correct write and almost no constraints, so most agents pass. An oversold night requires choosing among guests under rate, channel, group and permission rules, which separates agents that follow policy from agents that only produce fluent text.

Can the reservation status field be the label?

No. A status of Canceled says nothing about whether a charge was wrongly posted, whether the replacement booking exists or whether the channel was updated. Grade the full set of linked records against an expected end state.

How many counterexamples should a packet set include?

Pair each recovery case with at least one near-miss night from the same property where the right answer is different, such as a night the shortfall closed by itself. Without them, an agent can score well by relocating someone every time.

Do illustrative cases have any place in the dataset?

They are useful for documentation and for testing the grader, as long as they are labeled illustrative and kept out of any claim about real performance. Training and evaluation results should rest on approved real cases.

Sources

  • Mews documents that a property can allow overbooking by allocating more inventory than physically exists. Source
  • The Mews Connector API exposes service overbooking limits with a limit value and covered time units; the operation is in beta. Source
  • Mews defines reservation states, channel confirmation numbers and a cancel operation that requires reason notes and charges cancellation fees only when requested. Source
  • OPERA Cloud sell limits allow overbooking by adding rooms to sell on top of physical rooms at house, room class or room type level. Source
  • OPERA Cloud cannot move canceled, no-show, charged or paid reservations to another property; a move issues a new confirmation number and needs a reason. Source
  • Booking.com says an overbooked property must find same-or-better accommodation, cover the cost and should arrange transport. Source
  • Booking.com says its bookings are confirmed immediately and must be honored by the property. Source
  • tau-bench scores an episode by comparing the agent's final database state with the state produced by replaying ground-truth actions, and by checking required outputs. Source

See if your company qualifies

A short company assessment. No data uploads are needed.

See if your company qualifies