# When Game Discovery Has a Deadline
## Allocating attention during a short discovery event
Author: Jason Doyle
Published: 22 October 2026
Version: 1.0.0
Status: Self-published technical report; not peer reviewed.
Licensing: Report text and original data are CC BY 4.0. Python software is
MIT. See [LICENSES.txt](./LICENSES.txt) for file-level terms.
> Disclosure: These views are my own and do not represent any current or
> former employer. This paper uses public documentation and published
> research. It does not describe non-public Steam architecture or ranking
> models.
> Evidence status: The accompanying study is entirely synthetic. Its
> response probabilities, audience labels and attention budgets are
> modelling assumptions. They are not estimates of Steam traffic or player
> preferences. Raw results and executable code are included for review.
## Executive summary
Steam Next Fest gives players a week to find upcoming games with playable
demos. A short event creates a practical question for a discovery system:
how much attention should be spent learning about unfamiliar games when
there is limited time to use the answer?
The October 2026 edition runs from 19 to 26 October. Steam's documentation
describes initial randomisation, with exceptions for named lists and
personalised segments. After a few days, logged-in users see carousels
personalised using behaviour during the festival. It does not disclose a
single exploration budget or a platform-wide switch to pure
exploitation.\[1\]\[2\]
Public player counts omit recommendation opportunities and outcomes for
games that were never shown. Evaluating the allocation would require
assignment records or an experiment.
This paper examines a narrower decision model. It compares greedy selection
and random-then-greedy schedules with audience-conditioned Thompson sampling.
Count quotas test whether reserving attention for under-observed games
increases relevant exposure. A terminal-tail control uses posterior means
when new feedback can no longer affect another decision within the
declared objective horizon.
The study uses 32 paired synthetic worlds for each of 16 conditions.
Catalogue sizes range from 1,000 to 15,000 items. Other conditions vary the
attention budget, feedback delay and existing observations. They also
change the quality of audience labels. Two conditions continue learning
after the event.
In the 7,000-item base condition, a one-observation quota increases cold
relevant-pair coverage from 61.96% under Thompson sampling to 73.76%.
Expected response falls from 59.03% to 57.79%. A three-observation quota
with a terminal tail produces 73.13% coverage but reduces response to
49.26%.
The trade-offs change with the budget and feedback delay. The
one-observation quota improves both measures against Thompson in some
conditions. Greedy from start reaches 71.71% response while covering
16.31% of cold relevant pairs. Short random-then-greedy schedules also
produce high response rates with much narrower coverage.
In the continuation conditions, adding a greedy window before event close
reduces post-event response against a control with the same final tail.
The immediate gain still raises combined response over the equally
weighted event and continuation.
The practical question is what another observation can change before the
objective horizon. A count target makes coverage measurable; its value
depends on the decisions that remain.
## 1. What the public documentation establishes
Steam Next Fest has a clear public purpose: help players find upcoming
games and let developers reach potential audiences. The October 2026
edition starts at 10:00 AM PDT on 19 October and ends at the same time on
26 October.\[1\]\[2\]
Steam says that the first few days are randomised except for explicitly
named lists and personalised segments. It then describes personalised
carousels for logged-in users, using behaviour observed during the event.
Categories and Steam tags also affect where games appear.\[1\]
This describes a transition in some discovery surfaces. The fraction of
impressions allocated to exploration remains unspecified, along with the
role of existing player histories in individual rankings. Internal
candidate generation and ranking objectives are also undisclosed.
The distinction matters when constructing a baseline. A simulator that
randomises every recommendation for two days and then becomes greedy is
an experimental policy with its own assumptions.
### 1.1 Catalogue totals are not assignment logs
Output Lag's launch article reports 7,131 demos. Its embedded counter showed
7,288 when checked on 22 October. The methodology identifies this as its
own count rather than Valve's. It measures concurrent players, checking
active demos hourly and the whole festival every four hours.\[3\]
Those figures give a reason to examine several-thousand-item catalogues.
They provide no estimate of recommendation impressions. Changes to a live
counter also cannot establish historical catalogue growth.
The public demo viewer provides another useful catalogue reference, but
its README warns that some demos may not officially participate in Next
Fest. Its classifications describe AI use during development; this study
does not use them as game-quality or preference labels.\[4\]
Press Preview adds a separate counting difference. Valve says it includes
registered games regardless of whether their demos are live. Appearance
during the event requires a released demo.\[1\] A press-preview entry count
and a live-demo count should therefore not be presented as competing
measurements of the same population.
### 1.2 The event can end before the value of learning ends
Valve allows demos to launch before Next Fest. It also encourages
developers to consider leaving them available afterwards.\[1\]
A recommendation made late in the week can therefore generate information
that remains useful outside the festival. The model needs to state
whether it values that information.
The study therefore reports an event-only objective alongside separate
continuation conditions. Both are explicit modelling choices.
## 2. There are two deadlines to specify
An event has an availability window. A learning system has an objective
horizon: the last future decision for which information has value.
For an event-only model, the objective can be written as:
```text
J_event(policy) = expected sum of responses to event displays
```
A model that also values later recommendations can instead use:
```text
J_extended(policy) =
expected responses during the event
+ weight * expected responses during a declared continuation
```
The continuation experiments use equal weight per display and the same
attention rate for another seven days. These assumptions define the
comparison; later Steam traffic could follow a different pattern.
Feedback timing creates a second constraint within either objective.
An action selected in batch `b` receives feedback at `b + 1 + d`.
Let the final decision batch have index `H - 1`. When
`b + 1 + d >= H`, the outcome cannot inform another decision in the
objective window.
For a model concerned only with additive responses, with no effect on
future candidate availability, that action has no remaining learning
benefit inside the window. Choosing the highest expected immediate
response is then appropriate if the decision beliefs are valid.
The argument applies to the terminal tail of the chosen objective.
An earlier stopping rule would also need to account for the decisions
that feedback could still improve. The simulation uses imperfect Beta
beliefs, so the policy ordering still needs to be measured.
## 3. The underlying ideas already have substantial prior art
Exploration under deadlines is established research.
Traca, Rudin and Yan study mortal bandits, where items appear and
disappear. Their life-regulated algorithms reduce exploration of arms
near disappearance. The paper includes regret analysis and experiments
using randomly logged Yahoo news recommendations.\[5\]
This is close prior art for accounting for remaining item life. It also
shows why a festival deadline needs care: a game may remain available
and useful to recommend after the event.
Wang and colleagues address item-centric exploration for cold starts.
Their method places a filter after an existing ranking stage. It compares
predicted user satisfaction with an item's Beta posterior satisfaction
rate, avoiding audience assignments predicted to be substantially worse
than that item's average. They report an online short-video
evaluation.\[6\]
The count quotas below omit that ranking model and audience-quality test.
They evaluate a simpler scheduling rule.
Zhang and colleagues address delayed long-term outcomes through progressive
feedback. Their impatient-bandit method combines Thompson sampling with a
Bayesian filter that uses earlier observations while preserving uncertainty
about later outcomes. Their application concerns podcast engagement over
60 days.\[7\]
The study here models delayed binary feedback without early signals.
It measures the effects of that delay, leaving the filter and its
production results outside the comparison.
Thompson sampling itself is a standard baseline. In the Beta-Bernoulli
case, an algorithm samples plausible response probabilities from its
current beliefs. It chooses the largest draw and updates the selected
pair when its response arrives.\[8\]
The contribution is an applied comparison under explicit constraints.
The policies use established ideas. The focused source review makes no
exhaustive novelty claim about discovery benchmarks.
## 4. Evidence debt needs a unit
An under-observed game may deserve further investigation. To judge
whether another display is worthwhile, a scheduler needs more than a
count of missing observations.
Consider a count-based target:
```text
D(c, i, t) = max(0, m - n(c, i, t))
```
Here `n(c, i, t)` counts delivered observations for item `i` in audience
`c`. The chosen floor `m` has an explicit unit: observations of that
audience-item pair.
The result is a coverage deficit relative to the chosen floor. Using it
as an evidence requirement would need validation for a particular
decision.
Three observations can contain three failures or three successes.
Observations from a different audience may say little about a niche
game's intended audience. A working model can have many observations
and still be confidently wrong about another population.
Pending responses introduce another distinction. They consume attention
before they become evidence. A scheduler that ignores them can allocate
the same nominal deficit repeatedly while waiting for feedback.
The quota policies reserve pending allocations when deciding what to show.
The reported floor coverage counts only observations delivered by the
deadline. A scheduled display can therefore satisfy a reservation while
its outcome remains unavailable.
### 4.1 Some count targets are impossible before a policy is chosen
Suppose there are `K` target audience-item pairs, no existing observations
and `M` displays. One display produces at most one observation.
The fraction of target pairs receiving at least `m` observations is bounded by:
```text
count-floor coverage <= min(K, floor(M / m)) / K
```
With 14,000 displays and a three-observation floor:
| Items, with one target pair each | Displays per item | Maximum floor coverage |
| --- | --- | --- |
| 1,000 | 14.00 | 100.00% |
| 3,000 | 4.67 | 100.00% |
| 7,000 | 2.00 | 66.66% |
| 15,000 | 0.93 | 31.11% |
These are optimistic upper bounds. Wrong-audience displays and late
responses can reduce achievable coverage. Candidate eligibility can
prevent some allocations entirely.
For independent uniform draws over items, the expected fraction never
shown is `(1 - 1 / K)^M`. That is approximately 13.53% at 7,000 items and
39.32% at 15,000 items for 14,000 draws.
This calculation assumes uniform catalogue sampling. The simulation below
uses audience-targeted candidates. The bound shows why an exploration
duration alone is insufficient to guarantee coverage: the number of
available displays matters too.
Meeting the floor for every item may be a poor use of attention even when
it is feasible. The product objective needs to justify the likely effect
of those observations on future decisions.
## 5. A reproducible decision-model study
The study asks two questions. How do the declared policies trade expected
response against exposure of relevant cold pairs? Does the answer change
when feedback or the objective extends beyond the event?
### 5.1 The synthetic world
Every item has one latent favourable audience among six. Its probability
of a positive response in that audience comes from `Beta(2, 2)`. Other
audience probabilities come from `Beta(1, 19)`.
The two distributions create differences in audience fit. A positive
binary response is a synthetic outcome, with no calibrated relationship
to wishlists or player satisfaction.
| Component | Base condition |
| --- | --- |
| Event | 28 batches, representing four batches per day for seven days |
| Attention | 14,000 synthetic display opportunities |
| Catalogue | 7,000 items; separate size conditions use 1,000, 3,000 and 15,000 |
| Audience demand shares | 40%, 25%, 15%, 10%, 6% and 4% |
| Item audience distribution | Uniform over the six latent favourable audiences |
| Observed audience label | Correct with probability 0.85; otherwise one of the other five audiences |
| Candidate opportunities | 48 distinct matching-label items plus 16 other distinct catalogue items per request |
| Decision beliefs | Separate Beta beliefs for each audience-item pair |
| Prior mean | 0.4 for a matching observed label; 0.1 otherwise |
| Prior concentration | 2, with a separate concentration-20 condition |
| Feedback | Available at the next batch boundary, plus any declared extra delay |
The additional 16 candidates exclude the first 48, but can also carry a
matching label. Every policy receives the same candidate set for each
request in a world.
The random baseline therefore already benefits from coarse audience
targeting. All policies allocate attention within the same retrieved
opportunities.
The working Beta beliefs are misspecified for the hidden audience-mixture
generator. Changing label accuracy affects candidate quality as well as
the accuracy of the fixed prior means. The metadata conditions vary
both mechanisms together.
The policies also learn each pair independently, without transfer between
similar games.
### 5.2 Policies
| Policy | Decision rule |
| --- | --- |
| Uniform | Choose randomly within each request's candidate set |
| Greedy from start | Choose the highest posterior mean from the first batch |
| Random 1 day | Random selection for one day, then highest posterior mean |
| Random 2 days | Random selection for two days, then highest posterior mean |
| Random 3 days | Random selection for three days, then highest posterior mean |
| Thompson | Sample each candidate pair's Beta belief and choose the largest |
| Thompson + terminal tail | Thompson until feedback cannot inform another decision inside the objective; then highest mean |
| Quota 1 | Prefer a matching-label pair below a count of one; otherwise Thompson |
| Quota 3 | The same rule with a count of three |
| Quota 5 | The same rule with a count of five |
| Quota 3 + terminal tail | Quota 3 with the same terminal-tail gate |
The random-then-greedy policies continue updating means after the switch,
unlike an explore-then-commit policy with frozen estimates.
Greedy selection can still expose unfamiliar pairs when candidate sets
change or means tie. It has no deliberate exploration bonus, although it
can learn from that incidental exposure.
When several matching-label pairs are below a quota, the quota policy
chooses among those with the lowest count, breaking ties with Thompson
draws. That count includes delivered observations and pending display
reservations, including allocations earlier in the current batch.
Testing counts of one, three and five checks sensitivity to the chosen
floor. The rule treats each missing count equally, irrespective of its
potential information gain.
The candidate oracle, retained in the raw results, knows the true response
probabilities and selects the best of each request's candidates. This
evaluator-only information provides an upper reference for the policies,
which learn from delivered outcomes.
### 5.3 Measures
Expected response is the average latent response probability of selected
items. Using this evaluator-only quantity reduces the noise from sampled
binary outcomes. The raw results also preserve the actual sampled
responses used for learning.
Cold relevant-pair coverage considers only items with no pre-event
observations in any audience. Among those items, it is the fraction of
audience-item pairs with response probability at least 0.5 that receive
an event display. Pairs on warm items are excluded even if the pair
itself has no observations. This measures exposure in the synthetic
population.
Delivered floor coverage is the fraction of items whose observed-label
pair has at least three delivered observations by event close. It is
reported for every policy, irrespective of that policy's quota.
The raw results also include item coverage and the display share of the
most-exposed 1% of items. Late-feedback fraction counts event responses
not yet delivered at close. An outcome delivered exactly at close counts
as received, although it cannot affect another event recommendation.
### 5.4 Paired worlds and sensitivity conditions
There are 32 independent world seeds per condition. Within a world, policies
share latent response probabilities and candidate opportunities. They also
share exogenous outcome draws. Only the outcome of the selected item is
delivered to a policy.
The protocol specifies 16 conditions. Four vary catalogue size while
holding event attention at 14,000. Two additional conditions keep attention
at two displays per item for the smallest and largest catalogues.
Other conditions add one or three days of extra feedback delay. Warm
conditions provide existing observations for 20% or 60% of items.
Metadata accuracy varies separately. There is also an equal-demand
audience condition and a stronger-prior condition.
Warm items are chosen independently of quality. Each receives 12 pre-event
observations, allocated according to audience demand. Those observations
are available to all policies and are outside the event budget. The warm
conditions test access to prior evidence; they omit fan-driven traffic.
Two continuation conditions retain the same items and response
probabilities for another seven days at the same attention rate. One has
no extra feedback delay; the other adds one day. Each includes an extra
control that applies a terminal tail at the event boundary as well as at
the actual end.
The complete study has 512 synthetic worlds and 5,696 policy runs.
Reported 95% intervals use 10,000 bootstrap resamples of world seeds.
Differences preserve seed pairing.
The intervals describe Monte Carlo uncertainty within the model.
Uncertainty about the model's suitability for Steam remains outside
them. The comparisons are exploratory, with no multiple-comparison
significance claims.
The initial ten-policy protocol was saved before the first experiment.
After its results showed strong random-then-greedy performance, a protocol
amendment added the missing greedy-from-start control. Every condition
and policy was rerun. World seeds and the original policy parameters
were unchanged.
Both protocol hashes are retained. The protocol was locked locally;
the experiment was not publicly preregistered.
## 6. Results under a fixed attention budget
### 6.1 Quotas increase coverage in the base condition
The base condition has 7,000 items and 14,000 event displays. Feedback
arrives at the next batch boundary, with no extra delay. All items begin
without observations.
Values below are means across 32 worlds. The coverage percentages
describe the defined populations, rather than confidence levels.
| Policy | Expected response, % | Cold relevant-pair coverage, % | Delivered three-observation floor, % |
| --- | --- | --- | --- |
| Uniform | 35.53 | 57.79 | 23.64 |
| Greedy from start | 71.71 | 16.31 | 7.72 |
| Random 1 day | 69.80 | 19.73 | 8.56 |
| Random 2 days | 65.16 | 29.25 | 10.58 |
| Random 3 days | 59.61 | 37.62 | 12.97 |
| Thompson | 59.03 | 61.96 | 18.37 |
| Thompson + terminal tail | 59.50 | 61.18 | 18.10 |
| Quota 1 | 57.79 | 73.76 | 16.10 |
| Quota 3 | 48.57 | 73.73 | 32.38 |
| Quota 5 | 44.29 | 73.68 | 32.38 |
| Quota 3 + terminal tail | 49.26 | 73.13 | 32.78 |
The one-observation quota raises cold relevant-pair coverage by 11.80
percentage points against Thompson, at a response cost of 1.24 points.
Increasing the quota to five produces a similar coverage mean while
reducing response further.
Quota 3 + terminal tail reaches the delivered three-observation floor for
32.78% of items, compared with 18.37% for Thompson. Its expected response
is 9.77 percentage points lower.
Paired differences against Thompson are:
| Policy compared with Thompson | Response difference, percentage points | Cold-coverage difference, percentage points |
| --- | --- | --- |
| Random 1 day | +10.78 [10.58, 10.98] | -42.23 [-42.58, -41.87] |
| Quota 1 | -1.24 [-1.38, -1.09] | +11.80 [11.59, 12.02] |
| Quota 3 + terminal tail | -9.77 [-9.98, -9.56] | +11.17 [10.97, 11.36] |
Brackets contain 95% bootstrap intervals over paired world seeds.
Figure 1. Exposure coverage and expected response are separate objectives. Bars are marginal 95% intervals, not a joint confidence region. Every point uses the same retrieval opportunities. Open full-size figure.
### 6.2 Greedy has the highest response in the base condition
Greedy from start produces 71.71% response, with 16.31%
cold relevant-pair coverage. Against Thompson, the paired response
difference is +12.68 points [12.45, 12.91]. Its coverage difference is
-45.65 points [-46.01, -45.29]. The most-exposed 1% receive 47.73% of
displays.
Random 1 day produces a 69.80% response rate, compared with 59.03% for
continuous Thompson sampling. Its cold relevant-pair coverage is much
lower: 19.73% against 61.96%.
The exposure distribution also changes. The most-exposed 1% of items
receive 42.63% of displays under Random 1 day, compared with 17.47% under
Thompson. This describes how much attention goes to a small set of items.
The model supplies audience-targeted candidates and informative prior
means before a policy acts. Greedy can then encounter unfamiliar items
through changing candidate sets. Those information sources are already
present in the greedy baseline.
With 60% warm items, greedy response reaches 81.60%, while cold
relevant-pair coverage falls to 0.05%.
Warm observations were allocated independently of item quality, so this
condition leaves the effects of an established fan audience untested.
Choosing between these allocations requires a product objective. Expected
response and cold discovery can support different choices.
### 6.3 Response stays similar at a fixed attention ratio
With event attention fixed at 14,000, all selected policies' response rates
fall as the catalogue grows.
| Items | Greedy from start, % | Random 1 day, % | Thompson, % | Quota 1, % | Quota 3 + terminal tail, % |
| --- | --- | --- | --- | --- | --- |
| 1,000 | 81.89 | 78.41 | 76.68 | 77.00 | 74.30 |
| 3,000 | 76.15 | 74.20 | 67.29 | 67.20 | 60.90 |
| 7,000 | 71.71 | 69.80 | 59.03 | 57.79 | 49.26 |
| 15,000 | 67.09 | 65.08 | 52.22 | 49.13 | 45.45 |
The one-day random schedule remains above Quota 3 + terminal tail on
response at every tested size. Increasing the catalogue alone produces
no crossover between those two policies in this range.
Figure 2. A fixed attention budget makes larger catalogues harder to learn in this model. Lines connect the tested sizes; they do not establish a threshold or predict Steam outcomes. Open full-size figure.
Holding attention at two displays per item produces a different comparison:
| Items | Event displays | Thompson response, % | Thompson cold relevant-pair coverage, % |
| --- | --- | --- | --- |
| 1,000 | 2,000 | 59.29 | 61.10 |
| 7,000 | 14,000 | 59.03 | 61.96 |
| 15,000 | 30,000 | 59.05 | 61.82 |
The response means are similar despite a fifteen-fold difference in
catalogue size. In this model, attention per item accounts for much of
the earlier size effect. The experiment holds that ratio fixed; it
provides no estimate of how real attention scales.
There are also positive quota results. At 1,000 items and 14,000 displays,
Quota 1 exceeds Thompson by 0.33 response points, with a paired 95%
interval of [0.11, 0.54]. Cold coverage improves by 0.68 points
[0.51, 0.87]. At 3,000 items its response difference is -0.09 points
[-0.26, 0.09], while coverage improves by 8.35 points [8.05, 8.64].
The interval at 3,000 items includes a response loss as well as a gain.
A non-inferiority claim would need a justified acceptable loss and an
evaluation designed around it.
## 7. Feedback and continuation change the decision
### 7.1 A terminal tail changes the cost of delayed feedback
In the base condition, replacing Thompson sampling with posterior means
in the final batch increases response by 0.48 percentage points
[0.47, 0.49]. Cold relevant-pair coverage falls by 0.78 points
[-0.83, -0.73].
One or three days of extra delay make the terminal tail longer. They also
leave 14.29% or 42.86% of event responses undelivered at close,
respectively. Every policy has the same late fraction because request
times and the fixed delay are shared.
The following differences are against Thompson within each condition.
Brackets contain paired 95% intervals.
| Condition | Terminal-tail response difference, points | Quota 1 response difference, points | Quota 3 + tail response difference, points |
| --- | --- | --- | --- |
| Base | +0.48 [0.47, 0.49] | -1.24 [-1.38, -1.09] | -9.77 [-9.98, -9.56] |
| One day extra delay | +2.43 [2.38, 2.48] | +0.24 [0.13, 0.36] | -3.28 [-3.43, -3.13] |
| Three days extra delay | +6.55 [6.38, 6.70] | +0.63 [0.54, 0.72] | +4.18 [4.08, 4.30] |
| 20% warm items | +0.46 [0.45, 0.48] | -1.48 [-1.61, -1.35] | -9.98 [-10.13, -9.83] |
| 60% warm items | +0.40 [0.39, 0.41] | -1.66 [-1.78, -1.54] | -8.89 [-9.03, -8.76] |
| Uninformative labels | +0.87 [0.85, 0.90] | -0.62 [-0.78, -0.46] | -10.77 [-11.06, -10.49] |
| Perfect labels | +0.44 [0.43, 0.45] | -1.16 [-1.30, -1.02] | -7.50 [-7.64, -7.37] |
| Equal audience demand | +0.57 [0.55, 0.58] | -2.03 [-2.16, -1.89] | -10.76 [-10.89, -10.62] |
| Prior concentration 20 | +0.43 [0.42, 0.45] | -2.23 [-2.41, -2.05] | -7.75 [-7.94, -7.56] |
Zero extra delay still includes the batch boundary. One day of extra delay
means feedback arrives five six-hour batches after a display.
The delayed quota results also require both outcome measures. Quota 1
improves cold coverage against Thompson by 12.04 points with one day of
extra delay and 12.05 points with three days. In those conditions its
response difference is positive too.
Quota 3 + terminal tail improves response against Thompson under three
days of extra delay, but cold coverage falls by 1.20 points
[-1.42, -0.97]. The tail-only control has a larger response gain of
6.55 points, with lower coverage. An improvement against a horizon-unaware
reference does not establish that the count quota caused the response
benefit.
The quota policies also count pending reservations, while ordinary
Thompson samples its delivered-data beliefs without such reservations.
These delayed comparisons do not isolate a quota's effect from
pending-work scheduling.
### 7.2 The event-tail window has a post-event cost
When the same items remain useful for another seven days, the actual
objective horizon moves. The end-only terminal-tail policy therefore
matches ordinary Thompson during the event and becomes greedy only near
the later end.
The extra control inserts a greedy window before event close, then
resumes Thompson afterwards. Both controls use the same tail at the
actual end. Comparing them isolates the added event boundary.
| Extra delay | Policy | Event response, % | Post-event response, % | Combined response, % |
| --- | --- | --- | --- | --- |
| None | Greedy from start | 71.71 | 79.05 | 75.38 |
| None | Random 1 day | 69.80 | 80.31 | 75.06 |
| None | Thompson | 59.03 | 73.51 | 66.27 |
| None | End-only terminal tail | 59.03 | 73.89 | 66.46 |
| None | Event and end tails | 59.50 | 73.67 | 66.59 |
| One day | Greedy from start | 67.76 | 81.74 | 74.75 |
| One day | Random 1 day | 65.99 | 81.26 | 73.62 |
| One day | Thompson | 54.59 | 72.21 | 63.40 |
| One day | End-only terminal tail | 54.59 | 74.11 | 64.35 |
| One day | Event and end tails | 57.02 | 73.18 | 65.10 |
Against the matched end-only-tail control, the extra event window reduces
post-event response by 0.22 percentage points, with a paired 95% interval
of [-0.34, -0.12]. With one day of extra delay, the reduction is
0.93 points [-1.04, -0.81]. Those comparisons show a later cost from the
earlier change in exploration.
The immediate gains outweigh that cost over the tested continuation.
Combined response rises by 0.13 points [0.07, 0.18] without extra delay
and 0.75 points [0.69, 0.81] with one day of extra delay. These results
use equal weight per display for the event and the following seven days.
The paired control comparison was added during manuscript QA using the
existing runs. Comparing only with ordinary Thompson would also change
behaviour at the final boundary and obscure the post-event cost.
Thompson is not an optimal finite-horizon policy. New observations can
have some future value while still costing more than they return.
The terminal-tail argument identifies when feedback has no remaining
use inside the objective; it leaves the earlier schedule open.
## 8. What a platform evaluation would need
Public popularity charts cannot supply the missing counterfactuals.
A demo with few concurrent players may have had few displays. It may have
received poorly matched displays, or players may have tried it and left.
The same count can be consistent with different mechanisms.
Li and colleagues' offline contextual-bandit work makes the partial-label
problem explicit. Outcomes are observed for the actions taken by a logging
policy, not every action a new policy might choose. Their replay evaluator
uses random logging under stated assumptions to obtain an unbiased
evaluation. They also discuss the modelling bias introduced by
simulators.\[9\]
A credible empirical follow-up would need a platform experiment or
appropriate assignment logs. Each record should identify the discovery
surface and eligible candidates, along with the chosen item and its
assignment probability. The evaluation needs a defined response outcome
and separate timestamps for display and feedback ingestion.
Existing audience evidence needs a defined cut-off. An algorithm cannot
be evaluated as cold-start if it receives future observations or private
labels unavailable to its comparator.
For a real player study, the outcome would need to be grounded in the
experience of trying a demo. A tile click can be an early signal.
Using it as a surrogate for later satisfaction needs separate validation.
The experiment would also need to define its trade-off. A minimum
acceptable player outcome can be a guardrail while discovery coverage is
examined separately. Equal exposure is a different policy objective and
should be justified as such.
Collection of player-level records would require appropriate privacy
controls. Nothing in this paper calls for scraping private developer
dashboards or assembling identities across public services.
The earlier game-discovery paper asks whether claims about a game carry
adequate source and scope information.\[10\] Better descriptive evidence
may improve retrieval. Attention allocation still needs its own evaluation.
## 9. The strongest counterargument
This model may be too simple to select a production policy.
Real recommendation systems can transfer information across related
items. They have richer player histories, and candidate retrieval can
adapt as evidence changes. A game can appeal to several audiences.
Playing a demo also consumes a different amount of time from seeing a
tile.
The study omits those mechanisms. Repeated-user fatigue is absent. Every
synthetic display has the same budget cost, and responses are stationary
conditional on an audience-item pair. A policy's choices do not change
later demand or developer behaviour.
These assumptions are especially important for catalogue scaling.
Independent pair beliefs make a new item expensive to learn about.
A production system with effective transfer may need much less item-specific
feedback. Conversely, a badly misspecified model can require more
careful investigation than a count floor suggests.
The common candidate generator isolates allocation within the model.
It also excludes a potentially decisive part of a real system. A policy
cannot choose a relevant game that its retrieval stage never supplies.
The study provides conditional trade-offs and an auditable test harness.
Choosing a production policy would require an evaluation that represents
the omitted mechanisms and actual player outcomes.
The baseline set is also limited to recognisable policies and a few
random-then-greedy schedules. More capable contextual-bandit methods
remain to be compared.
## 10. What this paper does not claim
This paper does not establish that Steam stops exploring too early.
It does not identify a hidden threshold at which a festival catalogue
becomes too large.
It does not introduce deadline-sensitive exploration or reproduce the
item-centric filter in the cited industry paper. The count quotas are
simple controls, not a novel information-gain algorithm.
It does not measure recommendation quality, wishlist conversion or
commercial success on Steam. The relevant-pair labels exist only in the
synthetic evaluator.
It does not establish that reducing exposure concentration helps every
player, or that every game deserves the same observation budget.
It makes no claim about long-term creative diversity or developer
incentives.
## Conclusion
Count-based coverage targets can exceed the available observation budget
before any ranking policy is chosen. The simple budget bound makes that
constraint explicit.
Within the synthetic model, increasing a quota can raise delivered count
coverage while reducing expected response substantially. A smaller quota
improves both headline measures against Thompson in some conditions.
The policy ordering changes with the available attention and feedback
timing.
The continuation comparison shows why the objective matters. An extra
event-tail window reduces post-event response against the matched
end-only-tail control. Its immediate gains still improve combined
response over the tested continuation. That ordering could change with
different continuation weights or duration.
Before choosing an exploration schedule, specify which future decisions
count and when feedback reaches them. A coverage target can then be
judged against its cost and the opportunities left to use what it reveals.
## Reproducibility
The [study bundle](./discovery-study.zip) contains the protocol, executable
policies and raw results. The same files are available in the
[artifact directory](./artifact/README.md).
The primary records are [individual runs](./artifact/results/raw.csv) and
[daily checkpoints](./artifact/results/daily.csv).
[Paired comparisons against Thompson](./artifact/results/paired-vs-thompson.csv)
retain uncertainty across world seeds. The
[matched continuation comparison](./artifact/results/paired-event-vs-end-tail.csv)
isolates the extra event-tail window.
Run the focused tests, create a fresh result directory with `study.py`,
then run `analyse.py` against that directory. Exact commands and
dependencies are in the artifact README. No network access or private data
is required for the experiment.
## About the author
Jason Doyle writes about reliable software systems and applied AI.
He publishes at [jasondoyle.ie](https://jasondoyle.ie).
## References
1. Valve, *Steam Next Fest*, accessed 22 October 2026,
[https://partner.steamgames.com/doc/marketing/upcoming_events/nextfest](https://partner.steamgames.com/doc/marketing/upcoming_events/nextfest).
2. Valve, *Steam Next Fest: October 2026 Edition*, accessed 22 October 2026,
[https://partner.steamgames.com/doc/marketing/upcoming_events/nextfest/2026october](https://partner.steamgames.com/doc/marketing/upcoming_events/nextfest/2026october).
3. Output Lag, *Steam Next Fest Live Chart*, accessed 22 October 2026,
[https://outputlag.com/news/steam-next-fest-live-chart/](https://outputlag.com/news/steam-next-fest-live-chart/).
4. SonOfSaris, *Viewer for Steam Next Fest*, README at commit
`c272801948b81dabe951d09d3022fb894cdde80d`, accessed 22 October 2026,
[https://github.com/SonOfSaris/ViewerForSteamNextFest/blob/c272801948b81dabe951d09d3022fb894cdde80d/README.md](https://github.com/SonOfSaris/ViewerForSteamNextFest/blob/c272801948b81dabe951d09d3022fb894cdde80d/README.md).
5. Stefano Traca, Cynthia Rudin and Weiyu Yan, *Reducing Exploration of Dying
Arms in Mortal Bandits*, UAI 2019,
[https://arxiv.org/abs/1907.02571v1](https://arxiv.org/abs/1907.02571v1).
6. Dong Wang and colleagues, *Item-centric Exploration for Cold Start
Problem*, ACM RecSys 2025 Industry Track,
[https://arxiv.org/abs/2507.09423v1](https://arxiv.org/abs/2507.09423v1).
7. Kelly W. Zhang and colleagues, *Impatient Bandits: Optimizing for the
Long-Term Without Delay*, Journal of Machine Learning Research 27(156),
2026, pp. 1-58,
[https://jmlr.org/papers/v27/25-0030.html](https://jmlr.org/papers/v27/25-0030.html).
8. Daniel Russo and colleagues, *A Tutorial on Thompson Sampling*,
arXiv:1707.02038v3,
[https://arxiv.org/abs/1707.02038v3](https://arxiv.org/abs/1707.02038v3).
9. Lihong Li, Wei Chu, John Langford and Xuanhui Wang, *Unbiased Offline
Evaluation of Contextual-bandit-based News Article Recommendation
Algorithms*, arXiv:1003.5956, 2010,
[https://arxiv.org/abs/1003.5956](https://arxiv.org/abs/1003.5956).
10. Jason Doyle, *Game Discovery Is an Evidence Problem*, 28 September 2026,
[https://jasondoyle.ie/whitepapers/game-discovery-is-an-evidence-problem/](https://jasondoyle.ie/whitepapers/game-discovery-is-an-evidence-problem/).
