# Build destinations — September 6, 2026

Quantitative follow-up on where advertised build expression is conventional,
associated with greater review volume within a host activity, and repeated among
several independently credited developers. This study does **not** use individual
review text or positivity for its rankings or conclusions. All work is local to the
September 5 snapshot; no network collection was performed.

## Read the measurement correctly

The strict classifier detects explicit build/deck promises, equipment loadouts,
configuration synergies and class combinations. It has **no genre prerequisite**.
The extended screen also detects broader customization, skill trees, lineup choices
and functional parts. Extended matches are candidates, not a verified upper bound
on true build prevalence. The direct-word variant is a further diagnostic.

Neither classifier verifies actual depth, viable alternatives, player expectations
or the heavy/light distinction in gameplay. No detected promise does not mean no
build system. This especially matters in fields where configuration may be assumed
and not explained in the pitch. Low wording prevalence is not an underserved market.
This is a new, stricter operational definition; do not directly compare its counts
with earlier classifiers as if the market had changed between runs.

Genres/activities are the 34 explicit tag groups in `design.py`. Main membership
requires a top-10 tag **and** matching activity language in the short description
or first 350 words of the full description. This removes some obviously unsuitable
tag matches, but reduces coverage and can miss valid games. These are confirmed
store-description samples, not exhaustive genre censuses. Top-5, top-20 and tag-only
versions are retained. Categories overlap; their counts cannot be summed.

## Population

The 75,860 records are currently paid, dated released games from 2019–2026 August,
outside the previous explicit-content and central idle/incremental/FMV/visual-novel
exclusions. No manual quality filter is applied. There are 27,645 eligible records
in 2019–2022, 35,806 in 2023–2025, and 12,409 in January–August 2026.

The headline comparison uses 2023–2025. Older and newer cohorts are separate
consistency checks, not an untouched historical forecasting backtest. Review counts
are current lifetime totals, not reconstructed launch outcomes. Dates, prices and
descriptions may change, early-access/edition histories remain incomplete, and
delisted/unlisted titles may be missing. The previously verified KOF XV date override
is applied through `portfolios-2026-09-06/source_data.py`; the snapshot is untouched.

## Audit and corrections

The 96-case development audit sampled strict matches, broader-only matches and
non-matches across 12 host categories with numerical outcomes and prices hidden.
It exposed cosmetic mixing, technical-engine references, literal wooden-deck
construction, blackjack house-rule configuration and missed lineup/ship choices.
These informed general rule revisions. It also exposed incorrect activity tags,
including prominently tagged platformers whose descriptions describe management
or turn-based strategy.

The original packet and single-reader labels are frozen. The revised strict screen
retains 26 audited cases, all labeled as clear advertised setups. This is a
development diagnostic, **not** independent precision/recall certification.
Unclear broader cases remain; no game was played to verify the advertisement.
The original packet is a retained development input; deleting it and drawing a
new sample with the revised classifier would not recreate the original audit.

## Comparisons

Primary outcomes are counts reaching 50, 100, 200 and 1,000 filtered reviews,
medians and the low-review tail. Reviews measure response scale, not sales,
profitability, player enjoyment or the causal contribution of builds.

The first exploratory comparison used within-host date/price cells with >=20
controls and progressively broader fallback. Several apparent advantages weakened
with the stronger comparison below; these exploratory ratios remain in
`host_results.csv` and must not be confused with the headline matched ratios.

For the main matched comparison, each build-promising game receives up to five
games without a detected strict build promise, from different developer credits,
in the same current US price band and within 90 days of its release (30 days for
2026). At least three peers are required; unsupported games are excluded. Peers
can be reused across different build games. A separate missing-price category is
retained, with a known-price-only sensitivity. The three main sparse destination
results are unchanged by removing missing prices.

The expected count sums each focal game's peers' threshold rate. For example,
Sports has six qualifying build games against an expectation of 3.2 across the
17 matched cases. This is a descriptive weighted baseline, not an estimate of
how a hypothetical game would perform without builds. Matching does not control
quality, budget, promotion, every subgenre, or all creator differences.

Sensitivity settings use 3/5/10 peers, 60/90/180-day windows for 2023–2025,
14/30/60-day windows for 2026, same description-length bins, alternate build
definitions, tag-rank changes, known-price-only records, removal of leading
developers/publishers or top three build games, and creator-history exclusions.
Some stricter settings lose coverage; counts and coverage are always retained.

Creator identities use normalized reported names, **not Steam creator-page IDs**.
The earlier portfolio study showed that page IDs can merge unrelated developers.
The main per-game identity is the first reported credit; history flags consult all
developer credits. These are not verified team sizes, corporate ownership or
human identities. Earlier currently 1,000-review games are an observational
sensitivity, not release-time fame. No earlier visible credit does not prove an
inexperienced or unknown maker.

The familiar-build flag uses the fixed top-10 tag set: Action Roguelike, Roguelike
Deckbuilder, Deckbuilding, Auto Battler, Action RPG, Party-Based RPG, Tactical RPG,
Strategy RPG. The rogue flag uses the separately saved roguelike tags. Lack of
these tags is not proof of absent roguelike mechanics. The splits show where the
association resides in the catalog, not causal mediation.

## Sparse screen

The initial screen is preserved in `sparse_screen.csv`. The final screen replaces
the sparse-cell comparison with nearest-peer matching and requires, in 2023–2025:

- At least 100 confirmed host games.
- Strict build share <=15%, with at least ten strict matches.
- At least five games reaching 100 reviews from at least five developer names.
- At least 80% matching coverage, five qualifying matched games, and >=1.5 times
  the matched expected count.

All passing and failing rows are retained. Five pass: Action RPG, Sports,
Board/dice/tabletop, City/colony building and Real-time strategy. The RPG/RTS
references demonstrate why rare *wording* must not be equated with an unexpected
mechanical mixture. The less conventional leads require the detailed checks in
the report. Threshold selection and multiple exploratory comparisons preclude
treating these as proven inefficiencies or underserved markets.

## Reproduction and artifacts

Use `../.venv/bin/python` from this directory:

1. `prepare.py` builds the revised classifications, host indicators and specification.
2. `audit.py` materializes saved manual development judgments.
3. `analyze.py` generates full within-host exploratory comparisons and membership lists.
4. `followups.py` generates matched comparisons, peer IDs and creator-history splits.
5. `match_sensitivity.py` and `text_check.py` generate the stronger sensitivities.
6. `finalize.py` produces the final screen, overview, removal and known-price checks.
7. `charts.py` renders the two PNG/SVG figures.
8. `validate.py` checks count accounting, peer identity/price/date eligibility,
   expected-count reproduction, final screen rules, sampled classification,
   frozen audit material, source hashes and collection state.

Main files: `overview.csv`, `final_sparse_screen.csv`, `matched_results.csv`,
`matched_games.csv`, `match_sensitivity.csv`, `matched_text_length.csv`,
`matched_removals.csv`, `known_price_sensitivity.csv`, `jan_aug_supply.csv`,
`build_members.parquet`, `candidate_successes.csv`, `sports_all_cases.csv`,
`audit_blind.json`, `audit_labels.csv`, `audit_recheck.csv` and `validation.json`.
`findings.md` is the full report, also delivered substantively in chat.