Title: R.E.P.O. (AppID 3241660) — semiwork Type: Full-census second-title sample. Supersedes the short version of 2026-11-18 (SAMPLE_REPORT_REPO_SHORT_2026-11-18.md) as the depth inventory for this title; the short version remains the dated October snapshot. Purpose: test, on a second unrelated title, whether the patterns found in the PEAK full census (v0.2) are a category pattern or a one-title accident — the stated precondition for any category claim in Sprint Item 2. Date of this sample: 2026-11-19 Price point this format is testing: $49/month (Solo, 1 title) / $99/month (Studio, up to 5 titles) Research note: Research-phase sample. No sales claim is made. Every figure below cites its source and snapshot date. Where a figure is an estimate, the method and its limits are stated next to it. Sections that could not be verified are labelled as such, rather than filled in.
Correction 1 (to the R.E.P.O. short sample, §2): The short sample assumed Japanese-language reviews "could only begin accumulating in volume after 29 Oct 2026", when official Japanese arrived. The full census disproves that. Japanese players had already written 1,764 Japanese-language reviews before localisation (26 Feb 2025 – 28 Oct 2026), playing in English or via the community Japanese MOD (61 reviews explicitly mention MODs, all pre-localisation). The pre-localisation Japanese demand was not a trickle waiting to be unlocked; it was a standing, measurable audience that the developer could not read without Japanese. The short sample's §2 framing must not be quoted.
Correction 2 (production method, applies to all titles): In IUserReviewsService/GetAppReviews, omitting the languages parameter does not return all languages — it silently returns English only (for R.E.P.O.: 141,890 reviews, identical to languages=["english"]). The all-language total requires languages=["all"] (350,812). This is the same class of silent failure as the input_json wrapping issue recorded in the classification pilot (2026-11-19 04:24). Both are now recorded production rules: wrap in input_json, and always name the language set explicitly.
| Metric | Value | Source / snapshot |
|---|---|---|
| Total Steam reviews (all languages) | 350,812 (336,963 positive / 13,849 negative, 96.1% positive) | Steam Web API query_summary, languages=["all"], 2026-11-19 |
| English-language reviews | 141,890 (137,181 / 4,709, 96.7% positive) | Same, languages=["english"] |
| Japanese-language reviews | 1,804 (1,670 positive / 134 negative in query_summary; 1,669 / 135 in the fetched census — a 1-review difference, likely a vote changed between calls) | Same, languages=["japanese"] |
| Japanese share of reviews | 0.514% (1,804 / 350,812) | Calculated from the two rows above |
| Japanese positive rate | 92.6% vs 96.1% all-language and 96.7% English | Same, 2026-11-19 |
| Release / price | 26 Feb 2025 (Early Access) / US $9.99; Japan Steam price ¥1,200 | Steam listing via trackers and Japanese coverage (short sample, 2026-11-18) |
| Japanese support | First official Japanese (interface + subtitles, 21 languages) in v0.5.0 "The Sunstone Update", 29 Oct 2026 — ~20 months after launch | SteamDB patch notes; short sample §3 |
| Japanese review volume by month (census) | 2025: Mar 59 / Apr 99 / May 150 / Jun 121 / Jul 119 / Aug 164 / Sep 101 / Oct 102 / Nov 241 / Dec 96; 2026: Jan 74 → Oct 32 per month, Nov (to 18th) 37 | Full census, §2 |
Reading: Japan is 0.514% of R.E.P.O.'s review volume — a higher share than PEAK (0.356%) — and Japanese reviewers rate it 3.5 points lower than the global average (PEAK's gap: 6.2 points). The direction of the gap is the same on both titles; its size is title-specific. As with PEAK, the score itself is now free in Steam; why the gap exists, and whether its causes are moving, is §2.
Population: 1,804 Japanese reviews, posted 2025-02-27 to 2026-11-18 (UTC). Two stratifications (R.E.P.O. has two meaningful cut points, so both are reported):
Counting: Two layers, both keyword-based and reproducible. Strict counts require an explicit word. Broad counts use a wider keyword set per theme and overlap (one review can count in several themes). These are mention rates, not sentiment scores. A structural caveat PEAK did not have: the median Japanese review for R.E.P.O. is 20 characters (PEAK pilot median: 163). A large share of this census is one-liners ("最高", "良", "ok"). The classification is therefore thinner per review: counts are reliable, but there is less text per review to interpret, and quote selection draws on a minority of substantive reviews. Median playtime at review is 1,695 minutes (~28 hours) — the reviewers are engaged players writing briefly, not drive-by accounts.
| Indicator | All 1,804 | Early A 1,004 | Recent A 800 | Negative reviews only (135) |
|---|---|---|---|---|
| Friends explicit (友達/友人/フレンド) | 303 (16.8%) | 166 (16.5%) | 137 (17.1%) | 10 |
| Horror / scary explicit (怖い/ホラー/恐怖) | 191 (10.6%) | 100 (10.0%) | 91 (11.4%) | 7 |
| Difficulty explicit (難易度/難しい etc.) | 108 (6.0%) | 63 (6.3%) | 45 (5.6%) | 11 |
| "Japanese" explicit (日本語) | 60 (3.3%) | 36 (3.6%) | 24 (3.0%) | 0 |
| MOD explicit | 61 (3.4%) | 41 (4.1%) | 20 (2.5%) | 4 |
| Voice chat explicit (VC/ボイチャ etc.) | 51 (2.8%) | 35 (3.5%) | 16 (2.0%) | 5 |
| Pause / save / session explicit | 32 (1.8%) | 22 (2.2%) | 10 (1.3%) | 9 |
| No public matchmaking explicit (野良/マッチング) | 31 (1.7%) | 15 (1.5%) | 16 (2.0%) | 4 |
| Performance narrow (重い/最適化/フリーズ/クラッシュ etc.) | 30 (1.7%) | 19 (1.9%) | 11 (1.4%) | 9 |
| Bug explicit (バグ) | 20 (1.1%) | 16 (1.6%) | 4 (0.5%) | 5 |
| Motion sickness explicit (酔い incl. 画面酔い) | 15 (0.8%) | 9 (0.9%) | 6 (0.8%) | 2 |
| Price / value explicit | 13 (0.7%) | 10 (1.0%) | 3 (0.4%) | 1 |
| Theme | All | Early A | Recent A |
|---|---|---|---|
| Core praise words | 641 (35.5%) | 369 (36.8%) | 272 (34.0%) |
| Co-op / friends mention | 528 (29.3%) | 296 (29.5%) | 232 (29.0%) |
| Update / new content mention | 321 (17.8%) | 193 (19.2%) | 128 (16.0%) |
| Solo mention | 276 (15.3%) | 147 (14.6%) | 129 (16.1%) |
| Horror / atmosphere | 204 (11.3%) | 107 (10.7%) | 97 (12.1%) |
| Difficulty / balance (broad) | 180 (10.0%) | 111 (11.1%) | 69 (8.6%) |
| Extraction / loot mention | 149 (8.3%) | 92 (9.2%) | 57 (7.1%) |
| Session length / time | 107 (5.9%) | 57 (5.7%) | 50 (6.3%) |
| Language / UI / settings (broad) | 99 (5.5%) | 60 (6.0%) | 39 (4.9%) |
| Network / matchmaking (broad) | 82 (4.5%) | 52 (5.2%) | 30 (3.8%) |
| Voice / proximity chat | 80 (4.4%) | 53 (5.3%) | 27 (3.4%) |
| Bug / performance (broad) | 78 (4.3%) | 52 (5.2%) | 26 (3.3%) |
Note on the session row: the broad session set overlaps heavily with general time words; the strict pause/save row in §2a (1.8%) is the quotable figure.
| Focus | All | Negatives | Recent A | Post-localisation (n=40) |
|---|---|---|---|---|
| New map / new stage / Sunstone mentions | 9 | 8 of 9 | 7 (7 of 60 recent negatives) | 5 (5 of 11 post-loc negatives = 45%) |
| Pitfall / fall deaths (落とし穴/奈落/落下 etc.) | 66 (3.7%) | 16 | 27 | 2 |
| Client timeout / connection errors | 12 | 6 | 0 | 0 |
New map / Sunstone (negative, post-localisation):
Language demand (positive, pre-localisation): see §2d-1 for three dated examples (2025-04-17 MOD, 2025-04-19 request, 2026-06-07 kicked-as-foreigner).
Technical class that disappeared (early window):
Horror balance as appeal (positive):
PEAK v0.2 §8 stated: "a second full census is required before claiming a category pattern." That census now exists. Result:
| Pattern element | PEAK | R.E.P.O. | Verdict |
|---|---|---|---|
| Japanese positive rate below global | 88.4% vs 94.6% (−6.2pt) | 92.6% vs 96.1% (−3.5pt) | Replicates (direction same, size title-specific) |
| Complaint centre moves over time | Early bugs/performance/language → recent single-biome balance | Early timeout/bugs → recent/post-loc single-map balance | Replicates |
| Language mentions concentrate before/around localisation, then stop | All 10 in early window, 0 in recent window | All 60 pre-localisation, 0 post-localisation (n=40 caveat) | Replicates |
| One new map/biome dominates recent negatives | Deep Forest: 11/52 recent negatives (21%) | Sunstone/new map: 5/11 post-loc negatives (45%); 7/60 recent-A negatives (12%) | Replicates in structure (share varies with window size) |
| Early technical complaint class shrinks | Performance strict 3.0% → 0.2% | Bug strict 1.6% → 0.5%; timeout 12 → 0 in recent window | Replicates |
| Japanese share of reviews | 0.356% | 0.514% | Title-specific (no pattern claimed) |
| Pre-localisation Japanese-language demand | Small (Japanese added ~1 month after launch) | Large and long: 1,764 reviews written in Japanese over 20 months without official support | Difference, and itself a finding: the later the localisation, the larger the unread Japanese record a monthly service would have been reading |
Conclusion: on two full censuses, the core Radar claim — the Japanese complaint centre moves, language demand is visible in Japanese reviews long before it is answered, and a free language score shows none of this — is a category pattern for these co-op titles, not a PEAK-only anecdote. This still does not prove anyone will pay $49/month (Sprint Premise 1, confidence 25%, unchanged — this is internal production, not customer data).
Proves:
Does not prove:
IUserReviewsService/GetAppReviews v1, keyless, input_json wrapping, filter=Recent, languages=[japanese], day_range=0, cursor pagination — 20 calls (19 pages + 1 empty terminator), 1,804 unique reviews. Snapshots for all-language (languages=["all"]), English and Japanese totals: 3 further calls. Total ≈23 calls, far below the 100,000 calls/day API terms limit. Cash cost ¥0. No 429s.Status note: This report completes the Sprint Item 2 inventory requirement "second title, full census" and the precondition for the category-pattern claim. It does not move any sprint gate — gates move only on customer signals in the signal hierarchy (payment > prepayment > post-price registration > intent).