Evidence-backed game research, from a game name to traceable player insight.
PlayerVoice resolves multilingual game names, retrieves official facts and public community evidence, then turns relevant evidence into inspectable insights for players, creators, and analysts. 四种使用视角共享同一个 Intelligence Core；结论保留来源、样本范围与不确定性。
Game titles and aliases vary by language, while store facts and player opinions come from different sources. PlayerVoice connects those steps without treating missing data as a negative finding or presenting a self-selected public sample as representative.
Game Name → Resolver → Alias Validation → Query Plan → Real Community Evidence
→ Relevance Filter → Deduplicated Corpus → Structured Insight → Original Source
Official facts are retrieved as a separate Fact Layer; they are not mixed with player evidence. The runnable path is:
flowchart LR
A[Game name] --> B[Resolver]
B --> C[Alias validation]
C --> D[Multilingual query plans]
D --> E[Official fact retrieval]
D --> F[Community source adapters]
E --> G[Fact Layer]
F --> H[Normalized evidence]
H --> I[Relevance filter]
I --> J[Deduplicated corpus]
J --> K[Annotations and bounded Insights]
G --> L[Saved Intelligence Core]
K --> L
L --> M[Player]
L --> N[Creator]
L --> O[Analyst]
L --> P[Compare]
M -. evidence IDs .-> Q[Evidence Store]
N -. evidence IDs .-> Q
O -. evidence IDs .-> Q
P -. evidence IDs .-> Q
Q -. source URL or reference .-> R[Original source]
The diagram describes the current implementation: Modes render a saved core snapshot; they do not launch four independent collection pipelines. See Architecture, Data Schema, Specification, and Issue backlog.
| Slice | What it does | Example |
|---|---|---|
| Foundation (Issues 1–5) | Resolve identity, validate aliases, build safe multilingual plans, retrieve sourced Steam facts; does not collect UGC. | examples/foundation/ |
| Voice Slice (Issues 1–10) | Collect community evidence, record coverage, filter and deduplicate, annotate, and generate evidence-linked Insights. | examples/e2e/ |
| Mode Layer (Issues 10.5–14) | Render a saved Voice Slice snapshot for one audience or compare multiple games. | examples/modes/ |
The two end-to-end commands and a no-recollection Mode render are shown in Quick Start. Voice Slice collection uses live public routes and can be partial because of credentials, rate limits, or source availability.
All four Modes load the same Intelligence Core snapshot, including its facts, evidence store, annotations, Insights, and source coverage.
| Mode | Target user | Main output | Evidence handling | Uncertainty handling |
|---|---|---|---|---|
| Player | Players evaluating fit or practical requirements | Fact and player-evidence summary, preference profile, dimension-level fit | Claims link to Fact/Insight IDs and underlying evidence | Unsupported preferences stay uncertain; no universal recommendation score |
| Creator | Game creators and researchers | Praise, complaints, supported controversy, research questions and gaps | Each claim cites evidence; controversy requires distinct evidence for both positions | Weak disagreement is not promoted to controversy; questions are not findings |
| Analyst | Product and research analysts | Evidence-to-interpretation chains, requests, needs, expressed intent, hypotheses | Each stage carries epistemic status and evidence IDs | Public statements are not observed behaviour; business impact remains hypothesis |
| Compare | Players, creators, or analysts comparing games | Aligned Fact and taxonomy dimensions with sample coverage | Cells retain evidence, source/language mix, dates, and confidence | Thin or imbalanced samples are insufficient or caution-labelled; no winner score |
Runnable JSON and Markdown outputs are in examples/modes/. Individual implementation reviews: Player, Creator, Analyst, and Compare.
The checked-in Issue 6–10 runs use retrieved public community records, not synthetic demo data. Counts below describe those particular snapshots, not player populations or current platform-wide totals.
| Game | Raw community items | Unique after deduplication | Relevant corpus | Checked-in report |
|---|---|---|---|---|
| Infinity Nikki | 52 | 32 | 26 | evidence_report.md |
| PUBG: BATTLEGROUNDS | 43 | 23 | 20 | evidence_report.md |
| Counter-Strike 2 | 46 | 16 | 10 | evidence_report.md |
These runs show why the relevance and provenance steps matter: ambiguous aliases are scoped, unrelated search results are excluded, and duplicate returns retain their query links without inflating the corpus. The checked-in three-game comparison marks all community dimensions insufficient where its evidence floor is not met; it does not force a ranking. See the Issue 1–5 review, Issue 6–10 review, and Mode Layer review for methods and audits.
- Multilingual game resolution and evidence-backed alias discovery with ambiguity-safe search.
- Multilingual, source-aware query planning and official/store Fact retrieval.
- Community adapters for Steam and Bilibili; an OAuth-aware Reddit adapter.
- Relevance filtering, reviewable exclusions, source-ID deduplication, and a normalized corpus with retained query provenance.
- Versioned taxonomy annotation and evidence-linked, bounded Insights.
- Player, Creator, public-corpus Analyst, and Compare renderers on the shared Intelligence Core.
Every rendered conclusion can be followed as:
Insight → Evidence ID → stored original text and provenance → Original Source URL/reference
The Evidence Store review describes storage and lookup. The report renderer validates references rather than merely printing evidence-looking IDs.
Keep these layers and labels distinct:
| Label | Meaning |
|---|---|
| Fact Layer | Sourced, verifiable claims such as platform support or system requirements. |
| Player Evidence Layer | Public community records and Insights derived from them; not official facts. |
| Observed | What a source states, including an explicit request or expressed intent. It does not prove the described action occurred. |
| Inferred | A taxonomy label, pain point, need, or opportunity interpreted from evidence. |
| Hypothesis | Possible product or business relevance that requires separate validation. |
Analyst output labels behavioural statements as expressed intent, not observed churn. A complaint or association is not evidence of causation.
Requires Python 3.12 as configured in CI. The commands below use the existing scripts/gamepulse.py CLI.
py -3 -m venv .venv
.\.venv\Scripts\Activate.ps1
py -3 -m pip install -r requirements.txt
py -3 -m unittest discover -s tests -v
# Foundation: identity, aliases, query plans, and facts; no community collection.
py -3 scripts/gamepulse.py foundation --game "无限暖暖" --locale zh-CN
# Voice Slice: live community collection; source results may be partial.
py -3 scripts/gamepulse.py voice-slice --game "无限暖暖" --output runs/infinity-nikki
# Render Player Mode from the checked-in snapshot; no recollection.
py -3 scripts/gamepulse.py mode --mode player --input-dir examples/e2e/infinity-nikki --preferences "我每天只有一个小时，喜欢剧情和画面，不喜欢重 PvP" --output runs/player-infinity-nikkipython3 -m venv .venv
source .venv/bin/activate
python3 -m pip install -r requirements.txt
python3 -m unittest discover -s tests -v
python3 scripts/gamepulse.py foundation --game "无限暖暖" --locale zh-CN
python3 scripts/gamepulse.py voice-slice --game "无限暖暖" --output runs/infinity-nikki
python3 scripts/gamepulse.py mode --mode player --input-dir examples/e2e/infinity-nikki --preferences "我每天只有一个小时，喜欢剧情和画面，不喜欢重 PvP" --output runs/player-infinity-nikkiThe committed baseline is 149 tests; the CI workflow runs the same unittest suite. voice-slice accesses configured/public source routes. Reddit requires OAuth credentials; Bilibili may return partial or rate-limited results. Failures are recorded as unavailable/partial rather than interpreted as no player feedback. The Mode command reads the existing snapshot and does not need source credentials.
| State | Scope |
|---|---|
| Implemented | Resolver and alias validation; bilingual query planning; Steam Fact retrieval; Steam/Bilibili collection adapters; relevance decisions; SQLite Evidence Store, deduplication, and query links; versioned baseline annotations; evidence-linked Insights and coverage; all four shared-core Modes and CLI outputs. |
| Partially implemented | Transparent taxonomy and bilingual rules have limited long-tail/slang recall; Bilibili availability is variable; Reddit adapter execution needs OAuth and was unavailable in the cited validation. |
| Interface/spec only | Restricted-source connections without approved credentials/provider/export; authorized private-data ingestion and private/public joins. Contracts and unavailable/fallback boundaries do not mean a connected service or enterprise workflow exists. |
| Deferred / not implemented | Dashboard, API server, vector database, private enterprise ingestion, Patch Intelligence or causal patch analysis, and Issue 15+ features. |
No adapter bypasses login, CAPTCHA, rate limits, or platform access controls. Authorized exports or approved providers are required where a source restricts access; graceful fallback and truthful coverage are intentional behavior. No Issue 15 work is included in this repository state.
- Modular Python architecture with one shared Intelligence Core.
- Unified typed domain models and checked-in JSON Schemas.
- Provenance-first evidence storage and executable Insight traceback.
- Multilingual resolution, query planning, and baseline annotation.
- Ambiguity-safe alias validation and deterministic failure states.
- Deterministic, network-independent tests; 149/149 baseline.
Design and implementation references: Architecture, Data Schema, taxonomy, source adapter contract, taxonomy gap analysis, and UGC annotation notes.
Synthetic behavioural telemetry is retained only under demo/ for offline SQL demonstrations. Third-party source content remains subject to its platform and author rights. The code and original documentation are licensed under MIT.