Side Quest Nexus
Checkpoint icon: a plum lantern holding a crystal, on a gold tile with a faint constellation Checkpoint In private beta

It’s dangerous to go alone. Take this.

Game answers that know the patch. And who wrote them down.

Checkpoint is a research system, not a chatbot. It builds a structured knowledge base for each game, keeps it in step with each patch, and shows its sources on every answer.

Ask it the way you'd ask a friend who plays. It looks things up, checks what changed, and leads with a straight answer. Then it shows its work.

A general chatbot Illustrative
What changed in patch 1.3.11?

Patch 1.3.11 cut all trade income by 25% and removed merchant upkeep.

  • No source
  • Patch: unknown
Same question
Checkpoint Live patch
Checkpoint answering 'What changed in patch 1.3.11?': the release date, then the trade changes, each cited to the official patch notes.

Why it's different

Not a chatbot. A research system with a memory of the game.

A model can talk about a game. Checkpoint can look things up, and say where they came from.

  1. Structure first

    Items, bosses, laws and buildings are typed entities with attributes and relations. “Which items do X” is a lookup, not a guess.

  2. Patch history built in

    Every value carries the patches it applies to. Patch notes link to the things they change. Open betas stay apart from the live game.

  3. Research, in order

    Structured data first, wiki text second, the web third. An editor pass merges them, and answers say when the web contributed.

  4. Shows its work

    Every claim is cited inline. Sources are listed by name under each answer.

How an answer is built, in four layers From the bottom: sources (community wiki, official patch notes, web) feed an ontology of typed entities with patch ranges. A research agent queries it with structured lookup, wiki text and web research, then an editor pass merges the results into the answer, whose citation chips trace back to the sources. ANSWER 1 2 3 RESEARCH AGENT Structuredlookup Wikitext Webresearch 1st 2nd 3rd Editor pass ONTOLOGY patch range Boss Item Law Building SOURCES Community wiki Official patch notes Web

The knowledge base

A knowledge base for every game, built from its own community.

Each game gets its own ontology, designed for how that game works.

Weapons mean one thing in one game and laws mean another in the next. So an architect agent reads how the wiki is organised and designs the entity types, attributes and relations to fit. Values carry patch ranges, so Checkpoint can say what something did before and what it does now.

Checkpoint's ontology explorer for one game: a graph of entity types such as country, law, building, technology, company and good, with entity counts and the relations between them.
One game's knowledge base
  • 7,450entities
  • 28entity types
  • 81%attribute coverage

One attribute, across patches: “Damage bonus” Illustrative

  1. 1.0–1.215%
  2. 1.3–1.812%
  3. 1.9 → live10%

Onboarding

From a game's name to a working knowledge base, run by agents.

One human approval. Everything after it runs on its own, and every step is recorded.

Agents find the game, scout its wikis and official channels, and audit each source's licence and terms, including AI and text-mining reservations. A person approves the sources. Then agents discover how the wiki is structured, design the ontology, ingest the pages, build the patch history and fill the gaps. A full onboarding costs a few dollars of model spend, so adding the next game is a question of demand, not budget.

  1. Steam intake
  2. Source scouting
  3. Licence audit
  4. Gate 1Human approval
  5. Template discovery
  6. Architect
  7. Ingest
  8. Patch history
  9. Gap filling

Steward and schema agents keep improving it

Agent step Person

A layer that changes itself, safely

It edits itself. Nothing unrecorded, nothing irreversible.

Every change to a game's knowledge is traced, versioned and revertible.

The structure is a projection of a versioned spec, so a rollback is exact, not approximate. Each change is attributed to the agent that made it (architect, autoschema, triage, fill, guard, steward) or to a person. When quality checks raise a warning, a steward agent investigates and proposes or applies a fix inside the same guard.

Change logIllustrative
VersionMade byChangeStatus
v1architectInitial entity types, attributes and relationsGuard passed
v2autoschema15 new types, new attributesGuard passed
v3autoschemaRelations, 24 value-link rulesGuard passed
v4fillFilled missing attributesRolled back to v3answers regressed
v5stewardFixed leftover markup in valuesGuard passed

Evaluation

How we know an answer is better than googling it.

An evaluation system built for a hard question: is this answer actually better, and did the last change help or hurt?

A knowledge base can look healthy while its answers get worse. So Checkpoint is tested from several angles.

The evaluation loop for every change A proposed change goes through deterministic checks, then a noise-aware evaluation in which two unlabelled answers, A and B, are shown to a judge in both orders and disputed facts are verified. A decision follows: the change stands, or it is rolled back and revised. Revise and try again Proposed change Deterministic checks Noise estimate Noise-aware eval A B Blind,both orders Verifiesdisputed facts Judge sees unlabelled answers Better,not worse? Stands No Yes Rolled back automatically

Real questions

A benchmark written the way players ask, in Reddit and Google phrasing, written before any answer is seen. Each game also has its own golden set.

A strong opponent

Each question is answered head-to-head against a frontier model with web search and no Checkpoint tools. The bar is better than googling it.

A blind, careful judge

A pairwise judge sees the two answers unlabelled, runs in both orders to cancel position bias, and checks disputed facts with its own web searches.

Noise, accounted for

Models vary from run to run. We estimate that variance, so noise isn't mistaken for progress.

A gate on every change

Before a semantic-layer change stands, a guard runs deterministic checks, then a noise-aware eval. Regressions are rolled back automatically.

Down to the value

Data-quality checks run across every value and chunk, every answer keeps a trace of its research steps, and cost is tracked per model.

Credit

Built on community work. Credited like it.

Wikis and guides take years of unpaid care. Checkpoint puts them first.

Every answer names its sources and links to the pages used, with licence lines where they apply. A citation ledger counts how often each source is cited and how many answers it fed. Visits are counted anonymously: the site, the game and the day, never who. The numbers are public every month on Where answers come from.

The knowledge is theirs. We say so on every answer.

How credit flows A community wiki feeds an answer that carries a citation chip. Citations are counted in a ledger, which publishes to a public sources page. A player correction goes to a person for review and then to the wiki's editors, who improve the wiki. A dashed branch from the ledger, which we are building towards, leads to letting sources claim themselves. Community wiki used and cited Answer 1 2 every citation counted Ledger published monthly Publicsources page Claim your source Building towards Playercorrection Reviewed by a person Wiki editors

  • Credit on every answer

    Sources named and grouped. Wiki text carries its CC BY-SA line.

  • Rules respected

    Licences and AI reservations are audited, and a person approves each source.

  • Corrections go home

    When a player flags an error, Checkpoint traces it to the wiki page and drafts a suggested edit for that wiki's editors. A person reviews it first.

  • Building towards

    Sources claiming themselves and setting their preferences, and, as Checkpoint grows, sharing what it earns with the wikis and creators it relies on. That's our aim; we'll say more when there's something real to share.

The Sources section under an Elden Ring answer: five Elden Ring Wiki pages with their sections, the line 'Text from the Elden Ring Wiki (CC BY-SA)', then four web guides grouped by site, each dated and with an outbound link.
Sources named under every answer.
The public 'Where answers come from' page: this month's answers, sources cited and visits sent, then a table of the sources answers drew on, with the games and answer count for each.
The public page, updated monthly.

Trust and status

Private beta, and plain about it.

No ads. Your data isn't sold. Your chats are private, and you can delete them. Checkpoint runs on cost controls: answers are cached per patch, cheap models handle labelling, and daily spend caps and per-account limits apply. Players can request games from inside the app.

  • No ads, and your data isn't sold
  • Your chats are private, and you can delete them
  • Daily spend caps and per-account limits apply

If you run a wiki, write guides, or work at a studio or publisher, write to hello@sidequest.nexus.