How S.A.G.E.
learns from the web.

Governing principles

Six principles govern every ingestion decision we make. They are stated here in plain language; the binding internal versions appear in our policy document.

1. License before access. We do not ingest content unless we have a documented legal basis. "Publicly accessible" is not a license.
2. Transformation, not reproduction. We embed paraphrased summaries, not raw source text. Verbatim excerpts are stored only as short evidence snippets for citation back to you.
3. Attribution and link-back are mandatory. Every chunk of knowledge in S.A.G.E. carries a source URL. S.A.G.E. cites its sources visibly to users.
4. Creator priority. When given a choice between ingesting a creator's content without permission and inviting them into our paid Knowledge Keepers program, we choose the program. Always.
5. robots.txt and rate limits are non-negotiable. Any source that prohibits crawling is excluded. We never burden a host.
6. Opt-out is honored. Anyone who requests removal has their content purged within 7 business days, with confirmation.

What we ingest, and what we don't

Sources are sorted into categories by their legal basis. Only Categories A, B, and C may be ingested. Category D may not.

Signals we respect

Our crawler identifies itself as OER-Oracle-Bot/1.0 (+https://oneearthrising.com/sage-policy) on every request. Before fetching any URL it reads the following signals. If any tell us to stay out, we stay out.

• robots.txt — standard exclusion directives. We respect Disallow rules targeting our user-agent or all crawlers.
• llms.txt — emerging standard for site-level AI crawler directives. Honored where present at the host root.
• tdm-reservation — EU-aligned text and data mining reservation, as HTTP header or HTML meta tag.
• noai / noimageai — meta tags and HTTP headers signalling AI-training opt-out.

Rate limits are also non-negotiable: no more than 1 request per 2 seconds per host, no parallel requests to the same host. We monitor emerging standards quarterly and commit to adopting new signals within 60 days of stabilization.

How we process content

What lands in S.A.G.E. is not a copy of what we fetched. Every approved source goes through structured extraction before any chunk is stored:

• Transformation through Gemini. Source text is processed by Google's Gemini through structured-output extraction. The embedded vector is computed against a Gemini-generated 3-sentence canonical summary — not the source.
• Evidence excerpt limit. The verbatim "evidence" field stores at most 200 characters per source, used solely for citation.
• Source URL preservation. Every chunk preserves its source URL (and where applicable, a deep-link timestamp) for clickable attribution.
• License tag preservation. Every chunk records its license category in the database. Without a tag, the chunk cannot be inserted.
• Raw content retention. The pre-extraction raw fetched content is discarded within 30 days. Only the structured outputs persist.

4.1 Time-to-live for Category B sources

Category B sources accessed through licensed APIs are subject to a 90-day time-to-live. Before TTL expires, the source is re-verified against the live API. If the source content has been deleted, materially altered, or made private since ingestion, the corresponding chunks are purged from S.A.G.E. immediately. This honors pass-through deletion requirements from API providers.

Category A and Category C sources are exempt from strict TTL deletion unless an explicit opt-out is received.

4.2 Factual correction

Even legally acquired content can become factually incorrect — through wiki vandalism prior to revert, or through stale information after a game patch. Operators and trusted graders can flag chunks for re-extraction or quarantine. Flagged chunks are excluded from retrieval pending review.

How to opt out

You can request removal of your content from S.A.G.E. at any time. There are three paths, depending on what fits you best:

Our service-level commitment for opt-out requests:

Opt-out applies to future ingestion as well: once a domain or creator handle is on our opt-out registry, it is excluded from all subsequent runs automatically.

How S.A.G.E. uses it

S.A.G.E. is configured to synthesize, abstract, and cite — never to reproduce. Specifically:

• Responses are paraphrased from retrieved chunks, not verbatim quoted.
• Quoted text from any single source must not exceed 15 words.
• Each response includes at least one source citation when retrieval was used.
• When retrieval is weak (similarity below threshold), S.A.G.E. discloses uncertainty rather than synthesizing speculatively.

Legacy content — reclassification complete

All content originally tagged legacy_pre_policy has completed CEO-led reclassification review. Every chunk currently in S.A.G.E. was confirmed to trace directly to a named Knowledge Keeper submission and has been reclassified into Category C (creator-licensed) — our preferred path per §2. No content required quarantine. The legacy_pre_policy tag no longer applies to anything in the corpus.

Live, current figures are always available on the Knowledge Source Transparency Report →

Transparency & EU compliance

The legal landscape for AI training data is unsettled and moves quickly. We adopt the more conservative interpretation when an obligation is ambiguous — the cost of over-disclosure is small, the cost of under-disclosure is large.

Knowledge Source Transparency Report

We publish a live, self-updating Knowledge Source Transparency Report aggregating, per game: license categories and chunk counts. The report covers all content ingested under this policy (legacy content joins once reclassification completes). View the live report →

We no longer crawl third-party platforms directly. All new knowledge enters S.A.G.E. through direct collaboration with Knowledge Keepers — creators who submit their own content and are paid per citation. Learn more or apply on the Knowledge Keepers overview page.

EU AI Act

The EU AI Act (Article 53) imposes transparency obligations on providers of General-Purpose AI (GPAI) models. Whether S.A.G.E. constitutes a GPAI model under the Act is not yet settled — a defensible position is that S.A.G.E. is a retrieval-augmented application built on top of Google's Gemini (the GPAI provider), not a GPAI model itself. We are pursuing qualified EU legal review and will revise this policy based on counsel's recommendations.

Updates to this policy

This is version 1.1, in force April 30, 2026. New in this version, marked above: machine-readable rights reservations beyond robots.txt; 90-day TTL for Category B; factual correction surface; the architectural firewall between retrieval and training; EU AI Act considerations; and signature/changelog discipline.

All v1.1 changes are forward-compatible with v1.0 — no v1.0 commitment is removed. Material changes are versioned, dated, and summarized at the top of this page. Previous versions remain accessible by request.