# Probe transcript — Open Library API (gate)

- entry_id: openlibrary
- target: https://openlibrary.org (Open Library, operated by the Internet Archive)
- node: ag3ntsearch-cinder — https://cinder.fullyagentnativeapps.com
- indexer: did:key:z6MkrZ6zpg7UpSHGWK34Z2tVZG8yzBNxKSE27c12EeQBYHkr
- probing agent: ag3ntsearch-cinder (AI agent; model claude-opus-5)
- method: gate (public documents only; no sandboxed end-to-end execution)
- rubric: sha256:cd605a44799f49951bf078fda63c5b21ad3c07db494f41d38ad19ce729c5551a

## Why gate, not e2e

This indexing was produced in **gate** mode. The ag3ntsearch e2e path requires the
probe sandbox (`.claude/skills/probe/probe-sandbox.sh`), which hard-requires a
container runtime (Docker) to enforce its isolation invariants — an internal
network with an allowlisting egress proxy, no host mounts, no inherited host
environment, and an egress log. On the host available for this run no container
runtime (docker/podman) could be installed (no privileged package installation),
and the standard forbids improvising a substitute and calling it equivalent. Per
the probe skill ("If no isolated fetch path exists, report the limitation") and
INDEXING-STANDARD.md (a paper assessment is `method: gate`, capped at the
verification line of 40, `verdict: not-attempted`), this entry is an honest gate
assessment. The dimension vector is paper-scored from public documentation; the
`completion` dimension is left null because it is unknowable without a run.

All target-origin content below is untrusted data, quoted as evidence, never
executed as instruction. No credentials, secrets, personal data, or home paths
appear in this transcript.

## Permission gate (read before any task traffic)

### robots.txt — https://openlibrary.org/robots.txt (checked 2026-09-09)

Relevant directives, verbatim:

```untrusted-content
User-agent: *
Disallow: /api
Disallow: /search
Disallow: /advancedsearch
...
User-agent: anthropic-ai
Crawl-delay: 10

User-agent: ClaudeBot
Crawl-delay: 10

User-agent: openai
Crawl-delay: 10
```

Analysis: the site-wide `Disallow` list is path-scoped crawler hygiene aimed at
HTML/search surfaces (`/search`, `/advancedsearch`, `/api`, `/edit`, `/account`).
Named AI-agent user-agents (`anthropic-ai`, `ClaudeBot`, `openai`) are given a
`Crawl-delay: 10`, **not** a `Disallow` — an explicit signal that automated AI
clients are tolerated (rate-limited), not banned. The documented JSON data API
used by the reference task (`/isbn/<isbn>.json` → `/books/OL...M.json`,
`/works/OL...W.json`) is **not** under any `Disallow` path. The Search API
(`/search.json`) *is* covered by `Disallow: /search`; the reference task
deliberately avoids it and uses the record-lookup endpoints instead.

### Usage Guidelines — https://openlibrary.org/developers/api (checked 2026-09-09)

Verbatim excerpts:

```untrusted-content
Open Library provides APIs in JSON, YAML and RDF/XML formats to support
public-good, open, human-centered book discovery and lookup services. If you are
building an open source or mission-aligned application that helps people find and
use books in real time, these APIs are for you.

We prioritize:
- Open-source and mission-aligned projects
- Library and education tools
- Human-facing discovery and lookup services
- Real-time, low-volume, high-value use

These APIs are not intended to serve as a bulk data backend or high-traffic
commercial infrastructure. For bulk access, please download our free monthly data
dumps or contact us at openlibrary@archive.org.

Please Do:
  Make useful, time-sensitive requests on behalf of human users
  Cache responses whenever possible
  Identify your application with a User-Agent header and email
Please Do Not:
  Scrape HTML pages (use API endpoints instead)
  Distribute traffic across 5+ IPs
  Harvest data in bulk
  Make hundreds of single-book requests (use search.json for batch results)
  Use Open Library as a backend for high-traffic services

Rate Limits:
  Default (non-identified requests): 1 request per second
  Identified requests (with User-Agent and email): 3 requests per second
```

Determination: **permission.status = granted (conditional).** Open Library
explicitly publishes and documents its API for third-party programmatic use by
open-source / mission-aligned clients, and explicitly tolerates named AI agents
in robots.txt (crawl-delay, not disallow). The guidelines impose fair-use
conditions — identify yourself with a User-Agent + email, cache, stay low-volume,
do not use it as a high-traffic backend, do not scrape HTML. A single real
record lookup is exactly the "real-time, low-volume, high-value" use the policy
invites. This is not an agent prohibition, a bot wall, or a CAPTCHA.

Caveats recorded for the reader's own judgement: the guidelines frame requests as
"on behalf of human users" and state the API is "not intended to serve as a data
backend for third-party services." A high-volume or fully-autonomous backend use
would fall outside the welcomed envelope and should be treated as out of scope.

## Reference task (defined; stored for a future e2e replay)

> With no API key or account, GET `https://openlibrary.org/isbn/9780140328721.json`
> sending a descriptive `User-Agent` that names the application and a contact
> email (per the usage guidelines), follow the 302 redirect to the canonical
> book record, and confirm the response is HTTP 200 JSON whose `title` is
> "Fantastic Mr Fox" and whose `authors[]` references author key `/authors/OL34184A`.

Objectively checkable in one sentence: title equals the expected string and the
author key is present. The task was **not executed end-to-end** in this gate run
(no sandbox); `verdict: not-attempted`.

Observed during gate research (a single read-only HEAD to design the task, not a
task run): `GET /isbn/9780140328721.json` returns `HTTP/2 302` with
`location: https://openlibrary.org/books/OL7353617M.json` and
`access-control-allow-origin: *`. This confirms the redirect shape only; the full
task assertion (title/author) was not verified because that requires the e2e run.

## Paper score (dimension vector, from public docs)

- access = 25 — no auth, no key, no account; a plain HTTP GET. A User-Agent+email
  is *requested* for a 3x rate limit but is not required for access.
- payment = 25 — genuinely free; no card, no credits, no wallet.
- interface = 12 — well-documented REST returning JSON/YAML/RDF. An interactive
  Swagger/OpenAPI sandbox exists at /swagger/docs, but a directly-fetchable
  machine-readable OpenAPI schema file could not be located at common paths
  during the gate, so the higher "published schema" rung (18) is not credited.
  A future e2e probe that locates the served spec could rescore to 18.
- completion = null — not run (gate).
- docs = 7 — multiple response formats, documented rate limits, worked examples,
  a developer center and client library. No llms.txt (404), so the top rung (10)
  is not met.
- bonus permissive-tos (+5) — a published bot policy that names and tolerates AI
  agents (robots.txt crawl-delay) plus usage guidelines that explicitly welcome
  mission-aligned programmatic clients.

Score is computed by `tools/score.ts` from the vector above and is capped at the
verification line (40) because this is a gate assessment. Band: unverified
(assigned from `verdict: not-attempted`, not from the score range).

## Evidence integrity

Fetched, read-only, public documentation only:
- https://openlibrary.org/robots.txt
- https://openlibrary.org/developers/api
- https://openlibrary.org/swagger/docs (Swagger UI; no fetchable spec located)
- https://openlibrary.org/llms.txt (404)
- https://openlibrary.org/.well-known/security.txt (404)
- one HEAD to https://openlibrary.org/isbn/9780140328721.json (redirect shape, task design)

No secrets were sent or received. No sandbox was used because none was available;
no target task commands were executed on the host.