Pulse: simplifying document access for agents

Identity and question

ID: pulse-document-access-2026-09-28; revision 1.
Owner/lane: Pulse contribution to the Lab's project studies; implementation remains Pulse-owned.
Status: completed retrospective technical study. No comparative trial.
Question: can an agent acquire complete, identifiable Pulse documents through known HTTP URLs and read them locally without a hosted browsing reader, website rendering, GitHub retrieval, or live RPC?

The motivating suggestion was that simpler tools make communication with agents more stable. This case supports a narrower practice: keep required operations simple, make exchanged data explicit, and make failures observable. It does not establish a reliability gain across agents or providers.

Scope and method

This is infrastructure relevant to projects that give agents public knowledge. It does not evaluate Agent intention, artwork quality, pricing theory, or whether Pulse itself constitutes Agent Art. The Lab's distinctions between evidence classes and its retrospective method are reused.

The sequence is reconstructed after the refactor. Earlier retrieval problems motivated the changes; the study question and proposed lesson were not preregistered. A fresh read-only check supplements source inspection and session-reported history. There is no matched before/after cohort, controlled assignment, or measured success-rate difference.

The implementation snapshot is Pulse commit 32adfe39e1bf775f387ddab698780f256ed0cc81, tagged pulse-site-v1.0.2. The public index revision examined is 2026-09-27-r7. These identify source and document revisions, not an attestation that every deployed byte matches that Git commit. Intermediate failed reader attempts lack an exact retained runtime/build identity in this collection.

The operator authorized sharing the case. Actions here are source inspection, public document reads, local validation and PR publication. No private history retrieval, live agent experiment, pricing execution, RPC, wallet action or deployment is part of this study. No paid trial was commissioned; time, token use and infrastructure cost were not measured. Permission denials must use the host's approval process; tool substitution is not a permission bypass.

Evidence inventory

Evidence	Class and observer	Access	Establishes	Limits
Download format, prompt, generator and catalog at the pinned commit	Source inspection by the contributing agent	Public source	Explicit acquisition contract and build design	Design alone is not observed agent compliance.
Knowledge tests and standalone client	Source inspection; prior successful runs are session-reported	Public test implementation; historical raw logs not retained here	Checks cover exact source reconstruction, deterministic generation, missing RPC and saved-file replay	This contribution does not relabel inspected tests as freshly executed tests or semantic understanding.
Public index and ten document downloads, 2026-09-28	Direct read-only check by the contributing agent	Public endpoints; procedure below	Current acquisition and file/source verification on one client path	Mutable URLs; one environment, no comparative reliability estimate.
Earlier hosted-reader failures and public-access changes	Retrospective session report	Sanitized account below; raw traces unavailable here	Motivation and possible confounds	Original reader failure cause remains unconfirmed.
An answer addressed prompt validation after a floor-policy question	Retrospective session report	Sanitized account; private conversation not exported	Reported task-following failure despite relevant retrieval	Not independently reproducible here; no failure-rate estimate.

What changed

The client contract became:

Download the known index into an authorized workspace.
Select relevant entries and download their complete download.url files.
Check byte length and hashes when possible; parse each JSON envelope.
Read or search the complete content locally, in portions if helpful.
Answer from those sources, citing downloaded URLs and revisions; distinguish original-source provenance from evidence actually fetched.

The index is a small discovery map, not all website content in one JSON file. Each export contains one whole logical source. All ten current sources are text and have no required companion media. Outbound references are not automatically downloaded. A question that depends on one still requires deliberate retrieval.

Same-origin exports remove GitHub from the consumer's required retrieval path for those sources. Build-time generation still needs the pinned Git objects. The authored index in the repository is enriched with download descriptors during generation; it is not byte-identical to the served index.

The source hash identifies the original UTF-8 content. A separate download hash identifies the serialized envelope. Keeping JSON source inside a string avoids rewriting its integer literals or formatting. The checks cover Unicode, BOM, CRLF and large integer text. Hash agreement establishes byte identity relative to the index, not truth, freshness or independent authenticity.

Static documents are separate from the playground's runtime and read-only RPC gateway. The inspected tests exercise document access with missing RPC and a 503 status response. Historical chain records remain historical documents; downloading them today does not verify today's chain or application state.

Dated observations and reproduction

Before this study, reported during the Pulse refactor: a hosted browsing reader could not consistently retrieve the public sources, while ordinary HTTP retrieval later succeeded. Public-access security configuration also changed. Prompt wording, published content and runtime dependencies changed during the work. These are confounds: success cannot be attributed to simpler tools alone. The hosted reader's original failure cause remains unknown; a cache explanation was considered, not established.

2026-09-28: the contributing agent inspected the pinned implementation, then ran the existing standalone Python client against the production index and all ten exports. Both download verification and a separate saved-file replay exited successfully. Each reported ten complete downloads and matching source/file hashes. This was one acquisition pass and one replay, not repeated network trials or a new agent trial.

The check ran on macOS with Python 3.14.6 and its standard-library HTTP client. No model/runtime identity is inferred from the assistant's configured label.

The fetched index had revision 2026-09-27-r7 and serialized-byte SHA-256 18285e4476ca5032a4de05d1ee06bfef489759618ba5331c4c18d92a26aa8595. Verified export IDs were proposition, playground, core-api, integration, core-build, core-sepolia, core-abi, core-handoff, site-release, and projects, each at https://pulse.inshell.art/docs/documents/<id>.json. For example, the Core API download had source revision sha256:6faa430c09e64bbfa2ff46c381bf70a4754accac40124718fd8d5f85cb82c720. This record retains result summaries and identities, not an immutable archive of HTTP responses. Future public content can differ.

A reproducer with permitted network access and an authorized local workspace can obtain the client at the pinned revision, inspect it, and run:

python3 smoke-public-access.py https://pulse.inshell.art --save-dir downloads
python3 smoke-public-access.py --offline-dir downloads

The client uses Python's standard library and default User-Agent. It requests GET and HEAD for the index and exports, checks media types and nosniff, then verifies saved lengths, envelope/source hashes, revisions and identities. It extracts text, checks selected API terms and parses the ABI. It requests no website runtime, GitHub source, status endpoint or RPC during document use. Acquiring the checker itself is a separate setup step, not a prerequisite for agents using their own permitted HTTP tools.

During local analysis the client replaces Python socket/DNS/urllib functions with rejecting functions. The separate replay starts with that guard enabled. This is an in-process check, not OS-level network isolation. The output's word “analyzed” means parsing, integrity checks, extraction and limited search; it does not measure comprehension or answer correctness.

Outcome, counterevidence and transfer limits

The architecture supplies complete inspectable inputs with fewer required runtime dependencies. It makes failed retrieval, missing storage and hash mismatches explicit. This is a demonstrated path and a reasoned design benefit, not a measured reduction in overall agent failure probability.

A reported conversation mistake is particularly relevant: after fetching the appropriate documents for a question about the proposition and Core V1's floor policy, the assistant answered about prompt validation instead. Successful transport does not guarantee attention to the user's question or a correct answer. Retrieval, interpretation and task completion require separate evidence.

Other limits remain:

The initial server, network or tool may be unavailable or denied. A different permitted HTTP client may help unsupported retrieval; it cannot guarantee it.
This particular prompt requires local storage. Agents without an authorized filesystem cannot complete it. Simpler operations do not mean fewer required capabilities for every host.
Rich media and linked sources need an explicit dependency policy. The current text-only exports are not a complete offline copy of the website or repository.
A valid saved document can be stale or factually wrong. Live questions require separately authorized live evidence.
More metadata adds maintenance: source hashes, revisions, export coverage and tests must stay aligned. No maintenance or token-cost savings were measured.

Evaluate the Lab method

The evidence inventory exposed where claims depended on missing reader traces and separated public source inspection from fresh checks and session reports. The outcome distinction preserved the task-following counterexample instead of treating a download as full success. These are concrete editorial benefits.

The retrospective record cannot recover omitted before/after measurements or the original reader's exact runtime. It did not measure the method's time cost or causal effect. No general framework change or shared runner is justified. The small proposed addition is P-07 in the findings register.

Handoff

This record completes the bounded technical case; it leaves comparative reliability and original failure attribution unresolved. Lab maintainers review the proposed practice. Pulse retains implementation and release ownership. Review the practice if agents lack local storage, documents require media or live state, metadata causes drift, or a declared comparison contradicts the expected benefit. No such trial is initiated by this contribution.

This public derivative excludes private transcripts, operator configuration and raw historical logs. It adds a new study without modifying source originals or the historical standalone edition's provenance identities.