# Pith

> Reviews of arXiv papers for AI agents: verdicts, claims, flaws, compiled records, citation graphs.

- name: science.pith/pith
- version: 1.1.0
- connection class: R0 (autonomous)
- trust: 83/100 flags: no-repository
- quality: 90/100 (strong)
- owner: not claimed
- descriptor: https://api.protogrid.dev/v1/servers/science.pith%2Fpith
- tools: https://api.protogrid.dev/v1/servers/science.pith%2Fpith/tools

**An agent can connect right now, no human step.** Remote endpoint, no authentication, reachable on the last probe.

## Connection (mcpServers)

```json
{
  "mcpServers": {
    "pith": {
      "type": "http",
      "url": "https://pith.science/mcp"
    }
  }
}
```


## Trust

- hygiene: 80
- liveness: 100
- freshness: 100
- provenance: 55
- drivers: +dns-namespace +website +reachable +uptime-100% +updated-27d-ago -no-repository
- Derived from observable signals (official registry feed and our own credential-free probes); no code audit performed.

## Quality 90/100 (strong)

### protocol: 88 (weight 25)

- pass `protocol.modern`: supports 2026-07-28 (server/discover)
- pass `protocol.stateless`: answers without a session
- pass `protocol.transport`: streamable HTTP
- warn `protocol.list_ttl`: tools/list declares ttlMs 0, so clients re-fetch it on every use

### authorization: n/a (weight 15)

- n/a `auth.prm`, `auth.as_metadata`, `auth.cimd`: no remote uses OAuth
- n/a `auth.secret_in_url`: no templated URL

### tool hygiene: 100 (weight 30)

- pass `tools.descriptions`: every tool has a description
- pass `tools.description_length`: descriptions are concise
- pass `tools.schemas`: every tool has an object input schema
- pass `tools.annotations`: 24 of 24 tools declare readOnlyHint or destructiveHint
- pass `tools.directory_hints`: every tool declares readOnlyHint, destructiveHint, idempotentHint and openWorldHint
- pass `tools.token_cost`: about 4,569 tokens to load every tool
- pass `tools.api_dump`: tools are not a one-to-one API dump

### stability: 75 (weight 15)

- warn `stability.changes`: 13 tool changes in 30 days
- pass `stability.rug_pull`: no tool changed its meaning under the same name

### dependencies: n/a (weight 15)

- n/a `deps.known_vulns`, `deps.mcp_sdk_version`, `deps.resolvable`: no npm or PyPI package

- tools: 24, about 4,569 tokens to load them all
- tool set hash: a5720b9bb9f585776309c1a95a6f10ec
- tools last changed: 2026-10-09T21:35:48.703Z
- badge: [![protogrid quality](https://protogrid.dev/badge/science.pith/pith.svg)](https://protogrid.dev/servers/science.pith/pith)
- Checks run on what our credential-free, read-only probes observe; tools are never called and no code audit is performed.

### Tool changes

- 2026-10-09 `get_review` changed (outputSchema)
- 2026-10-03 `paper_connections` changed (description)
  - before: What other arXiv papers do with this paper, and what it does with the arXiv papers it cites: builds on, uses the method or tool, uses the data, compares with, cites as support or background, or contests. Each group counts distinct papers and quotes the citing authors' own sentences. Labels come from each citing paper's LaTeX source typed by Pith's classifier and are shown only at intent probability 0.8 or more (contests at 0.7 or more); the rest are counted as unlabelled. Measured agreement of every label with independent model families: https://pith.science/connected-record. Use this for how a result has been taken up, challenged or built on across arXiv.
  - after: What other arXiv papers do with this paper, and what it does with the arXiv papers it cites: builds on, uses the method or tool, uses the data, compares with, cites as support or background, or contests. Each group counts distinct papers and quotes the citing authors' own sentences. Labels come from each citing paper's LaTeX source typed by Pith's classifier and are shown only at intent probability 0.8 or more (support at 0.95, contests at 0.7); the rest are counted as unlabelled. Measured agreement of every label with independent model families: https://pith.science/connected-record. Use this for how a result has been taken up, challenged or built on across arXiv.
- 2026-10-02 `ask_field` added
  - Read the papers closest to a question and return, for each, whether its own abstract sentences answer yes, no or not addressed, with the deciding sentence quoted. Call when a user asks whether something holds across the literature. Candidates are chosen by meaning, so a paper that answers in different words can be missed. Needs a personal key; counts against a daily allowance.
- 2026-10-02 `check_my_draft` added
  - Resolve each reference in a draft, flag retracted works, and for arXiv papers show how later papers cite them: as support, and whether any contest them, with the contesting sentence quoted. Call before a submission or when a user wants their reference list checked. Needs a personal key; counts against a daily allowance.
- 2026-10-02 `my_reviews` added
  - Pith's machine review verdict and the count of published signed human reviews for each paper claimed on the account behind your key. Needs a personal key.
- 2026-10-02 `my_citations` added
  - For the papers claimed on the account behind your key: how many papers cite each, whether they cite it as support, extension or disagreement, and the citing sentence quoted. Call when the user asks who is building on or disputing their work. Needs a personal key.
- 2026-09-26 `find_passages` changed (description, outputSchema)
  - before: Author sentences Jev labeled with one role, across every stored record, each with its probability and paper. Abstract-tier records cover limitation, result and main_claim; full-text records add condition, citation and measurement.
  - after: Author sentences the classifier labeled with one role, across every stored record, each with its probability and paper. Abstract-tier records cover limitation, result and main_claim; full-text records add condition, citation and measurement.
- 2026-09-26 `verify_citation` added
  - Whether the cited paper's own sentences support what a sentence attributes to it: supports, partly supports (a condition is missing), contradicts, or not addressed, with the single best sentence from the paper quoted. Call this before citing a paper for a specific claim, or to check a citation in a draft or an answer. It reads the paper's abstract, and its full text where Pith holds it; the examined field says which. On real citations its verdict agreed with two blind model families 139 times in 160, most differences being partly supported versus not addressed (https://pith.science/connected-record).
- 2026-09-26 `check_claim` added
  - Which arXiv papers support a scientific statement, which contradict it, and which study it without deciding, each with the paper's own main claim quoted and the classifier's probability. Call this before asserting a scientific result, when a user asks whether something is established or disputed, or when two sources disagree. Candidates are papers whose title and abstract match the statement's words, plus papers Pith's connected record links to them as stating the same or a conflicting result; a paper that states the result in other words can be missed, so an empty answer is not evidence of absence. Reported supporters and contradictions agreed with two blind model families 110 times in 112 (https://pith.science/connected-record).
- 2026-09-26 `paper_connections` changed (description)
  - before: What other arXiv papers do with this paper, and what it does with the arXiv papers it cites: builds on, uses the method or tool, uses the data, compares with, cites as support or background, or contests. Each group counts distinct papers and quotes the citing authors' own sentences. Labels come from each citing paper's LaTeX source typed by Jev and are shown only at intent probability 0.8 or more (contests at 0.7 or more); the rest are counted as unlabelled. Measured agreement of every label with independent model families: https://pith.science/connected-record. Use this for how a result has been taken up, challenged or built on across arXiv.
  - after: What other arXiv papers do with this paper, and what it does with the arXiv papers it cites: builds on, uses the method or tool, uses the data, compares with, cites as support or background, or contests. Each group counts distinct papers and quotes the citing authors' own sentences. Labels come from each citing paper's LaTeX source typed by Pith's classifier and are shown only at intent probability 0.8 or more (contests at 0.7 or more); the rest are counted as unlabelled. Measured agreement of every label with independent model families: https://pith.science/connected-record. Use this for how a result has been taken up, challenged or built on across arXiv.

Full history: https://api.protogrid.dev/v1/servers/science.pith%2Fpith/changes

## Remotes

- https://pith.science/mcp (streamable-http, auth none, reachable true, uptime30d 1)

## Packages

_none_

## Tools (24)

- `analyze_paper_claims`: The structured argument of one specific paper: what it actually
    asserts, the premises it assumes, the machinery connecting those
    premises to the claim, and the falsifier, as short labeled statements
    in argument order (premise, machinery, claim, falsifier). Use this for
    requests about a paper's claims, premises, assumptions, reasoning,
    logical dependencies, logical structure, argument structure, or
    conclusion. Each entry includes its text, role, type, and status.
- `ask_field`: Read the papers closest to a question and return, for each, whether its
    own abstract sentences answer yes, no or not addressed, with the deciding
    sentence quoted. Call when a user asks whether something holds across the
    literature. Candidates are chosen by meaning, so a paper that answers in
    different words can be missed. Needs a personal key; counts against a daily
    allowance.
- `check_claim`: Which arXiv papers support a scientific statement, which contradict
    it, and which study it without deciding, each with the paper's own main
    claim quoted and the classifier's probability. Call this before asserting a
    scientific result, when a user asks whether something is established or
    disputed, or when two sources disagree. Candidates are papers whose title
    and abstract match the statement's words, plus papers Pith's connected
    record links to them as stating the same or a conflicting result; a paper
    that states the result in other words can be missed, so an empty answer
    is not evidence of absence. Reported supporters and contradictions agreed
    with two blind model families 110 times in 112
    (https://pith.science/connected-record).
- `check_my_draft`: Resolve each reference in a draft, flag retracted works, and for arXiv
    papers show how later papers cite them: as support, and whether any
    contest them, with the contesting sentence quoted. Call before a
    submission or when a user wants their reference list checked. Needs a
    personal key; counts against a daily allowance.
- `citation_neighbors`: One hop from a paper: every reference entry with its resolved work
    identity, the papers that cite it, and, where a record exists, the
    citing sentences with their labeled intent. An edge means the paper
    names the other work, not that the work supports the citing claim.
- `find_passages`: Author sentences the classifier labeled with one role, across every stored
    record, each with its probability and paper. Abstract-tier records
    cover limitation, result and main_claim; full-text records add
    condition, citation and measurement.
- `get_claims`: Alias of analyze_paper_claims: the same ledger of what the paper
    asserts, its premises, machinery, and falsifier, in argument order.
- `get_falsifier`: Alias of get_paper_falsifier: the same falsifier and watcher count.
- `get_paper_falsifier`: The falsifier for one specific paper's core claim: the concrete
    observation, experiment, or computation that would settle or disprove
    it, plus how many people are watching for that result. Use this when the
    user asks what would falsify the paper, prove it wrong, test its central
    claim, settle the claim, or show that the conclusion fails.
- `get_record`: The compiled record of one paper: the authors' own sentences, arranged
    with typed labels. It carries the main claim in the authors' words, each
    abstract sentence's role, results with the qualifications the authors
    state beside them, the conditions of the study, measurements with where
    they appear, the arXiv listing abstract compared with the manuscript's
    abstract, resources including what is not stated, the authors' own
    limitations, how the paper uses its references, and what was and was not
    examined. No sentence in it was written by a model; every label carries
    its probability. Use this for what a paper states and under what
    conditions, including papers Pith has not reviewed, where get_review
    reports no review. Accepts an arXiv id or URL, DOI, or Pith Number.
- `get_review`: Pith's review of one specific paper, in one call: the verdict,
    confidence, novelty and correctness scores, one-line summary, weakest
    assumption, and the paper's argument as short labeled statements
    (premise, machinery, core claim, falsifier) already extracted, so the
    paper's premises and logical structure can be stated without reading
    the PDF. Use this for any question about a specific paper: what it
    asserts, what it assumes, how the argument runs, how sound it is, or
    what Pith concluded. Accepts an arXiv id or URL, DOI, or Pith Number.
- `get_reviewer_record`: A reviewer's published signed reviews and, when present, their
    calibration record (Brier score, open predictions, positivity skew).
    Call this when the user asks about a named Pith reviewer or how
    trustworthy a reviewer's track record is.
- `get_signed_reviews`: Published human signed reviews of one paper: reviewer identity, axis
    deltas against the machine review, claim verdicts, defects, predictions,
    and who paid. Call this when the user asks what human experts said about
    a paper, on top of the machine review.
- `latest_reviews`: The papers Pith reviewed most recently, newest first. Call this when
    the user asks what is new, notable, or recently reviewed in a field or
    overall. days caps how far back to look (max 90).
- `my_citations`: For the papers claimed on the account behind your key: how many
    papers cite each, whether they cite it as support, extension or
    disagreement, and the citing sentence quoted. Call when the user asks who
    is building on or disputing their work. Needs a personal key.
- `my_reviews`: Pith's machine review verdict and the count of published signed human
    reviews for each paper claimed on the account behind your key. Needs a
    personal key.
- `paper_bridge`: Whether either paper cites the other, the resolved works both cite
    (on or off arXiv), and the papers that cite both.
- `paper_connections`: What other arXiv papers do with this paper, and what it does with the
    arXiv papers it cites: builds on, uses the method or tool, uses the data,
    compares with, cites as support or background, or contests. Each group
    counts distinct papers and quotes the citing authors' own sentences.
    Labels come from each citing paper's LaTeX source typed by Pith's classifier and are
    shown only at intent probability 0.8 or more (support at 0.95, contests at 0.7);
    the rest are counted as unlabelled. Measured agreement of every label with
    independent model families: https://pith.science/connected-record. Use this
    for how a result has been taken up, challenged or built on across arXiv.
- `paper_packet`: One paper's compiled record, small enough to hold beside others: the
    main claim and contribution in the authors' words, each abstract
    sentence with its role and probability, paper-level labels, and, for
    full-text records, conditions, measurements, results and limitations,
    with totals so a capped list is not mistaken for the whole paper.
- `record_coverage`: Sources read and levels run for one paper's record, or the number of
    stored records by tier. Call before treating an empty result as a
    finding: an empty list means nothing in the examined scope matched.
- `resolve_paper`: Turn a pasted arXiv id, arXiv URL, DOI, or Pith Number into the
    arXiv id the other tools key on. ``resolved`` is false when the input
    did not become an arXiv id.
- `search_claims`: Full-text search over every scientific claim Pith has extracted,
    across all reviewed papers. Call this when the user asks who claims
    something, whether a claim appears in the recent literature, or what
    evidence bears on a statement. Returns each matching claim with its
    paper, verdict, and review link.
- `search_papers`: Search Pith-reviewed papers by title and abstract. Call this when the
    user names a research topic, phenomenon, method, or paper title without
    an identifier, or asks whether Pith has reviewed anything on a subject.
    Each result carries the paper's verdict, summary, and review link.
- `verify_citation`: Whether the cited paper's own sentences support what a sentence
    attributes to it: supports, partly supports (a condition is missing),
    contradicts, or not addressed, with the single best sentence from the
    paper quoted. Call this before citing a paper for a specific claim, or to
    check a citation in a draft or an answer. It reads the paper's abstract,
    and its full text where Pith holds it; the examined field says which. On
    real citations its verdict agreed with two blind model families 139 times
    in 160, most differences being partly supported versus not addressed
    (https://pith.science/connected-record).