Answer Service

Agents asking what I have argued should not have to infer it. The Answer Service returns the claim, the exact passage it rests on, the sha256 of the source document, and the conditions under which the claim was made.

Two surfaces, deliberately separated. The agentic claim graph is a static projection of the corpus into agent capability structure. The Answer Service is the retrieval interface over it. The graph is published as content addressed JSON-LD. The service runs off site, read only, no authentication, no keys.

Read only by construction. The service creates no claims and affirms no claims. Retrieval is not authorship.

Endpoints

Home	kaal-answer-service.wulf577462.chatgpt.site
Answer API	/api/answer
MCP JSON-RPC	/api/rpc
A2A endpoint	/a2a
Agent card	/.well-known/agent-card.json
ARD catalog	/.well-known/ai-catalog.json
Status	/api/status

Two MCP tools are exposed. answer_question takes a question and an optional limit between 1 and 20. graph_status returns graph and service scope.

Call it

curl -s -X POST https://kaal-answer-service.wulf577462.chatgpt.site/api/answer \
  -H 'content-type: application/json' \
  -d '{"question":"What has Wulf A. Kaal written about verified reputation?","limit":5}'

What comes back

Every response carries a contract version, a coverage status, a retrieval summary, an authority block naming the graph and its source authorities, and an explicit limitations array. Each claim carries:

claim_id and canonical_url, resolving into the published claim layer
Attribution: author, ORCID, attribution status, and class
The exact source passage, with passage_sha256 and source_sha256
Source title, full citation, source hash, paper graph identifier
Scope conditions, the conditions the claim was argued under
Topics, capabilities, confidence, and verification status

Scope conditions are not decoration. A claim about reputation voting inside a DAO of DAOs is not a claim about reputation systems generally, and the response says so.

The graph behind it

Graph identifier kaal:agentic-claim-graph:pilot-100:v1, state public_pilot, built 31 July 2026.

Claims projected	100
Canonical corpus claims	5,033
Nodes	248: 100 Claim, 100 SourcePassage, 37 ScholarlyArticle, 10 AgentCapability, 1 Person
Edges	800 typed edges across capability, source, support, and authorship relations
Selection	Deterministic agentic capability relevance v1, capped at ten claims per paper
Attribution class	source derived only

index.json, the entry point
graph.jsonld, the graph itself
manifest.json, digests and counts
vocabulary.json, the capability terms

Capability vocabulary

Ten agent facing capability classes carry the projection. Claims attach to capabilities, not to marketing categories.

agentic-commerce-and-settlementcoordination-and-incentivesdecentralized-infrastructuregovernance-and-adjudicationlearning-and-feedbacklegal-and-institutional-governanceprovenance-and-knowledgereputation-and-identitysafety-and-accountabilitytrust-and-verification

Invariants

These hold for the service and the graph, and they are published as machine readable assertions in index.json:

Read only. The service creates no claims and affirms no claims.
Only source derived claims directly attributable to Wulf A. Kaal are returned.
A Colloquium validation pool may settle a procedural claim about what a review concluded. That claim is authored by the Colloquium validation process, not by me.
Any substantive derivative stays unpublished and unattributed until exact hash bound affirmation. The queue is currently empty by design.

Procedural claim contract: procedural-claim-v1.schema.json. Current state: procedural-claims.json and derived-candidates.json.

Verify it yourself

The graph is content addressed. Recompute the digest and compare it against the manifest:

curl -s https://wulfkaal.github.io/agentic-claim-graph/v1/graph.jsonld | shasum -a 256

Expect 7914a0e4ce2df76d0d4314255720cbeaeb483bdabc7a28d3618f7bfbd140ac0b, the value recorded in manifest.json. Any claim returned by the service can be checked the same way, passage against source, using verify.py.

Limits

This is a pilot of 100 claims, not the full 5,033 claim corpus. Absence from a response is not evidence of absence from the corpus.
Exact passages are preserved from the canonical claim records. Per claim PDF matching was not recomputed by this pilot.
Quotes carry exact text without page locators.
The service is a retrieval surface. It reports what was argued and under what conditions. It does not adjudicate whether the argument is correct.

Elsewhere

Agents, what this domain offers a machine reader
LLM, the flat index of every machine surface
Claim layer, all 5,033 claims
llms.txt, the machine index
Colloquium, the validation layer

Wulf A. Kaal · ORCID 0009-0008-7840-1847 · SSRN