CAPTURE WAVE1 CONTRACT

Repository: thelaplage/counterpedia-acquisition
Bound main SHA: eac638def41922fdaacfa04d1b0302de9d2ed5e7
Target branch: feat/demo-corpus-capture-wave1-v0-1
P0 manifest entries: 112
Subjects: 12
State: acquisition / pre-adjudication only

Purpose

Convert the research dossiers' P0 web-reconnaissance sources into durable acquisition evidence:

requested locator → HTTP attempt → exact captured bytes → sha256 content address → CaptureReceipt → attempt/run log → deterministic passage anchors → research edition-lineage observation

This wave does not decide what any source proves.

Constitutional boundary

The existing counterpedia-acquisition boundary remains binding:

• agents may discover, capture, observe, and propose;

• source identity remains acquisition-local and URL-derived under the existing code;

• exact bytes remain content addressed;

• each capture attempt retains its own receipt/outcome;

• redirects do not rewrite source identity;

• failures do not create fake receipts;

• there is no admission, approval, publication, or standing authority here.

This wave MUST NOT:

• redesign SourceRecord, SourceEdition, Artifact identity, or Counterpedia namespace rules;

• mint canonical cp:* IDs;

• write ClaimGraph/admission state;

• assign epistemic confidence;

• generate article prose;

• treat source class/priority as truth weight;

• silently replace origin bytes with a mirror;

• infer equality of two artifacts from title/URL similarity.

P0 manifest semantics

The manifest is research/run input, not canonical source identity.

dossier_source_id such as TIT-S02 is an inventory key only.

capture_mode values:

• FETCH_DIRECT — attempt the requested locator directly.

• DISCOVERY_PAGE / DISCOVERY_INDEX — capture the page itself; child artifacts remain separate.

• MULTI_ARTIFACT_SOURCE — one research source ID requires multiple exact child artifacts; runner records the landing capture and a child-resolution queue, but MUST NOT recursively crawl without explicit bounded resolver rules.

• DISCOVERY_REQUIRED — the dossier identifies an exact source object but current direct origin locator is unresolved. Do not invent one.

• ORIGIN_REQUIRED — a mirror/recon copy exists, but governed capture must locate or explicitly distinguish the origin.

• EXPAND_LINKED_PACKAGE — capture landing page, then operator-approved exact linked documents individually.

• SAME_ARTIFACT_SECTION — no second fetch; passage anchor is against another already captured artifact.

• BLOCKED_PRIMARY_BYTES — STOP for that source until acceptable institutional/court/origin bytes are found.

Run output contract

A run directory must be self-contained and immutable after closure:

run.json
program.json
manifest.snapshot.json
attempts.jsonl
receipts.jsonl
lineage_hints.jsonl
anchors.jsonl
failures.jsonl
objects/

run.json

Must include:

• run UUID

• start/end UTC

• git repo + exact commit

• manifest sha256

• Python version

• acquisition package version/commit

• extractor names + versions

• configured max bytes / timeout

• final counts by disposition

exact bytes

Use the existing FilesystemObjectStore.

Never write partial response bodies.

Never trust Content-Length as byte count.

passage anchors

Anchors are derived extraction artifacts, not claims.

Every anchor must bind to:

• exact artifact sha256;

• capture ID;

• source inventory key;

• extractor ID/version;

• locator type (html_text_span, pdf_page_text, pdf_page_figure, etc.);

• page/index/span;

• extracted passage text hash;

• extraction timestamp UTC.

PDF extraction:

• deterministic text extraction only;

• no OCR in v0.1;

• if page text is empty/image-only, emit ANCHOR_UNAVAILABLE_IMAGE_ONLY, then permit figure/page anchors without inventing OCR text.

HTML extraction:

• anchor against exact HTTP bytes, not browser-rendered text.

• normalized visible-text extraction may be emitted only with extractor/version metadata.

edition lineage

The acquisition lane may record research lineage hints only:

• corrects

• amends

• updated_version_of

• same_artifact_section

• candidate_same_document

A lineage hint becomes useful only after the exact artifact digests on both sides are known.

It MUST NOT mint or assert canonical SourceEdition identity.

STOP conditions

For an individual entry:

• no acceptable origin locator where origin is required;

• HTTP non-200 under current fetch contract;

• byte limit exceeded;

• timeout;

• malformed transport metadata;

• redirect to a non-http(s) target;

• content cannot be extracted without OCR where OCR is not allowed;

• mirror found but origin equivalence not byte-proven.

Record the disposition and continue other entries.

For the entire run:

• manifest validation failure;

• duplicate capture_key;

• non-http(s) fetch locator;

• repo main mismatch during implementation preflight;

• any code path that creates admission/standing/publication state;

• any attempt to fabricate a receipt for a failed capture.

Acceptance target

The wave is successful even if some sources remain blocked.

Success means:

1. every one of the 112 P0 manifest entries has a deterministic final disposition;

2. every successful fetch has exact bytes + verified sha256 + real CaptureReceipt;

3. every failed/blocked item remains visible;

4. no mirror is silently promoted;

5. passage anchors exist where deterministic extraction permits them;

6. image-only pages are page/figure anchored or explicitly marked unavailable;

7. edition/correction relationships remain typed research hints;

8. no adjudication occurred.