Skip to main content

1. Purpose and status

This report sets out how the authors checked each statement in the indexer-hardening guide (indexer-hardening.mdx and its checklist). For every claim it gives the kind of evidence behind it, where that evidence is in this repository or upstream, and what is still open. All numbers below are generated from the repository’s files by guide/review/build.py, and make verify-guide fails if this report no longer matches them. This is the authors’ own check. Nobody independent has reviewed the guide. No stellar-core or platform engineer has reviewed it yet, and nothing here should be read as saying one has. Of the guide’s 462 claims, 0 are confirmed or corrected by a reviewer, 0 are disputed and 462 are unreviewed. The guide is written for developers.stellar.org. It will not be published there until a stellar-core or platform engineer has reviewed it: every fact claim confirmed or corrected and none disputed. This report is meant to make that review short: it tells the reviewer what each claim rests on and which claims need an expert’s judgement most.

2. Method

2.1 What a claim is

Each normative sentence of the guide has a claim tag ({/* claim: C-nnnn */}) right after it. Each tag has one record in guide/claims.yaml: the sentence, whether it is a fact or a recommendation, its sources, its verification and its reviewer_status. The field table also has records, one for each line’s verdict, each core check and each “you may / must not assume” sentence. scripts/py/check_guide_claims.py (in make repo-checks) fails when the page and the records differ, or when a sentence with “must”, “never”, “should” or “do not” has no tag. guide/REVIEW.md describes the record format. The records hold 462 claims (407 facts, 55 recommendations) and 1093 sources: 691 code lines, 241 corpus entries, 130 probes, 12 CAP and 15 SEP sections, and 4 links to upstream issues that the report leaves out.

2.2 The evidence classes

claims.yaml stores two verification values: executed (190 claims) and code_reference (272). This report splits executed by what the record cites, which gives the classes used in every table below. The rule is in guide/review/verification.py. Corpus entries: how the bytes were produced and checked. The claims cite 70 corpus IDs. An ID that names a group, such as a set of SEP-23 vectors, stands for each of its members, so the IDs cover 125 entries of corpus/index.json (54 ledger, 43 strkey, 20 api, 8 xdr) with 133 variants: 94 local-core, 29 history, 8 synthetic, 2 core-golden. The checks on each source kind are in docs/manifest.md sections 4.4 and 11.
  • All entries. xdrcheck-corpus verify decodes each blob with explicit limits and re-encodes it, and the bytes must match exactly. It also checks the hashes: each blob’s name is its SHA-256, and each decoded sidecar matches its stated hash and a fresh render. Every path and value an entry’s expect names must resolve and be equal. The ledger header hash and the transaction hashes are recomputed under the stated passphrase. The participants, state changes and movements an entry expects must match the reference extractor’s output.
  • local-core variants come from the pinned stellar-core release, run on a local network with a scenario file. Of the cited variants: 94 normalised-identical. Normalised-identical means two runs give the same ledger header hashes and the same bytes once stellar-core’s own golden-comparison normalisation is applied. The raw meta bytes are not byte-reproducible (see manifest section 4.4). xdrcheck-corpus gen --check regenerates a run and compares it.
  • Golden agreement. Of the 24 cited local-core entries, 2 match a transaction in stellar-core’s own golden ledger set, and 6 more have a golden operation of the same shape for each of their operations. Across the corpus the figures are 20 and 15 of 138 local-core entries. This agreement is about shape: union arms, lengths and shape tokens. It never covers values or bytes (corpus/mirrors.json).
  • history variants are captured verbatim from the public ledger data set. xdrcheck-corpus refetch checks the object hash, the decompressed hash and the offset again. Of the cited history variants, 12 have regenerated: true: the object was written by a later stellar-core than the ledger’s own era. Such an object is real stellar-core output for that ledger, but not the bytes a node wrote at the time. 17 have regenerated: false.
  • Hash chain. 2 claims about when a result code occurs rest on the pubnet scan. That scan reads every ledger from 2 to 65,505,999 (65,505,998 ledgers) and refuses 0. It checks 65,505,997 previous-ledger hash links and finds 0 mismatches (evidence/mainnet-scan).
  • synthetic and core-golden variants: a synthetic variant follows a recorded recipe, and its entry says why stellar-core cannot emit it. A core-golden variant must equal a frame of stellar-core’s golden file at the pinned tag.
Probes, by family. Every probe source fits one or more families below. If one fits none, the build fails. Some probe sources record design reasoning or cite an existing docs page rather than a measurement. They back only these recommendations: C-7041, C-7042, C-7043, C-7044, C-8017, C-9027. Their verification is code_reference, so the tables count them as code references. The result files of the SDK decoding matrix are not committed: evidence/sdk-matrix/run.sh regenerates them, and the corpus entries and conformance columns cite the vectors that matter. Code references. Each code source names a file and lines at a fixed commit: The claims cite 173 distinct line ranges in 51 files. review/code-refs.tsv (guide/review/code-refs.tsv) records whether each range exists at its pinned commit: 173 of 173 do. build.py --check fails when a claim cites a range the file does not hold, or one resolved at another commit. The file is regenerated from clones of the three repositories with build.py --resolve-refs. The field table’s own 94 code references can be printed with their source lines by scripts/py/check_fields.py --print-refs, and it fails on a reference that does not exist. A line that exists is not proof that it says what the claim says. Only reading it shows that, which is the reviewer’s task. Links in the guide and in this report use the release tag v30.0.0 for stellar-core, which is commit 0e49751aa9ddf892302cca02cea7f3f5d97df9f1. For stellar-xdr and rs-soroban-env they use the commit itself.

2.3 The field table is checked against the corpus

The field-by-field table on the page is never written by hand. xdrcheck-corpus guide-table generates it from corpus/fields.yaml joined with the corpus manifest: each line’s status, stage, code link and “demonstrated by” entries. make verify-guide OFFLINE=1 then fails when:
  • the committed table differs from a fresh build;
  • a sentence differs from its claim record (guide-claims --check);
  • the checklist differs from the linter’s rule registry (xdrlint-docs --check);
  • this review package, this report included, is out of date (guide/review/build.py --check).
make repo-checks (scripts/py/check_fields.py) and xdrcheck-corpus verify also fail when a manifest entry lists a line under a status other than the line’s own. It also runs the stellar-docs repository’s own format and link checks on the pages, in a clone at the pinned docs commit 48259204e3a3431896d4847358e27d8e50813099. The table has 53 rows; a row whose parts differ is split into one line per part. 66 lines are published (49 validated, 0 partially validated, 17 passed through) and 3 are held back until they are established. 33 of the published lines cite a corpus entry that demonstrates them. 33 rest on a code reference alone (review/unverified.md (guide/review/unverified.md)).

2.4 Pinned versions

From pins.toml and the pins block of corpus/fields.yaml:

3. Results in numbers

3.1 Claims per section and evidence class

The reviewer_note fields of 4 claims (C-2411, C-2412, C-4711, C-8013) hold notes the authors wrote. No reviewer wrote them.

3.2 What the guide says core does not validate

These are the claims a reviewer should check first. A ledger can show that core accepted a value. No ledger can show that core has no check for it, and a code reference can miss a check that sits in a file or stage the claim does not cite. The list is derived, not hand-picked (review/high-risk.md (guide/review/high-risk.md) gives each sentence). It holds 68 claims:
  • 17 “passed through” verdicts of the field table, 2 of them on a code reference alone;
  • 51 other statements that core does not check, verify, consume or accept something, 19 of them on a code reference alone.
Of the 68: 45 cite a corpus entry, 2 a probe without an entry, 0 record no executed source, and 21 rest on code alone. In each table, claims on code alone come first.

Passed-through verdicts

Other “core does not check” statements

4. Section by section

The guide has four top-level sections: validated versus unvalidated fields, recommended caps, participants from meta, muxed and signed-payload addresses including the SEP-23 forms; then a checklist that carries the three linter checks to other languages.

4.1 Validated versus unvalidated fields

What it claims. Validation runs in stages, D0 to H. A transaction in a ledger has passed every stateless check against the last closed ledger. A failed result can come from any stateful check. Then, field by field, the section says whether core validates the field, validates part of it, or passes it through untouched, with the stage, the code, the result code and what a reader may and must not assume. Further subsections cover passed-through fields, three places where the common shorthand is wrong, what a failed transaction still leaves in the ledger, which parts of meta are consensus, and how strict decoders are.
  • Corpus entries cited (47): cap-0015/fee-bump-inner-failed, cap-0015/fee-bump-inner-success-inner-fee-nonzero-pre-p25, cap-0015/fee-bump-muxed-fee-source-inner-failed, cap-0015/fee-bump-soroban-inner-fee-zero-from-p25, cap-0021/extra-signers-two-signed-payload-keys, cap-0027/account-merge-to-muxed-destination, cap-0027/muxed-source-and-destination-max-ids, cap-0027/muxed-source-id-above-2-pow-63, cap-0034/tx-too-late-with-fee-charged, cap-0035/clawback-from-muxed-account, cap-0038/claim-atoms-pool-and-order-book-mixed, cap-0038/pool-deposit-use-count-on-asset-trustlines, cap-0038/pool-issuers-order-by-bytes-not-strkey, cap-0040/extra-signer-signed-payload-local, cap-0040/signed-payload-signer-added, cap-0040/signed-payload-signers-1-byte-and-64-byte, cap-0040/tx-authorised-by-signed-payload-signature, cap-0046/auth-unmatched-entries-succeed, cap-0046/failed-soroban-tx-with-sac-transfers-in-diagnostics, cap-0058/create-contract-v2-constructor-golden.c2190f875f6f59fd, cap-0066/p22-restore-with-backfilled-restored-changes, cap-0067/forged-token-event-from-unrelated-contract, cap-0067/legacy-layout-mint-with-admin-topic.pubnet-64870500, cap-0067/memo-id-and-memo-return-in-events.memo-id, cap-0067/memo-id-and-memo-return-in-events.memo-return, cap-0067/memo-text-that-looks-like-an-address, cap-0067/sac-transfer-to-muxed-account, cap-0067/sep41-map-form-burn-approve-events, cap-0067/to-muxed-id-absent, cap-0067/to-muxed-id-bytes-from-memo-hash, cap-0067/to-muxed-id-text-from-memo, cap-0067/to-muxed-id-text-non-utf8-memo, cap-0067/to-muxed-id-u64-from-muxed-destination-over-memo, cap-0067/to-muxed-id-void-custom-token, cap-0073/sac-trust-creates-trustline, cap-0087/ml-dsa-sized-byte-arguments, classic/create-account-funded-from-root, classic/failed-and-successful-path-payment-amounts, classic/hash-x-signer-preimage-0-and-64-bytes, classic/manage-data-binary-value-and-delete.binary-value, classic/operations-100-signatures-20, protocol-01/earliest-ledger-v0-envelope-text-memo, protocol-13/upgrade-ledger-header-13-envelopes-v0, xdr-strict/api-transaction-parse-trailing-bytes, xdr-strict/bool-value-2, xdr-strict/length-prefix-exceeds-input, xdr-strict/union-discriminant-undeclared.scaddress-type-6.
  • Probes, by family (claims citing each): local-core 37, captured 18, sdk-matrix 9, print-xdr 4, scan 2.
  • Code: 151 distinct line ranges in 48 files (stellar-core 126, rs-soroban-env 18, stellar-xdr 7).
How it was verified.
  • Stages and the general rule (C-001 to C-011, C-6101 to C-6104, C-6162 to C-6171) cite the stage functions at v30.0.0. Every cited range exists at the pinned commit (section 2.2). The field table’s 257 claims share these stage definitions.
  • Validated and partially validated verdicts rest on the core check each line cites. Where a corpus entry lists the line, the entry shows the check at work: a ledger with the stated result code, or a value at the edge of what the check allows. 33 published lines have no such entry yet. They are listed with their code links in review/unverified.md (guide/review/unverified.md), and question 1 of review/questions.md (guide/review/questions.md) asks the reviewer to check them first.
  • Passed-through verdicts (section 3.2): where an entry exists, it shows an attacker-shaped value accepted into a ledger. Examples: a muxed ID above 2^63 (cap-0027/muxed-source-id-above-2-pow-63), a non-UTF-8 memo (cap-0067/to-muxed-id-text-non-utf8-memo), a forged token event (cap-0067/forged-token-event-from-unrelated-contract), unmatched authorization entries in a successful transaction (cap-0046/auth-unmatched-entries-succeed). Lines without an entry rest on code.
  • Failed transactions and the time-bound race (C-6144 to C-6146, C-6161, C-6172): local-core runs, a captured pubnet ledger (cap-0034/tx-too-late-with-fee-charged) and the pubnet scan.
  • Meta consensus (C-6147 to C-6150): code for the hashing and the node flags. For the regenerated public data set, the lake object metadata of the captured ledgers.
  • Decoder strictness (C-6151 to C-6160): the SDK decoding matrix, and stellar-core’s own print-xdr transcripts under evidence/core-print-xdr (evidence/core-print-xdr/README.md). These show how core’s decoder behaves offline, not what core admits into a ledger, and the guide says so in C-6151.
What remains open. The held-back lines and the open points inside published rows (section 5). What it claims. A meta reader must never cap below what core admitted. A pre-submission reader may cap tighter if it documents that choice. Each decoder counts depth in its own unit. The section then gives what core’s decoder, its envelope admission check and the Soroban host admit, the defaults the SDKs ship with, and what to do at a bound.
  • Corpus entries cited (4): cap-0003/p10-upgrade-liabilities-50936-changes, a contract event nested as deep as the Soroban host converts, a limits/ entry at the Soroban host’s value limit, xdr-strict/length-prefix-exceeds-input.
  • Probes, by family (claims citing each): local-core 2, captured 2, sdk-matrix 19, print-xdr 6, network-config 2, design 4.
  • Code: 18 distinct line ranges in 10 files (stellar-core 7, rs-soroban-env 5, stellar-xdr 6).
How it was verified.
  • Units per nesting level, per decoder (C-7005 to C-7011): measured by bisection. For the SDKs, in the depth group of the SDK decoding matrix. For stellar-core, with the release binary’s print-xdr, whose transcript is evidence/core-print-xdr/out/depth.txt.
  • Limits core admits (C-7013 to C-7018): each limit is read from code at the pinned commits. The decoder bound and the host value limit were also measured: the decoder bound with print-xdr, and the host value limit in a local-core scenario run in which a contract emits a nested vector in an event and the next level fails (a contract event nested as deep as the Soroban host converts).
  • Derived bounds (C-7019 to C-7023) are marked as derived on the page. They are arithmetic on the measured units and the code limits. The page labels each one as measured, derived or read from code.
  • SDK defaults (C-7024 to C-7031) are measured on the releases in section 2.4.
  • Sizes (C-7038, C-7045, C-7051, C-7052) come from a captured pubnet ledger and a read of the network configuration at a stated ledger.
  • Recommendations (C-7033 to C-7048) and the summary table (C-7049 to C-7054) follow from the facts above. C-7041 to C-7044, on batching and on what to do at a bound, rest on design reasoning, and their records say so.
What remains open. The envelope admission depth that core releases enforced before Protocol 28. The deepest nesting a failed transaction’s envelope can hold in each era. The cost per level of authorization sub-invocations (question 3; section 5).

4.3 Participants from meta

What it claims. Who a transaction affected comes from meta: fee processing, transaction-level changes, operation changes, Soroban changes and consensus events. It does not come from the envelope. The section gives a step-by-step algorithm, a table of cases where envelope and meta disagree, and a worked example of a transaction refused at apply.
  • Corpus entries cited (14): cap-0001/account-merge-seqnum-too-far, cap-0015/fee-bump-muxed-fee-source-inner-failed, cap-0027/muxed-source-and-destination-max-ids, cap-0034/tx-too-late-with-fee-charged, cap-0038/claim-atoms-pool-and-order-book-mixed, cap-0046/auth-unmatched-entries-succeed, cap-0067/to-muxed-id-text-from-memo, cap-0073/sac-trust-creates-trustline, cap-0084/sac-transfer-to-muxed-contract, cap-0088/first-ledger-with-signed-ms, cap-0088/upgrade-ledger-header-30-whole-second-value, protocol-13/upgrade-ledger-header-13-envelopes-v0, protocol-20/upgrade-ledger-lcm-v0-header-20, protocol-23/upgrade-ledger-lcm-v1-header-23.
  • Probes, by family (claims citing each): local-core 3, captured 5, sdk-matrix 3, replay 1, docs-page 1.
  • Code: 27 distinct line ranges in 12 files (stellar-core 20, rs-soroban-env 6, stellar-xdr 1).
How it was verified.
  • Each row of the disagreement table (C-8025 to C-8033) and most algorithm steps name a corpus entry. A ledger entry states the participants, state changes and movements a correct reader produces, and xdrcheck-corpus verify (the extractor check) fails when the reference extractor’s output differs from them.
  • The worked example (C-8034, C-8035) is the captured pubnet ledger cap-0034/tx-too-late-with-fee-charged. Its values are read from the ledger’s bytes.
  • Protocol-boundary advice (C-8019 to C-8024) rests on the SDK decoding matrix on a Protocol 30 ledger, captured upgrade ledgers (protocol-13/upgrade-ledger-header-13-envelopes-v0, cap-0088/first-ledger-with-signed-ms), the code that sets the meta version on an upgrade ledger, and a replay of an indexer that halts on one error.
  • The authorizer rule (C-8013, C-8014) rests on cap-0046/auth-unmatched-entries-succeed. There, a nonce entry is created only for the matched entry.
What remains open. Some claims here are marked executed, but their records cite no corpus entry or probe (section 5, item 8). Question 9 asks whether the event rule (C-8009 to C-8011) is complete.

4.4 Muxed and signed-payload addresses, including the SEP-23 forms

What it claims. The section covers every SEP-23 strkey form with its version byte, payload and length, and the XDR types that render to each. It covers storage widths, muxed accounts and the event routing ID, muxed contracts (CAP-84), signed-payload signers (CAP-40), how to decode addresses safely, and how to carry 64-bit IDs.
  • Corpus entries cited (24): cap-0021/extra-signers-two-signed-payload-keys, cap-0027/api-muxed-account-id-encoding, cap-0027/muxed-source-and-destination-max-ids, cap-0027/muxed-source-id-above-2-pow-63, cap-0027/muxed-source-id-zero, cap-0040/signed-payload-signer-added, cap-0040/signed-payload-signers-1-byte-and-64-byte, cap-0040/tx-authorised-by-signed-payload-signature, cap-0067/legacy-layout-mint-with-admin-topic, cap-0067/to-muxed-id-text-from-memo, cap-0067/to-muxed-id-u64-from-muxed-destination-over-memo, cap-0067/to-muxed-id-void-custom-token, cap-0084/api-muxed-address-field-order, cap-0084/sac-transfer-to-muxed-contract, classic/api-offer-id-encoding, classic/muxed-id-max-and-offer-id-max-roundtrip, strkey/api-encode-signed-payload-empty, strkey/sep-0023-v1.4.0-vectors, strkey/sep-0023-v1.4.0-vectors.valid-01, strkey/sep-0023-v1.4.0-vectors.valid-03, strkey/sep-0023-v1.4.0-vectors.valid-05, strkey/synthetic-gap-vectors, strkey/synthetic-gap-vectors.signed-payload-64-byte, strkey/truncated-muxed-address-56-chars.
  • Probes, by family (claims citing each): local-core 2, captured 7, sdk-matrix 9, docs-page 1.
  • Code: 17 distinct line ranges in 9 files (stellar-core 10, rs-soroban-env 1, stellar-xdr 6).
How it was verified.
  • The form table and the worked examples (C-9001 to C-9003, C-9032 to C-9038) come from the SEP-23 v1.4.0 test vectors, stored as corpus entries (strkey/sep-0023-v1.4.0-vectors) and decoded by every SDK column. The widths are measured on synthetic vectors (strkey/synthetic-gap-vectors).
  • Decoder behaviour on edge cases (C-9005, C-9021, C-9026, C-9028, C-9029): groups 9 and 3 of the SDK decoding matrix, and strkey/truncated-muxed-address-56-chars.
  • Muxed accounts, the routing ID and muxed contracts (C-9007 to C-9021): local-core runs with muxed sources and destinations, the maximum ID and ID zero, and captured pubnet and testnet ledgers for each to_muxed_id form, including the CAP-84 muxed contract on testnet (cap-0084/sac-transfer-to-muxed-contract).
  • Signed-payload signers (C-9022 to C-9026): a captured testnet ledger adding the signer (cap-0040/signed-payload-signer-added), a captured ledger with two signed-payload extra signers (cap-0021/extra-signers-two-signed-payload-keys), and a local-core run with payloads of 1 and 64 bytes (cap-0040/signed-payload-signers-1-byte-and-64-byte). That core refuses an empty payload rests on code.
  • Storage advice (C-9003, C-9004, C-9006) and 64-bit IDs (C-9030, C-9031) follow from the SEP text and the matrix’s ID-encoding vectors.
What remains open. Question 7 (whether an explicit void routing ID should mean “no ID”) and question 8 (whether SEP-23 should state the 1 to 64-byte payload bound).

4.5 The checklist, carrying the three linter checks to other languages

What it claims. One item for each check an indexer in any language should pass, grouped by mistake. Each item names the xdrlint rule that checks it in Go code, or explains why it is checked by hand and how. How it was verified. The checklist has no records in claims.yaml: it restates the guide’s claims as checks. go run ./cmd/xdrlint-docs --check (in make verify-guide) generates it from the linter’s rule registry. Every one of the 21 rules has an item. 21 of the 35 items are checked by the linter and 14 by hand. make lint-fixtures runs the linter’s tests, which check each rule against fixture code with seeded violations. Every corpus entry the checklist names exists (scripts/py/check_guide_entries.py, run by build.py --check). The guide sections each item links to are checked by the docs repository’s link check. What remains open. Manual items are checked only by the procedure each one describes. They have no automated test.

5. Known limits and open questions

Each item states what is not established and what would establish it. review/questions.md (guide/review/questions.md) holds the 11 questions put to the reviewer. The page lists the first four items under “Not yet established”. The field table holds back these lines: VF-14b, VF-45b, VF-50c.
  1. A time-bound failure in a ledger after CAP-34 (VF-14b; C-6173, C-6172; question 11). The corpus shows txTOO_LATE with the fee charged only in Protocol 13. The pubnet scan finds no txTOO_LATE after the upgrade to Protocol 15, and the corpus has no txTOO_EARLY. To close it: a local-core run at Protocol 30 that closes a ledger whose close time passes a queued transaction’s maxTime (and one before its minTime), recording whether the transaction lands with a failed result or is dropped. Then a scan query for txTOO_EARLY.
  2. An archived-entry index that names a live entry (VF-45b; question 4). To close it: a local-core run at Protocol 30 submitting a Soroban transaction whose archivedSorobanEntries marks a read-write entry that is live, recording core’s answer (refused at submission, failed at apply, or applied).
  3. The deepest nesting core admits in a failed transaction’s envelope, per era, and the depth of authorization sub-invocation trees (VF-50c, VF-49; C-7012, C-7014, C-7023; question 3). To close it: bisect the nesting of an argument in an InvokeHostFunction transaction that fails, submitted to the pinned release, and record the last depth admitted. For eras before Protocol 28, repeat with the core release that was live then. Measure the decoder cost per level of a sub-invocation tree in every decoder column.
  4. Plain enum fields at admission, and non-canonical encodings in meta (C-6156, C-6160; question 10). To close it: submit to a local core an envelope with an undeclared value in a plain enum field, and one with a boolean encoded as 2. Record whether core admits either into a ledger, and read the admission code with a core engineer.
  5. Open points inside published rows. Where a clawback whose from is a muxed form of the issuer fails at apply (VF-20, VF-34; question 5). Whether unused CAP-71 delegate signatures are tolerated (VF-48; question 2). The deepest authorization tree in a successful and a failed transaction (VF-49). To close them: one local-core run for each, recording the result code and the meta.
  6. How the regenerated public data set was produced (C-6150; question 6). The object metadata names stellar-core 23.0.1. The configuration flags it ran with are not recorded. To close it: a statement from the data set’s maintainers, or a regeneration of one captured ledger with stated flags and a byte comparison.
  7. Conventions the code cannot settle (questions 7 to 9): whether an explicit void to_muxed_id means “no routing ID”, the signed-payload bound in SEP-23, and the event rule for token movements. To close them: the reviewer’s answer, and where needed a SEP change proposed by the maintainers.
  8. Gaps in the records. Executed, but no executed source recorded: none. Code reference, but the record cites no code line (13): C-6136 (cap+corpus), C-6160 (probe), C-7041 (probe), C-7042 (probe), C-7043 (probe), C-7044 (probe), C-8017 (probe), C-9003 (sep), C-9004 (sep), C-9006 (sep), C-9011 (cap), C-9016 (cap), C-9027 (probe). To close them: add the corpus entry or probe each one rests on, or change its verification to what the record shows. Only the authors change records, and this report changes no verdict.
  9. Limits of the method itself. A line that exists may not say what the claim says. A ledger shows that a value was accepted, not that no check exists. Golden agreement covers shape only. Local-core meta matches only after normalisation. Regenerated history is not the bytes of its time. print-xdr shows the decoder, not admission. To close them: only the engineer’s review closes the first two for code-only claims, and section 3.2 lists those first.

6. For the reviewer

Check one claim in five minutes

  1. Find the claim. On the page, the tag follows the sentence. In guide/claims.yaml, search for claim_id: C-nnnn. review/claims.csv (guide/review/claims.csv) has the same record with a link to every cited source line at its pinned commit.
  2. Read the sources. A code source is a file and line range at the commit in its at field. The CSV links open it upstream.
  3. If it cites a corpus entry, look at the bytes:
    The entry’s manifest, under corpus/manifest/, names the blob, its SHA-256, the source (a local-core run, a captured ledger or a recipe) and the values a correct reader derives.
  4. If it cites a probe, open the run’s folder (section 2.2, “Probes, by family”). Each README gives the command that runs it again.
  5. Check the whole set at once:

Record the review

Each answer changes one claim’s reviewer_status in guide/claims.yaml:
  • confirmed;
  • corrected: the page sentence and the record’s sentence change together;
  • disputed: the objection goes in reviewer_note.
guide/REVIEW.md gives the full procedure. Then run .venv/bin/python guide/review/build.py to rebuild the package and this report, and make verify-guide OFFLINE=1. The counts in the first section and the tables follow the records. The docs pull request is opened only when no claim is disputed and every fact claim is confirmed or corrected.