@conarium-ai/core 0.2.32 → 0.2.34

This diff represents the content of publicly available package versions that have been released to one of the supported registries. The information contained in this diff is provided for informational purposes only and reflects changes between package versions as they appear in their respective public registries.
package/CHANGELOG.md CHANGED
@@ -1,5 +1,65 @@
1
1
  # Changelog
2
2
 
3
+ ## 0.2.34 — 2026-08-19 — the prior-art claim that got narrower
4
+
5
+ A comparative claim in this package was too strong, and the correction came from
6
+ reading a competitor's source rather than from a gate of ours catching it.
7
+
8
+ - **`docs/PRIOR-ART.md` no longer says nobody else has arrived at the idea.**
9
+ Vaara (`vaaraio/vaara`, AGPL-3.0) specifies a coverage reconciliation in
10
+ `credential-broker-spec.md` §D — joining each used credential to the receipt it
11
+ should have produced, and reading a mismatch as a bypassed broker — with its
12
+ limits stated in §E. The row is scored from source: the tree at `befdced` was
13
+ cloned, its contiguity tests run, and searched for code implementing that
14
+ design. None was found, so the column reads **"specified, not found
15
+ implemented"** rather than yes.
16
+ - **That row was scored twice, and the first scoring is left on the page.** It
17
+ first read "Yes", on the strength of the design section alone — scoring a
18
+ specification as behaviour, which is the confusion this file's own column
19
+ definitions warn about two sections above. The correction is recorded where the
20
+ row is, not quietly.
21
+ - **What survives is stated in the narrower form.** The claim stands on
22
+ implementation, which is what every other row is scored on. What it can no
23
+ longer carry is that nobody else has had the idea. Anyone quoting the file
24
+ should quote *"we have not found another implementation that masks values
25
+ before the model sees them and then reconciles what it disclosed"* — not
26
+ *"nobody else reconciles"* and not *"none of the ten has the third column"*.
27
+ - **The Microsoft row said "Merkle-chained" and their tutorial describes a linear
28
+ `parent_receipt_hash`.** Overstating a competitor's mechanism is the same defect
29
+ as overstating our own, so it is corrected in the same pass.
30
+ - **`README.md` says eleven projects, and says which way the eleventh moved the
31
+ claim.** A comparison page that only ever gains rows it wins is not evidence.
32
+ - **`docs/security/NPM-PROVENANCE.md` documents the three release artefacts the
33
+ full publish path is configured to produce** — the published tarball, a
34
+ CycloneDX bill of materials, and a build attestation written to GitHub's
35
+ attestation store — and says which of them any given release actually carries.
36
+ The artefact path first ran on 19 August in `artefacts` mode against 0.2.33,
37
+ which now carries the tarball and the attestation but no bill of materials.
38
+ - **The changelog had stopped one version short of the package again**: 0.2.33
39
+ shipped before its entry was written, and that entry is below. The publish
40
+ workflow now refuses a version whose section does not exist, so the next
41
+ occurrence is a red gate rather than a discovery.
42
+
43
+ ## 0.2.33 — 2026-08-19
44
+
45
+ Written on 2026-08-19, after 0.2.33 had already shipped without it. The gap is
46
+ recorded rather than backdated.
47
+
48
+ - **The provenance page stopped saying there was nothing to verify.**
49
+ `docs/security/NPM-PROVENANCE.md` still read *"This is not a published
50
+ release"* eight releases after the first one, and pointed readers at
51
+ `gh attestation verify`, which answers `404` for a package published with npm
52
+ provenance — that command reads GitHub's attestation store, while
53
+ `npm publish --provenance` writes to the npm registry and its transparency log.
54
+ A reader checking the honest way would have read that 404 as "no attestation",
55
+ which is the opposite of the truth.
56
+ - The working commands replaced it: `npm view … dist --json`, which returns the
57
+ `attestations` block naming `slsa.dev/provenance/v1`, and `npm audit
58
+ signatures`. The `gh` command stayed on the page as a warning with its reason,
59
+ because a reader who tries it deserves to know why it fails.
60
+ - Found by an external review, not by our own claim gate. The gate locks phrasings
61
+ already caught; this file was never in the list of surfaces it reads.
62
+
3
63
  ## 0.2.32 — 2026-08-19
4
64
 
5
65
  Tool descriptions now say what the tools do to the world, not just what they are
package/README.md CHANGED
@@ -421,11 +421,17 @@ is verifiable without trusting us; this particular figure is our own measurement
421
421
  and [Governance Report 001](https://conarium.dev/report-001.html) lists its limits.
422
422
 
423
423
  That claim is hedged on purpose, and [`docs/PRIOR-ART.md`](docs/PRIOR-ART.md) is the
424
- evidence behind it: ten projects checked on 6 August 2026, what each one has, the
425
- closest academic prior art ([Sello / *Notarized Agents*](https://arxiv.org/html/2606.04193v1),
426
- which names this gap better than we did), and nine things we could **not** verify. If
427
- you know of an implementation combining all three, open an issue and it will be
428
- corrected.
424
+ evidence behind it: eleven projects — ten checked on 6 August 2026 and Vaara added
425
+ on 19 August — what each one has, the closest academic prior art
426
+ ([Sello / *Notarized Agents*](https://arxiv.org/html/2606.04193v1), which names this
427
+ gap better than we did), and nine things we could **not** verify. If you know of an
428
+ implementation combining all three, open an issue and it will be corrected.
429
+
430
+ ⚠️ The Vaara row narrowed this claim rather than confirming it. That project
431
+ *specifies* a coverage reconciliation in its design documents; a search of its
432
+ tree found no code running it, so the row reads "specified, not found
433
+ implemented". The idea is not ours alone — the running code, as far as this scan
434
+ reaches, still is, and the file says so above its own table.
429
435
 
430
436
  ---
431
437
 
package/docs/PRIOR-ART.md CHANGED
@@ -7,6 +7,11 @@ We make one comparative claim about Conarium:
7
7
  > enforcement, and (3) coverage reconciliation against the data source's own
8
8
  > query counters.
9
9
 
10
+ ⚠️ **Read that as a claim about implementations.** Since 19 August this file
11
+ carries a project that *specifies* part 3 without our finding it implemented —
12
+ see the Vaara row and the note under the table. The idea is not ours alone; the
13
+ running code, as far as this scan reaches, still is.
14
+
10
15
  A claim like that is unfalsifiable if we just assert it, so this file is the
11
16
  evidence behind it: what we searched, on what date, what we found, and — the part
12
17
  most comparison pages leave out — **what we could not verify.** If you know of an
@@ -16,6 +21,14 @@ implementation that does all three, open an issue and this file will be correcte
16
21
  official site, plus targeted web searches. **Source-code-level search was not
17
22
  performed on any repository** — see Limitations.
18
23
 
24
+ **Amended 19 August 2026.** Vaara was added after its author raised a related
25
+ mechanism on the SCITT mailing list. It is the one row scored from source rather
26
+ than from documentation: the repository was cloned at `befdced`, its contiguity
27
+ tests were run (`PYTHONPATH=src python -m pytest tests/credential/test_contiguity.py`
28
+ → 18 passed), the cited design sections were read, and the tree was searched for
29
+ code implementing them. That search is why the row was rescored the same day —
30
+ see the note under the table.
31
+
19
32
  ## What the three parts mean
20
33
 
21
34
  Precision matters here, because two of them are commonly confused:
@@ -37,7 +50,7 @@ perfectly intact.
37
50
 
38
51
  | Project | Enforcement | Portable signed receipt | Coverage reconciliation |
39
52
  |---|---|---|---|
40
- | [microsoft/agent-governance-toolkit](https://github.com/microsoft/agent-governance-toolkit) | Partial — Cedar allow/deny, DLP attribute ratchets; PII masking not documented | **Yes** — Ed25519, Merkle-chained, `verify_receipt_chain()` | No |
53
+ | [microsoft/agent-governance-toolkit](https://github.com/microsoft/agent-governance-toolkit) | Partial — Cedar allow/deny, DLP attribute ratchets; PII masking not documented | **Yes** — Ed25519, hash-chained (`parent_receipt_hash`), `verify_receipt_chain()`. This said "Merkle-chained" until 19 August; their tutorial describes a linear parent hash, and overstating a competitor's mechanism is the same defect as overstating our own | No |
41
54
  | [Signet](https://github.com/Prismer-AI/signet) | Partial — policy bound into the receipt; no masking | **Yes** — Ed25519, hash-chained, offline | No |
42
55
  | [Handshake.AI](https://handshake.ai/) | No | **Yes** — signed action receipts, DID-issued, offline | No |
43
56
  | [hoophq/hoop](https://github.com/hoophq/hoop) | **Yes — strongest here.** ML-based masking before bytes leave the gateway | No — session recording/replay, no signed offline-verifiable artifact found | No |
@@ -47,11 +60,42 @@ perfectly intact.
47
60
  | [lasso-security/mcp-gateway](https://github.com/lasso-security/mcp-gateway) | **Yes** — Presidio PII masking, injection filters | No — no signing mechanism documented | No |
48
61
  | [h33.ai](https://h33.ai/) | **Yes, and further** — FHE; the model never sees plaintext | **Yes** — ZK-STARK + Dilithium, verifiable "years later, offline, without the original vendor" | No |
49
62
  | [CertNode](https://certnode.io/solutions/ai-agents) | Weak — no blocking gate | **Yes** — ES256 JWS + RFC 3161 | No |
63
+ | [vaaraio/vaara](https://github.com/vaaraio/vaara) | Partial — `CredentialGateway` authorizes each tool call against a brokered credential and refuses without one; no value masking found. Its redaction is GDPR Art. 17 erasure inside the audit store, which is a different guarantee | **Yes** — signed decision record (ES256/HS256/RS256) over JCS-canonical blocks, evidence pinned by `evidenceRef.digest`, standalone checkers under `tests/vectors` that import none of their code | **Specified, not found implemented.** `docs/design/credential-broker-spec.md` §D designs the join — each used credential to a receipt on `attestationDigest`, a used credential with no matching receipt read as a bypassed broker — and §E states its limit: *"detection of a defeated broker, not a mathematical-completeness claim"*. At `befdced` we found no collector, join, CLI or test implementing it; `gateway.py` points at the design document for that residual rather than at code. **This is the only design for it we have found outside our own** |
50
64
 
51
65
  The landscape splits cleanly. One camp enforces well but issues no portable evidence
52
66
  (hoop.dev, Lasso). The other issues excellent portable evidence but does not enforce
53
- before the model (Signet, Circe, Acta, agentreceipts, Handshake, CertNode). Two
54
- projects do both — Microsoft's toolkit and h33.ai — and neither does the third.
67
+ before the model (Signet, Circe, Acta, agentreceipts, Handshake, CertNode). Three
68
+ projects do the first two — Microsoft's toolkit, h33.ai and Vaara — and none of
69
+ them was found running the third.
70
+
71
+ ### What the Vaara row cost the claim, and what it did not
72
+
73
+ This row was scored twice on 19 August, and the first scoring was wrong in a way
74
+ worth leaving on the page.
75
+
76
+ It first read **Yes** on coverage reconciliation, on the strength of
77
+ `credential-broker-spec.md` §D. That section designs the join and states its own
78
+ limit honestly, and reading it is enough to see that someone else has had this
79
+ idea and thought it through. It is not enough to see it run. A search of the tree
80
+ at `befdced` for a collector of used credentials, a join, a CLI or a test found
81
+ none, and `gateway.py` refers that residual to the design document rather than to
82
+ code. Scoring a specification as behaviour is the same confusion this file's own
83
+ column definitions warn about two sections above — *"not that the log records that
84
+ it happened, but that the value never left"* — committed by the person who wrote
85
+ them.
86
+
87
+ So the combined claim stands on implementation, which is what every other row in
88
+ this table is scored on. What it can no longer carry is the suggestion that
89
+ nobody else has arrived at the idea. Someone has, in public, with the limits
90
+ stated. The honest form:
91
+
92
+ > **We have not found another implementation that masks values before the model
93
+ > sees them and then reconciles what it disclosed against the source's own
94
+ > bookkeeping. One other project specifies such a reconciliation; we did not find
95
+ > it implemented.**
96
+
97
+ Anyone quoting this file should quote that, not "nobody else reconciles" and not
98
+ "none of the ten has the third column".
55
99
 
56
100
  ### Two near misses worth naming precisely
57
101
 
@@ -0,0 +1,73 @@
1
+ # Claims review — 0.2.33
2
+
3
+ - base: v0.2.32
4
+ - head: b8756f3
5
+ - surface: sha256:8eeabf25cea27362e070789c61c828eea291bd3c155362f098894d868f0f9b4b
6
+ - verdict: pass
7
+ - reviewer: Claude, reading on the maintainer's instruction — an assistant read, not an independent human one
8
+
9
+ One surface changed and one surface was added: `docs/security/NPM-PROVENANCE.md`
10
+ is rewritten and listed. Eighteen documents are now read at release time.
11
+
12
+ This release exists because of an external review of the repository, which
13
+ flagged the provenance page as stale. Reading it confirmed two defects that are
14
+ worse than staleness, and both are recorded below as findings rather than as
15
+ housekeeping — a reader who followed that page would have drawn a false
16
+ conclusion about whether this package can be trusted.
17
+
18
+ ## Findings
19
+
20
+ ### 1. docs/security/NPM-PROVENANCE.md:6 — the page denied a fact about itself
21
+
22
+ - Claim: `"This is **not** a published release."`, followed by a verification
23
+ section headed `"How to verify (after a provenance-backed release exists)"`
24
+ and a setup section headed `"One-time setup before the first publish"`.
25
+ - Mechanism: the npm registry, which records what was actually published.
26
+ - Why it falls short: measured on 0.2.32 —
27
+ `npm view @conarium-ai/core@0.2.32 dist --json` returns an `attestations`
28
+ block naming `https://slsa.dev/provenance/v1` plus a registry `signatures`
29
+ entry, and eight versions have shipped since the sentence was written. The
30
+ page described a pre-release state inside a released package, and it ships in
31
+ the tarball, so the reader most likely to open it is someone who already has
32
+ the published artefact in their hands.
33
+ - Disposition: rewritten. The page now states what exists, and the verification
34
+ section carries the measured output rather than a promise that something will
35
+ one day be verifiable.
36
+
37
+ ### 2. docs/security/NPM-PROVENANCE.md — the verification command cannot succeed
38
+
39
+ - Claim: `"or, with GitHub CLI, verify the attestation against this repo: gh
40
+ attestation verify --owner dogrucanemek-alt <downloaded.tgz>"`
41
+ - Mechanism: `gh attestation verify` reads GitHub's own attestation store, which
42
+ is populated by `actions/attest-build-provenance`.
43
+ - Why it falls short: `npm publish --provenance` writes to the npm registry and
44
+ the public transparency log, not to GitHub's store. Both forms were run
45
+ against the published 0.2.32 tarball and both answered **HTTP 404** —
46
+ `--owner dogrucanemek-alt` (which additionally targets an org endpoint for
47
+ what is a user account) and `--repo dogrucanemek-alt/conarium`. The failure
48
+ mode is the dangerous direction: a reader checking our provenance would see
49
+ 404 and reasonably conclude the attestation is missing, when it exists and
50
+ verifies by another route.
51
+ - Disposition: replaced with two commands that were run and do work —
52
+ `npm view … dist --json` and `npm audit signatures` (107 registry signatures
53
+ and 11 attestations verified on a clean install, exit 0). The `gh` command is
54
+ kept in the page as an explicit warning, with the reason, so the next person
55
+ who tries it does not repeat the diagnosis.
56
+
57
+ ### 3. test/denetci.mjs — the page was never on the claim surface list
58
+
59
+ - Claim: not a sentence — the absence of a reader. The page ships in the tarball
60
+ and its whole subject is how much of what we say can be checked without us.
61
+ - Mechanism: the `SURFACES` list in `test/denetci.mjs`.
62
+ - Why it falls short: it was unlisted, so both defects above survived eight
63
+ releases and were found by someone outside this process rather than by the
64
+ gate. This is the third instance of the same shape — `docs/PRICING.md` in
65
+ 0.2.31, the tool descriptions in 0.2.32, and now this — which is enough of a
66
+ pattern to state plainly: the list is hand-written, and things that ship
67
+ keep arriving outside it.
68
+ - Disposition: listed. The structural repair suggested by 0.2.32 — deriving
69
+ reviewable surfaces rather than maintaining a list — is still not done, and
70
+ this finding is the argument for doing it.
71
+
72
+ Nothing else on the claim surface moved: the surface hash covers eighteen
73
+ documents and only this one differs from 0.2.32.
@@ -0,0 +1,119 @@
1
+ # Claims review — 0.2.34
2
+
3
+ - base: v0.2.33
4
+ - head: 0198ccf
5
+ - surface: sha256:e80fad84574da3554959e855d5e9cb4db1a32b9e4cc3f58f3dd87db369dad5c3
6
+ - verdict: pass
7
+ - reviewer: Claude (Opus 5), then an independent pass by Codex and one by Cursor
8
+ before the version was pushed
9
+
10
+ Four surfaces changed: `docs/PRIOR-ART.md`, `README.md`, `CHANGELOG.md`,
11
+ `docs/security/NPM-PROVENANCE.md`. The release is itself a claims correction, so
12
+ most of the diff removes reach. The findings below are ordered by who found them,
13
+ because that ordering is the useful part of this record: **the first pass missed
14
+ the two that mattered, and both were found by readers who had not written the
15
+ text.**
16
+
17
+ ## Findings
18
+
19
+ ### 1. docs/PRIOR-ART.md — a design section scored as behaviour
20
+
21
+ - Claim: the Vaara row read **"Yes — the only other one here"** on coverage
22
+ reconciliation, and the file, `README.md` and `CHANGELOG.md` all said that
23
+ project *reconciles*.
24
+ - Mechanism: `docs/design/credential-broker-spec.md` §D at `befdced`, which
25
+ designs a join of used credentials to receipts on `attestationDigest`.
26
+ - Why it falls short: §D is a design document. A search of that tree for a
27
+ collector of used credentials, the join, a CLI or a test found none, and
28
+ `src/vaara/credential/gateway.py` refers that residual to the design document
29
+ rather than to code. Every other row in the table is scored on what a project
30
+ is found doing. Scoring a specification as behaviour is the confusion this
31
+ file's own column definitions warn about — *"not that the log records that it
32
+ happened, but that the value never left"* — committed by the person who wrote
33
+ them, in the row that changed the headline claim.
34
+ - Disposition: rescored **"Specified, not found implemented"** in `0198ccf`,
35
+ with the first scoring left visible on the page rather than quietly replaced.
36
+ `README.md` and `CHANGELOG.md` narrowed to match. The retired phrasings are
37
+ named so they are not requoted: *"nobody else reconciles"*, *"none of the ten
38
+ has the third column"*.
39
+ - **Found by: Codex.** Not by the first pass, which had read the same section and
40
+ taken it at its word.
41
+
42
+ ### 2. docs/PRIOR-ART.md — the summary paragraph contradicted the new row
43
+
44
+ - Claim: *"Two projects do both — Microsoft's toolkit and h33.ai — and neither
45
+ does the third."*
46
+ - Mechanism: the table directly above it.
47
+ - Why it falls short: the sentence sat unchanged three lines under a row that
48
+ had just been scored **Yes** on the third column. The paragraph immediately
49
+ after it was rewritten in the same edit; this one was not. A reader opening the
50
+ file from the post's first comment meets the contradiction before the
51
+ correction.
52
+ - Disposition: rewritten in `0198ccf` to name three projects and say none was
53
+ found running the third.
54
+ - **Found by: Cursor.** The first pass edited the paragraph below it and did not
55
+ re-read the paragraph above.
56
+
57
+ ### 3. docs/security/NPM-PROVENANCE.md — a limitation that expired the same day
58
+
59
+ - Claim: *"Those steps have not run yet — 0.2.34 is the first release that will
60
+ execute them"*, and *"every version published so far has only npm's
61
+ attestation"*.
62
+ - Mechanism: the workflow runs and the release pages.
63
+ - Why it falls short: both were true when written and false a few hours later.
64
+ The `artefacts` mode added in #32 was dispatched against 0.2.33 (runs
65
+ `32295097847`, `32295276511`), which verified the registry tarball, produced a
66
+ GitHub attestation and created that release. `v0.2.33` now carries
67
+ `conarium-ai-core-0.2.33.tgz` and `.tgz.intoto.jsonl`, and
68
+ `gh attestation verify` returns 0 for it.
69
+ - Disposition: replaced in `0198ccf` with what has actually run, including that
70
+ 0.2.33 carries no bill of materials because that step is skipped in artefacts
71
+ mode, and that the full publish path has still not run.
72
+ - **Found by: Codex.** Worth noting how it expired: this was a self-correction in
73
+ the first pass (finding 5 below), written carefully, and then invalidated by
74
+ our own action. A sentence about what has not happened yet has a shelf life.
75
+
76
+ ### 4. CHANGELOG.md — three assets described as an existing state
77
+
78
+ - Claim: *"releases now carry: the published tarball, a CycloneDX bill of
79
+ materials, and a build attestation"*.
80
+ - Mechanism: `publish.yml` produces the SBOM only in `publish` mode and uploads
81
+ only the assets that exist.
82
+ - Why it falls short: no release carries all three. 0.2.33 carries two; 0.2.34
83
+ was not on npm when this was written.
84
+ - Disposition: narrowed to what the publish path is *configured* to produce, with
85
+ 0.2.33's actual contents named. The neighbouring *"stops one version short …
86
+ again"* was in the present tense for a past event and is now *"had stopped"*.
87
+ - **Found by: Codex.**
88
+
89
+ ### 5. docs/security/NPM-PROVENANCE.md — an artefact described before it existed
90
+
91
+ - Claim: *"From 0.2.34 the publish workflow attaches three files to the GitHub
92
+ release"*, over a table of them.
93
+ - Mechanism: the steps added in #28, which had never executed.
94
+ - Why it falls short: it described an intention as an artefact — the defect
95
+ 0.2.33 was published to fix, with the sign reversed.
96
+ - Disposition: reworded in `d816ce1`, then superseded by finding 3 once the steps
97
+ actually ran.
98
+ - **Found by: the first pass**, which is recorded here mainly to be honest about
99
+ its hit rate: it caught the claim it had just written, and missed the two that
100
+ came from reading someone else's code and re-reading its own file.
101
+
102
+ ### 6. docs/PRIOR-ART.md — a competitor's mechanism overstated
103
+
104
+ - Claim: the Microsoft row read *"Ed25519, **Merkle-chained**"*.
105
+ - Mechanism: their own tutorial, which describes a linear `parent_receipt_hash`.
106
+ - Why it falls short: it is wrong, and wrong in their favour. A comparison table
107
+ that inflates a competitor is no more trustworthy than one that inflates its
108
+ author; both mean the rows were not checked.
109
+ - Disposition: corrected in `0198ccf`, with the previous wording named.
110
+ - **Found by: Cursor.**
111
+
112
+ ## Verification note
113
+
114
+ Vaara's contiguity tests were run during the first pass —
115
+ `PYTHONPATH=src python -m pytest tests/credential/test_contiguity.py` → 18
116
+ passed, at `befdced`. Codex could not reproduce that (`rfc8785` missing, no
117
+ package import environment) and correctly flagged that the claim had no command
118
+ recorded against it. The command and result are now in `docs/PRIOR-ART.md` beside
119
+ the claim rather than only in a session transcript.
@@ -3,58 +3,125 @@
3
3
  How a bank checks that an `@conarium-ai/core` tarball was built from this
4
4
  repository, not from a laptop zip.
5
5
 
6
- This is **not** a published release. The workflow
7
- [`.github/workflows/publish.yml`](../../.github/workflows/publish.yml) is
8
- dispatch-only. It does not run on push. The first publish is a human decision.
6
+ Releases are published from
7
+ [`.github/workflows/publish.yml`](../../.github/workflows/publish.yml), which is
8
+ dispatch-only — it does not run on push, and each release is a human decision.
9
+ Published versions carry a provenance attestation; the commands below are how
10
+ you check that for yourself rather than taking this page's word for it.
9
11
 
10
12
  ## What provenance is
11
13
 
12
14
  When `npm publish --provenance` runs on GitHub Actions with `id-token: write`,
13
- npm stores a signed attestation: package version ↔ git commit ↔ workflow.
15
+ npm stores a signed attestation binding package version ↔ git commit ↔ workflow.
16
+ The attestation lives in the npm registry and in the public transparency log,
17
+ which is what makes it checkable without us.
14
18
 
15
- ## How to verify (after a provenance-backed release exists)
19
+ ## How to verify
20
+
21
+ Ask the registry what it holds for a version:
16
22
 
17
23
  ```bash
18
- npm view @conarium-ai/core@<version> --json
24
+ npm view @conarium-ai/core@<version> dist --json
25
+ ```
26
+
27
+ A provenance-backed version answers with an `attestations` block naming
28
+ `https://slsa.dev/provenance/v1`, plus a registry `signatures` entry. Measured
29
+ against 0.2.32:
30
+
31
+ ```json
32
+ "attestations": {
33
+ "url": "https://registry.npmjs.org/-/npm/v1/attestations/@conarium-ai%2fcore@0.2.32",
34
+ "provenance": { "predicateType": "https://slsa.dev/provenance/v1" }
35
+ }
19
36
  ```
20
37
 
21
- Look for attestations / provenance on that version. Then:
38
+ To have npm check the signatures and attestations of everything you installed,
39
+ including this package:
22
40
 
23
41
  ```bash
24
- npx npm-provenance-check @conarium-ai/core@<version>
42
+ npm i @conarium-ai/core
43
+ npm audit signatures
25
44
  ```
26
45
 
27
- or, with GitHub CLI, verify the attestation against this repo:
46
+ That prints how many packages have verified registry signatures and how many
47
+ have verified attestations, and exits non-zero if a signature fails.
48
+
49
+ ⛔ **`gh attestation verify` does not work here, and the reason is worth
50
+ knowing.** That command reads GitHub's own attestation store, which is populated
51
+ by `actions/attest-build-provenance`. `npm publish --provenance` writes to the
52
+ npm registry and the transparency log instead, so GitHub answers `404` for a
53
+ tarball published this way — for both `--owner` and `--repo`. A 404 there is not
54
+ evidence of a missing attestation; it means you asked the wrong service. Use the
55
+ two commands above.
56
+
57
+ If a version has **no** attestation, it was not published through this workflow
58
+ (or predates it). Treat that as undeclared origin — not as "verified".
59
+
60
+ ## What a release page carries
61
+
62
+ The publish workflow is configured to attach three files to the GitHub release.
63
+ Which of them a given version actually carries is answered by its release page,
64
+ not by this table:
65
+
66
+ | Asset | What it is |
67
+ |---|---|
68
+ | `conarium-ai-core-<version>.tgz` | the published tarball, **downloaded from the registry**, not rebuilt |
69
+ | `conarium-ai-core-<version>.cdx.json` | CycloneDX bill of materials, read from the lockfile the release was built against |
70
+ | `conarium-ai-core-<version>.tgz.intoto.jsonl` | a build attestation over that tarball, also written to GitHub's attestation store |
71
+
72
+ The tarball is fetched rather than packed again on purpose. `npm pack` is not
73
+ byte-reproducible across environments — file modes, line endings and the gzip
74
+ implementation differ — so a rebuilt asset can be a lookalike of the release
75
+ rather than the release. Its sha512 is checked against the registry's
76
+ `dist.integrity` before anything is attested or uploaded.
77
+
78
+ **What has actually run.** The artefact path first executed on 19 August 2026 in
79
+ the workflow's `artefacts` mode — a dispatch that attaches files to a version
80
+ already on npm and skips publishing, tagging and the registry entirely. It was
81
+ run against **0.2.33**, whose release page had been left empty when that version
82
+ shipped. So 0.2.33 carries the tarball and the attestation and **no bill of
83
+ materials**: that step is skipped in artefacts mode, because the lockfile in the
84
+ tree belongs to a later version. The full publish path, producing all three, has
85
+ not run yet.
86
+
87
+ ⛔ **The `gh` command below works only for versions whose release page actually
88
+ carries the `.intoto.jsonl` asset** — check the page, do not assume from the
89
+ version number. That attestation is a second one, written to GitHub's store; it
90
+ does not replace npm's. For a version without it, the 404 explanation above still
91
+ applies.
28
92
 
29
93
  ```bash
30
- gh attestation verify --owner dogrucanemek-alt <downloaded.tgz>
94
+ # only for a version whose release page lists the .intoto.jsonl asset
95
+ gh attestation verify conarium-ai-core-<version>.tgz --repo dogrucanemek-alt/conarium
31
96
  ```
32
97
 
33
- If a version has **no** attestation, it was not published through this
34
- workflow (or was published before the workflow existed). Treat that as
35
- undeclared origin — not as "verified".
98
+ Two attestations over the same bytes are not redundancy for its own sake: one is
99
+ checkable without GitHub, the other without npm, and the bill of materials is
100
+ checkable without either.
36
101
 
37
- ## One-time setup before the first publish
102
+ ## Publisher configuration
38
103
 
39
104
  Auth is **npm Trusted Publishing (OIDC)**, not a stored token. npm hands the
40
105
  workflow a short-lived credential in exchange for its GitHub identity, so there
41
106
  is no long-lived publish secret in this repository to leak.
42
107
 
43
- On npmjs.com, as the package owner:
108
+ The trusted publisher is configured on npmjs.com against
109
+ `dogrucanemek-alt/conarium` and `publish.yml`. To reproduce it on a fork:
44
110
 
45
111
  1. `@conarium-ai/core` → **Settings** → **Trusted Publisher**
46
112
  2. Publisher: **GitHub Actions**
47
- 3. Repository: `dogrucanemek-alt/conarium`
113
+ 3. Repository: `<owner>/<repo>`
48
114
  4. Workflow file: `publish.yml`
49
115
 
50
- Then publish: GitHub → **Actions** → **publish** → *Run workflow* →
51
- type `publish` in the confirm box. The workflow runs lint, tests and checks
52
- before it publishes; a red gate stops the release.
116
+ Releasing: GitHub → **Actions** → **publish** → *Run workflow* → type `publish`
117
+ in the confirm box. The workflow runs lint, tests and the adversarial checks
118
+ before it publishes, and a claims review record must exist for the version being
119
+ shipped; a red gate stops the release.
53
120
 
54
- If npm rejects the OIDC exchange, do **not** fall back to adding an
55
- `NPM_TOKEN` secret. Publish by hand from a workstation for that release and fix
56
- the trusted publisher configuration afterwards — provenance is worth less than
57
- a leaked publish credential costs.
121
+ If npm rejects the OIDC exchange, do **not** fall back to adding an `NPM_TOKEN`
122
+ secret. Publish by hand from a workstation for that release and fix the trusted
123
+ publisher configuration afterwards — provenance is worth less than a leaked
124
+ publish credential costs.
58
125
 
59
126
  ## What this does not prove
60
127
 
@@ -62,3 +129,5 @@ a leaked publish credential costs.
62
129
  - Not that the commit is free of bugs.
63
130
  - Not that the operator's install matches the tarball (`npm ci` + lockfile
64
131
  does that locally).
132
+ - Not that the code does what its documentation says. Provenance binds an
133
+ artefact to a build, and stops there.
package/package.json CHANGED
@@ -1,6 +1,6 @@
1
1
  {
2
2
  "name": "@conarium-ai/core",
3
- "version": "0.2.32",
3
+ "version": "0.2.34",
4
4
  "mcpName": "io.github.dogrucanemek-alt/conarium",
5
5
  "//files-artefakt": "examples/ altindaki uretilmis dosyalar (ozel anahtar, makbuz/denetim JSONL) ELENIR. .gitignore npm'i baglamaz: izin listesinde 'examples' oldugu icin demoyu kosturmus bir makinede 'npm publish' _keys/audit-ed25519.pem'i (OZEL ANAHTAR) ve makbuzlari yayinliyordu. test/pack_artefakt.mjs bunu kilitler. Elemeler examples/ ile SINIRLI: test-vectors/ altindaki receipts.jsonl ve .pub.pem YAYINLANMAK ZORUNDA, yoksa ucuncu taraf konformans vektorlerini kosturamaz.",
6
6
  "//files": "Pakete NE gireceginin izin listesi. Liste olmadan npm her seyi gonderiyordu: 296 dosya / 1.5 MB — CI workflowlari, gitleaks ayari, 137 KB pazarlama gorseli, sitenin api/chat.js dosyasi. Bir bankanin guvenlik ekibi tarball'i acip bunlari gorur. scripts/ BILEREK var: conarium-reconcile calisirken scripts/pg-snapshot.sql'e atif yapiyor, olmazsa kullanici destek maili yazar.",