# VERIFY — pgp-artifacts (adversarial)

Verifier posture: skeptical, tried to REFUTE. Target:
`<research-workspace>/artifacts/pgp-artifacts--MERGED.md`
(plus `--A.md`, `--B.md`). Verdict: **AGREES = false** — the reconstruction is
cryptographically sound and essentially flawless on all Tengri artifacts, but it contains
**one confirmed factual error** (a wrong signature-creation timestamp on the ruled-out foreign key
`iJNtBJHC`) plus one minor fidelity caveat (line-ending claim). Both are listed precisely below.

Scratch/scripts: `verify/scratch/pgp/` (decode.py = independent OpenPGP packet parser + CRC24;
extract.py; imported keyring in `gpghome/`; live pastebin re-fetches in `/tmp/pbtest`).

---

## BOTTOM LINE

I attacked this from every offline and online angle I could. The armored data is **self-authenticating
and not fabricated**: every one of the 20 armored blocks has a valid CRC24, every packet decodes to
sensible values, every claimed timestamp matches the packet bytes, every GOOD/BAD verdict reproduces
under `gpg`, the fingerprint recomputes from a fresh live keyserver fetch, and all 12 live pastebins
are character-identical to the reconstruction. You cannot forge this by memory — a bad character would
break a CRC or a signature. **The Tengri-artifact portion of the reconstruction is confirmed.**

The single real defect is on a **non-Tengri, explicitly ruled-out** paste, so it does not touch any
authenticity conclusion — but the task says flag ANY discrepancy, so `agrees=false`.

---

## DISCREPANCIES (precise)

### D1 — CONFIRMED ERROR: `iJNtBJHC` signature time is wrong by +40 minutes
- Reconstruction states: **`2019-12-31 12:37:42 UTC`**
  - `pgp-artifacts--B.md` line 234 (master table) and lines 792–793 (§5c prose): "created 2019-12-31
    12:37:42 UTC"; §7 line 853 attributes the verdict to "packet decode".
  - `pgp-artifacts--MERGED.md` §7 line 960: "created 2019-12-31 12:37:42 UTC".
- Actual packet (authoritative): **`2019-12-31 13:17:42 UTC`**.
  - `gpg --list-packets` on the live paste: `version 4, created 1577798262`.
  - unix `1577798262` → `2019-12-31 13:17:42 UTC` (confirmed 3 ways: python `utcfromtimestamp`,
    BSD `date -r 1577798262`, and reverse `calendar.timegm`). `12:37:42` would be unix `1577795862`
    — a different value.
- Not a paste-edit artifact: the live paste's checksum is `=ZIkc` (crc_ok=True), byte-identical to
  what B recorded, so B/MERGED simply **misread the timestamp**. The +40-min offset is not a whole-hour
  timezone slip; it is a plain decode/transcription error, propagated A→(B)→MERGED.
- Impact: **none on authenticity.** `iJNtBJHC` is a foreign key (`20DD313C2A292DE8`, fpr
  `903F62CAC41267989E5BB6E920DD313C2A292DE8`), correctly ruled out as NOT Tengri. Only the stated
  time is wrong. (Everything else about this paste checks: issuer key id, issuer fpr, `=ZIkc`, RSA,
  SHA256, sigclass 0x01 all verified against the live packet.)

### D2 — MINOR: "byte-exact" line-ending claim overstated
- `pgp-artifacts--A.md` header claims message bodies were "re-fetched live from `pastebin.com/raw/<id>`
  … (byte-exact)", and the record calls the blocks "byte-identical to sources."
- The live pastebins are served with **CRLF** line endings; the reconstruction `.md` files store **LF**
  (verified: `MERGED.md` has no CRLF). So the stored copies are character-identical to the pastes but
  **not literally byte-identical to the served bytes** (line-ending encoding differs).
- Impact: immaterial to signature validity (clearsigned text is canonicalized before hashing, which is
  why the LF-stored copies still verify GOOD) and to the A-vs-B byte-identity claim (both store LF).
  Flagged only because the task weights verbatim/byte fidelity heavily and the word "byte-exact" is
  slightly stronger than what is true against the wire format.

---

## WHAT I VERIFIED AND CONFIRMED (the reconstruction is right about all of this)

### Cryptographic self-consistency (offline, my own OpenPGP parser)
- **All 20 armored blocks: CRC24 valid**, all packets decode cleanly. No smoothed-over/garbled base64.
- **Fingerprint recomputed from packet bytes** (SHA-1 over tag-6 primary-key packet) =
  `2A2C0AD4ED1CAD8353A30056D152D6C5666AB731` from all three key copies (=KdDd, =DI8b, =dxDi). Matches.
- **Key created** unix `1461495911` = `2016-04-24 11:05:11 UTC`. Matches.
- **Every Tengri signature timestamp matches its packet**, including the two the merge "corrected":
  - `P59Kf0cs`/`=/Vbg` → packet `1491162480` = **2017-04-02 19:48:00 UTC** (B right; A's 20:48:00 wrong). ✔
  - `5kQsJtRE`/`=MLam` → packet `1491131488` = **2017-04-02 11:11:28 UTC** (A's 12:11:28 wrong). ✔
  - All 14 GOOD + self-sigs (2016-05-02 ×2, 2019-09-10 ×2) + 4 third-party certs
    (040484988228DD98 2017-05-03, A49BDCD5BC224010 2017-05-07, D2FBEC71DB65A36D 2017-05-27,
    6966D8E21ACC064F 2017-11-23 w/ issuer-fpr `75711F66DFAB3662FBD290166966D8E21ACC064F`) all match.

### gpg verification (isolated keyring, imported =KdDd key)
- **12 clearsigned messages → Good signature**: fvRGu37W, X0pJtqfD, a1nqC79N, FAJNLpLZ, P59Kf0cs,
  d2qAQgBD, GJ7Gqhws, YmrA3rDS, 98DZzeYr, U7t25SVL, 137.txt(=at0e), tengri137.mp3.txt(=QWxa).
- **2 detached sigs → Good signature** against the actual binaries in `harvest/`:
  - `137.tar.gz.asc` vs `137.tar.gz` (sha256 `d109…3aa8`, 2,701,349 bytes) ✔
  - `tengri137.mp3.asc` vs `tengri137.mp3` (sha256 `b424…34d7`, 7,864,927 bytes) ✔
  - Both claimed hashes and both sizes reproduce exactly.
- **2 BAD artifacts → BAD signature**: `5kQsJtRE`(=MLam) and `CJsN0KWL`(=Ermz). Both reproduce BAD,
  on the reconstruction's stored copies AND on my independent live re-fetch.

### A vs B merge integrity (independent CRC-identity diff)
- **12 armored blocks shared between A and B are byte-identical**: =KdDd, =DI8b, =vL0l, =b2/C, =kYYv,
  =m2lf, =/Vbg, =Ermz, =Nw3q, =r1U0, =Th85, =60tK. (Exactly the merge's headline claim.)
- **Every MERGED armored payload traces to an A or B source; zero orphans.** A-only: =MLam, =dxDi.
  B-only: =3QH4, =KikD, =QWxa, =at0e. Matches the "complementary coverage" narrative (V3/V4/V5).

### Independent sources not in the reconstruction's provenance
- **Fresh live keyserver.ubuntu.com fetch (2026):** gpg import → fpr `2A2C…666AB731`, created
  `1461495911`; its invariant primary-key packet is **byte-identical** (525 bytes) to the
  reconstruction's copy. (The full served body differs only by keyserver re-ordering of the 4
  third-party sig packets — exactly as B predicted; same length 5821.)
- **bartman081523/tengri137-selenoprotein GitHub repo** (mentioned in B prose, NOT in provenance_urls):
  its `pgp_verification.txt` independently records `137.tar.gz.asc` "Signatur vom Do 18 Aug 2016
  20:51:18 CEST … RSA-Schlüssel D152D6C5666AB731" = **2016-08-18 18:51:18 UTC**, corroborating
  detached-sig #11's time and issuer.
- **All 12 live pastebins re-fetched → character-identical to the reconstruction** (normalized for
  CRLF/LF); the two gate files `BcMucDVh` and `u7pFiCpD` return **HTTP 404**, as claimed.

### Internal consistency / recomputed decodings
- Pi-Day matrix `FAJNLpLZ`: exactly **11 rows × 14 cols = 154 letters**; row 1 = `BURUMUTREFAMTU`,
  reversed `UTMAFERTUMURUB`. ✔
- BAD excerpt `5kQsJtRE`: exactly **404 letters**, and equals the **first 404 letters of
  themessage.txt** (head `MYVZZJCKTBWKOWUW…`). ✔
- `themessage.txt`: exactly **6448 letters**; tail `…QXUNYKNBMIHGDQHFZEEQJNBBLGAJDGRLUZMVKAWK`. ✔
  (6448 ≠ 6439 = 47×137, as the background warned.)
- `nothingisrandom.txt` digit string: exactly **1422 digits**. ✔
- `iA558C4F` is a **byte-identical duplicate of `P59Kf0cs`** (incl. line endings) — dup claim confirmed.
- Foreign key `khQff3xV`: issuer `950F7FD7DE0923E6`, `2011-04-18 01:37:37 UTC`, GnuPG v1.4.10, SHA1,
  `=TWIF` — all confirmed against the live packet; correctly ruled out.

---

## GAPS I could NOT independently close (agree with reconstruction's own §10)
- No surviving VALID signature over `themessage.txt` / `nothingisrandom.txt`: `BcMucDVh` and `u7pFiCpD`
  are dead 404 (re-confirmed live); the only signed excerpt (`5kQsJtRE`) is BAD. "ATTRIBUTED, not
  verified" is the correct grade. The GOOD-signature claims of contemporary readers remain SECONDARY.
- The 4 third-party certifier issuer keys, the onion-served originals, imgur album `tdYJJet`, and the
  Dropbox folders were not fetched (out of scope; agreed).
- Discord-era claims (Eternal Library / KIWI / GARLIC / IPv6 / `rjzdqt4z3z3xo73h.onion` /
  `tengri137@torbox3…`) have no signed artifact — correctly graded SPECULATION.

## Provenance tier check
PRIMARY label is **honest**: author's own key on public keyservers, contemporaneous pastebin
originals (still live), PGP-signed artifacts re-verified cryptographically this pass. Nothing PRIMARY
here is laundered folklore. The one folklore-adjacent numeric error (D1) is on a foreign key, not a
Tengri PRIMARY artifact.