# TgWitness Test Results & Qualification Record

Canonical page: https://phishdestroy.io/tgwitness/docs/reference/verification/

Read which TgWitness workflows were tested, on which revisions and environments, with synthetic checks and live Telegram results listed separately.

[Project context and measured scale](https://phishdestroy.io/tgwitness/docs/reference/overview/)

[Documentation index](https://phishdestroy.io/tgwitness/llms.txt)

The live Telegram test, automated checks, browser checks and the next qualification runs.

[TgWitness](https://github.com/phishdestroy/TgWitness/blob/main/README.md) / [Documentation](https://github.com/phishdestroy/TgWitness/blob/main/docs/README.md) · [Evidence and verification](https://phishdestroy.io/tgwitness/docs/reference/evidence/)

![Evidence & verification](https://phishdestroy.io/tgwitness/assets/chapter-evidence.svg?v=e49c65feab6c)

**Updated 10 November 2026.** Results are tied to the revision, environment and workload recorded for each run. [Current capabilities](https://github.com/phishdestroy/TgWitness/blob/main/docs/STATUS.md) describes the available workflows; this page records what has actually been checked.

These are selected validation results from the [development history](https://phishdestroy.io/tgwitness/docs/reference/overview/#testing-history-and-duration). Engineering tests, live acquisition, fault injection, recovery checks and artifact verification each have their stated scope.

On 8 November, the published installer at `b35f8e18f3d1` was inspected without executing or installing it: wheel and dependency-lock SHA-256 and RSA signatures passed, and the wheel contained the Secret Chat backend, interface, API routes, collector hooks and optional dependency declaration. See the [package check](https://phishdestroy.io/tgwitness/docs/reference/secret-chats/#package-availability). [CI for that revision](https://github.com/phishdestroy/TgWitness/actions/runs/42446658989) and the [installer run](https://github.com/phishdestroy/TgWitness/actions/runs/42446659252) completed successfully. Older audit reports retain the findings for their original snapshots; current implementation and qualification are recorded separately.

[Performance graphs](https://phishdestroy.io/tgwitness/docs/validation/) · [Signed English audit](https://github.com/phishdestroy/TgWitness/blob/main/website/validation/audit/REPORT.md) · [Downloadable measurements](https://github.com/phishdestroy/TgWitness/blob/main/website/validation/audit/MEASUREMENTS.csv)

| Latest evidence | Result | Scope |
| --- | --- | --- |
| Account security and exports · 10 Nov | Witness probe caught a second client on the same key; killed chat exports resumed; TDX quote checked offline | RunPod pod, three test accounts; revisions `6c7171228` to `196a5bcd1` |
| Phala / Intel TDX with live account · 9 Nov | All 6 TDX checks passed; live Telegram collection; export verified | CVM `cvm_mjbvBdjM`; commit `427621d4`; session imported; self-attested tier |
| drand checkpoint + verify · 9 Nov | drand round 33818717 in a signed checkpoint, confirmed via Cloudflare relay | RunPod collector; `--check-drand`; lower time bound for the checkpoint |
| session.fingerprint in journal · 9 Nov | `auth_key_id` logged at `session.fingerprint` on observation start | Verifier can match session file after revocation |
| Release audit · 6 Nov | 1,461 core tests passed; 37 platform skips | Source `92e539af40d8`; open findings retained |
| Fresh exports · 6 Nov | 127.713 and 141.781 records/s, complete cycle | Two accounts per server; text and metadata through WARP |
| Retained archives · 6 Nov | 48 archives; 3,598,381 records verified | Existing archives checked against pinned workspace keys |
| Trusted Live · 5 Nov | Guarded CPU recording and WARP checks | Official Desktop login and native export remain unqualified |
| Phala / Intel TDX · 4 Nov | All 16 verifier checks passed on later runs | Historical deployment; no Telegram account on the CVM |

**On this page**

- [Release audit and fresh exports · 6 November 2026](https://phishdestroy.io/tgwitness/docs/reference/verification/#release-audit-and-fresh-exports--6-november-2026)
- [Trusted Live CPU Ubuntu and WARP · 5 November 2026](https://phishdestroy.io/tgwitness/docs/reference/verification/#trusted-live-cpu-ubuntu-and-warp--5-november-2026)
- [Live Phala Cloud deployment · 4 November 2026](https://phishdestroy.io/tgwitness/docs/reference/verification/#live-phala-cloud-deployment--4-november-2026)
- [Live server journal inventory · 1 November 2026](https://phishdestroy.io/tgwitness/docs/reference/verification/#live-server-journal-inventory--1-november-2026)
- [Recent offline and fault-injection checks · 1 November 2026](https://phishdestroy.io/tgwitness/docs/reference/verification/#recent-offline-and-fault-injection-checks--1-november-2026)
- [Live Telegram test, whole account · 29 October 2026](https://phishdestroy.io/tgwitness/docs/reference/verification/#live-telegram-test-whole-account--29-october-2026)
- [Password-protected export opened with libarchive · 29 October 2026](https://phishdestroy.io/tgwitness/docs/reference/verification/#password-protected-export-opened-with-libarchive--29-october-2026)
- [Live Telegram test · 27 to 29 October 2026](https://phishdestroy.io/tgwitness/docs/reference/verification/#live-telegram-test--27-to-29-october-2026)
- [Synthetic local capacity measurement](https://phishdestroy.io/tgwitness/docs/reference/verification/#synthetic-local-capacity-measurement)
- [Full-text search on one million messages · 5 November 2026](https://phishdestroy.io/tgwitness/docs/reference/verification/#full-text-search-on-one-million-messages--5-november-2026)
- [Phala Cloud with live Telegram account · 9 November 2026](https://phishdestroy.io/tgwitness/docs/reference/verification/#phala-cloud-with-live-telegram-account--9-november-2026)
- [drand checkpoint anchoring and verify · 9 November 2026](https://phishdestroy.io/tgwitness/docs/reference/verification/#drand-checkpoint-anchoring-and-verify--9-november-2026)
- [Parallel session and session integrity tests · 9 November 2026](https://phishdestroy.io/tgwitness/docs/reference/verification/#parallel-session-and-session-integrity-tests--9-november-2026)
- [Session key created inside the TDX VM · 9 November 2026](https://phishdestroy.io/tgwitness/docs/reference/verification/#session-key-created-inside-the-tdx-vm--9-november-2026)
- [Account security and interrupted exports · 10 November 2026](https://phishdestroy.io/tgwitness/docs/reference/verification/#account-security-and-interrupted-exports--10-november-2026)
- [Qualification status](https://phishdestroy.io/tgwitness/docs/reference/verification/#qualification-status)

## Phala Cloud with live Telegram account · 9 November 2026

A fresh Phala Cloud deployment (`phala-live-6`) ran commit `427621d4` with a live Telegram account. The collector imported an existing session (`key_origin: imported`) and observed three channels for approximately three minutes. The panel exported a password-protected ZIP archive; `tgwitness verify` confirmed its integrity.

| Setup | Recorded value |
| --- | --- |
| Provider | Phala Cloud · `tdx.small` · US-WEST-1 |
| CVM | `cvm_mjbvBdjM` · app `408d07581ad77680b4285b041f5924e52464437a` |
| Commit | `427621d4740f267e834c5cf754bc1007d6d01750` · v0.9.1 |
| Collector fingerprint | `fcdbda2d3d69b43937a4883102e579bbd51a5915a1c85baa4ce154417ec7dd0f` |
| Intel collateral issued | 2026-11-09T20:40:09Z |
| TCB level | 2025-08-13T00:00:00Z · status UpToDate |
| dstack OS image | `dstack-0.5.12` · hash `8db40cb23008d6213668b5816702d100b37b859a8a5bfd18d938ae3a1b3715bd` |
| Account | dedicated test account · key_origin `imported` |
| Targets | three channels |
| Journal events | 389 |
| Export | Password-protected ZIP · 2.9 MB |
| Report formats | HTML + JSON + PDF (certificate `TgWitness workspace fcdbda2d3d69b439`) |
| source_tree_sha256 | `8c228874c7cfdd57ffac0184e5f0c12245ecaa0571e65ec84dee5ce08d0cbc6c` |

**Intel TDX attestation — all 6 checks passed:**

| Check | Result |
| --- | --- |
| Quote covers signed statement and challenge | PASS |
| Intel vouches: genuine TDX, TCB UpToDate, v4 quote, not debug, not migratable | PASS |
| RTMR3 measures the app-compose the collector reports | PASS |
| Measured compose equals the deployed commit's compose byte for byte | PASS |
| Quote comes from the deployed CVM with KMS keys and encrypted disk | PASS |
| MRTD and RTMR0-2 equal values computed from the published dstack-0.5.12 image | PASS |

**Verify output (excerpt):**

```
events: 389  files_checked: 35  sha256sums: 34
Report seal verified: report.html, language en.
JSON report seal verified: report.json.
PDF signature and Ed25519 seal verified: report.pdf
  certificate: TgWitness workspace fcdbda2d3d69b439
SHA256SUMS matches the signed manifest.
Chat channel-A: IDs 1-32, 8 not observed in 7 ranges
Chat channel-B: IDs 1-210, 69 not observed in 35 ranges
Chat channel-C: IDs 1-122, 2 not observed in 2 ranges
```

Assurance tier is `self_attested`: the session was imported and the archive carried no external timestamp. Signing in by QR code or phone code through the panel creates the Telegram session key inside the TDX VM (`key_origin: created`).

**Gap analysis summary:**

| Chat | ID range | Observed | Missing | Ranges |
| --- | --- | --- | --- | --- |
| channel-B | 1–210 | 141 | 69 | 35 |
| channel-A | 1–32 | 24 | 8 | 7 |
| channel-C | 1–122 | 120 | 2 | 2 |

Gaps are message IDs Telegram assigned that were never visible to this account. A gap records what the collector could not observe; attribution and cause are outside what the collector can determine.

## drand checkpoint anchoring and verify · 9 November 2026

A RunPod collector connected a dedicated test account. The checkpoint API returned a checkpoint signed with the collector key that embeds a live drand round from Cloudflare's relay; `--check-drand` confirmed the round against the network. A drand round is a lower time bound: the checkpoint was signed no earlier than that round. Existence by a time comes from RFC 3161, Bitcoin or a witness, so the tier stays as those checks set it.

| Item | Value |
| --- | --- |
| Collector | RunPod CPU instance · session imported · key_origin `imported` |
| Journal events | 387 |
| drand chain | `52db9ba70e0cc0f6eaf7803dd07447a1f5477735fd3f661792ba94600c84e971` |
| drand rounds in checkpoints | 33818717 (not before 2026-11-09T21:25:15Z) and 33819014 (not before 21:40:06Z) |
| session.fingerprint logged | `auth_key_id` logged at seq 18 |
| coverage.gap logged | seq 28 · reason `updates delivered while observation was paused` |
| Archive fingerprint | `f304aa2afd49a5f0b7a064bfb01167151387797b4723ceeb90357458a8f35673` |

**Code changes verified in this run:**

| File | Change | Verified |
| --- | --- | --- |
| `tgwitness/evidence.py` | `checkpoint()` embeds `drand_anchor` from live Cloudflare relay | ✓ round 33818717 in checkpoint |
| `tgwitness/verifier/journal.py` | `verify_checkpoint(check_drand=True)` confirms the round against relays | ✓ raises error on mismatch |
| `tgwitness/collector.py` | `session.fingerprint` event logs `auth_key_id` on observation start | ✓ seq 18 in journal |
| `tgwitness/store.py` | `session.*` added to `coverage` kind group | ✓ fingerprint visible in coverage |

## Release audit and fresh exports · 6 November 2026

The audit used v0.9.1 at source `92e539af40d8368d198ff5e040f1d876f2c1310e`, release `92e539af40d8-83e41104669e`. Installed-package smoke checks passed on three servers. The signed report retains the state observed at that time.

[![Acquisition and complete export rates for two fresh workloads](https://phishdestroy.io/tgwitness/assets/performance.svg?v=80440a98d0ea)](https://phishdestroy.io/tgwitness/docs/validation/)

| Workload | Records | Acquisition | Complete cycle | Complete-cycle rate |
| --- | --- | --- | --- | --- |
| Server A · two accounts | 27,030 | 189.043 s | 211.647 s | 127.713 records/s |
| Server B · two accounts | 29,299 | 182.513 s | 206.650 s | 141.781 records/s |

Both exports began without resume, excluded attachments and avatars, and ended with zero duplicate or reused records. The complete cycle includes preparation, acquisition, selection, packaging and archive verification. Server A also ran a protected-chat read check. Four FloodWait events on A and three on B were respected.

| Check | Result | Qualification |
| --- | --- | --- |
| Core pytest | 1,461 passed | 37 platform skips |
| Browser / browser UX / evidence / site | 102 / 66 / 15 / 9 passed | These programs overlap other tests; counts are not additive |
| Capture addon | 124 passed | Three Official Web checks blocked by the required Chromium sandbox |
| API authorization | 269 routes; 538 requests passed | Missing credentials and revoked access |
| Retained archives | 48 of 48 verified | 3,598,381 records; pinned workspace keys |
| Live fleet read checks | 24 of 40 profiles passed | 16 WARP routes failed qualification; those profiles were not connected |

The [signed report](https://github.com/phishdestroy/TgWitness/blob/main/website/validation/audit/REPORT.md) records all open findings, including three defects in an unpublished secret-chat snapshot absent from the audited main build. The earlier [public-history qualification](https://github.com/phishdestroy/TgWitness/blob/main/docs/EXPORT-BENCHMARK-2026-11-06.md) used a different revision and workload; its rates must not be combined with this comparison.

Verify the original report using the [pinned key and verifier](https://github.com/phishdestroy/TgWitness/blob/main/website/validation/audit/README.md#verify-the-original-files). Its signature protects the retained report from alteration. The report identifies the tested revision, workload, environment and verification results for that run.

## Trusted Live CPU Ubuntu and WARP · 5 November 2026

The retained integration run used a CPU-only Ubuntu 24.04.5 guest (4 vCPU, 8 GiB) under QEMU inside a RunPod CPU host. The collector used `14e642a5`; verification used `02072461`. The final fixes merged in [PR #114](https://github.com/phishdestroy/TgWitness/pull/114) as `755f78f6`, whose tree matched the tested PR head. This was not a TEE.

- A real WARP Connector reported healthy; proxied HTTPS reported WARP and Gateway enabled. Three probes saw the same egress address.
- Disconnecting WARP made the proxied request fail. Direct gateway-service HTTPS was rejected. In the recorder namespace, direct IPv4/IPv6, external DNS and Docker's embedded UDP resolver were blocked.
- A scoped read-only provider token could read health/connections and could not read the connector enrollment token. Signed provider observations included explicit unavailable samples. The witness ran on the operator's host; the test covers recording and retention.
- Official Desktop reached its unauthenticated welcome screen. Guarded keyboard input was rejected during capture. Three H.264 display segments decoded and their clock overlay was inspected. The 18-event preflight archive verified against the pinned recorder key and retained checkpoint.
- The final capture tests passed 114 cases on Windows and Linux CI, with real Chromium/FFmpeg. Five anchor tests and dependency advisory checks passed.

An initial 2 vCPU/4 GiB startup timed out under concurrent image-build load and stopped through the guard; it is not counted as success. Manual test-account login, native chat export, real hybrid comparison and Telegram-side logout remain pending. No supplied Telethon session was imported. Attachments, automatic OCR, local drand BLS verification and Trusted Live TDX attestation were not qualified.

## Live Phala Cloud deployment · 4 November 2026

**Historical result:** later deployments passed all 16 verifier checks, including Intel TDX and the quick tunnel. Restart evidence, signed address discovery and report timestamps were checked separately. Each test CVM was removed after its run.

| Setup | Recorded value |
| --- | --- |
| Provider | Phala Cloud · `tdx.small` · US-WEST-1 |
| Deployment | `--provider phala --quick-tunnel`; initially empty workspace |
| Initial collector | `df7d860c` · v0.8.1 |
| Collector with all fixes | `53a0f7e9` |
| Production guest | `dstack-0.5.12`; SSH access disabled |
| External services | ntfy.sh, GitHub and Intel PCS |
| Observed cost | Seven CVMs, about 1.9 US cents together; historical test pricing |

The sequence below separates initial failures, fixes and completed checks. No Telegram account was connected to these CVMs.

**Deployments 1–2 · provider differences and regression checks**

The first two deployments found five differences from the simulated Phala Cloud. Private Phala regressions reproduce each difference and fail without its fix:

| On Phala Cloud | Before the fix | Fix |
| --- | --- | --- |
| `compose_file.name` is required | HTTP 422, nothing created | The compose carries the CVM name |
| Unless given an empty one, Phala adds its own pre-launch script to the measured app-compose (v0.0.20: Docker login, `docker compose pull`, root password, SSH keys from `user_config`) | `tdx_compose` fails: code outside the published compose | The app-compose sets `pre_launch_script` to `""`; Phala then leaves it out |
| Without `image` Phala picks `dstack-dev-0.5.12`, the development build, with SSH access enabled | `server_state` fails. The quote alone passed on this image; only the provider record and `tdx_os` show it | The deployment asks for the production build of Phala's current release, `dstack-0.5.12` |
| With the dstack gateway off, the CVM details never get an `instance_id` | Deploy waits 30 minutes and fails | The instance ID is read from the CVM state |
| Restarts, stops and starts are CVM operations, not application revisions | `provider_actions` did not list restarts | The operations are read as well |

**Deployment 3 · attestation, access checks and restart evidence**

- The CVM ran `dstack-0.5.12` with SSH access off. The collector posted its signed address about 1.5 minutes after creation. Deploy then passed `tdx_binding`, `tdx_quote` (Intel TCB UpToDate, quote v4, FMSPC `20a06f000000`, not in debug mode, not migratable, collateral fetched from Intel), `tdx_compose_hash`, `tdx_compose` (byte for byte) and `tdx_identity` (KMS keys).
- The panel at its trycloudflare.com address answered 401 without a token or with a wrong one, and 200 with the panel token.
- `verify` with `--source` set to GitHub's archive of the commit and the fingerprint passed separately: server, `server_state`, firewall (no gateway, no public endpoint), provider log, statement, fingerprint, source hash, boot and the five `tdx_*` checks passed. Cloudflare and `tdx_os` were unavailable, so the result was INCOMPLETE.
- Two restarts through Phala's API: each time the CVM was back in under two minutes with the same collector key, a new signed address, and the old one answering 530. `verify` reported the new boot and both restarts as REVIEW. With `--evidence` it confirmed the journal signatures and three starts, each with its own Intel TDX quote from the same CVM.
- `destroy` removed only that CVM. The workspace was then empty, and the panel address stopped answering.
- The panel wizard's price check read `tdx.small` at $0.058 an hour. A wrong key got a plain "did not accept the API key" message.

**Follow-up fixes · signed address discovery and OS measurements**

- Without the address key, `verify` followed the newest well-formed post on the address topic, which is printed in the published compose. One forged post sent it to a host that does not exist, and `statement` failed: a forger could not pass as the collector, but could block third-party verification, for up to 12 hours with a forged future timestamp. The collector now also posts its address signed with its own key, `verify` accepts only such posts when the collector makes them, and posts are ordered by ntfy.sh's receive time. A private cloud-wizard regression replays the forged posts. A CVM runs this only from a commit that contains it, so this run's collector did not sign its address yet and `verify` said so.
- `tdx_os` could not run without `dstack-mr`, whose current version has `measure-os` rather than the `measure --json` the documentation named. The verifier now computes MRTD and RTMR0-2 itself from the published image and the VM's shape, ACPI tables included.

**Deployment 4 · all 16 checks, report seals and restart verification**

- Deploy downloaded the published `dstack-0.5.12` once, matched it to the OS image hash the CVM reported, and `tdx_os` passed: MRTD, RTMR0, RTMR1 and RTMR2 equal the values computed for QEMU 8.2.2, 1 vCPU and 2 GiB. The same computation gave a different RTMR0 for a claimed 4 GiB, 2 vCPUs, QEMU 9.2.1 or a second network card, and failed for a changed kernel, firmware or command line and for the development image.
- `verify` passed every one of its 16 checks, the quick tunnel's connector included, so the result was PASS. It wrote the HTML and PDF certificate, sealed with a qualified eIDAS time stamp from the Belgian government's time-stamping unit TSU 202301 (status granted on Belgium's trusted list), a FreeTSA time stamp and the verifier's signature; `check-report` confirmed all of them and the PDF's signature.
- After a restart, `verify --evidence` confirmed two starts, each quoted inside the CVM and each booted from the published `dstack-0.5.12`; the restart and the new boot were REVIEW.
- `destroy` left the workspace empty.

**Deployment 5 · collector-signed discovery and forged-post rejection**

The collector ran from commit `53a0f7e9`, which includes the fixes above.

- The collector posted its address signed with its own key, next to the operator's HMAC post. Two forged posts were then added to the topic, one with a zeroed HMAC and one signed by another Ed25519 key, both claiming a time in the year 2096. `verify` took the address signed by the collector key and passed all 16 checks; the seal again carried the qualified Belgian time stamp, FreeTSA's and the verifier's signature, and `check-report` confirmed them.
- After a restart the collector posted the new address signed: `verify` found it at once, the operator's **Find the current address** 27 seconds later from the HMAC post, and both ignored the forged posts. `verify --evidence` confirmed two starts booted from `dstack-0.5.12`.
- Deploy printed what Phala bills the CVM, disk included: $0.06078 an hour.

### Panel wizard and remaining scope

The panel wizard ran against Phala Cloud twice, once through its job API and once through the browser with a real Chromium: the price check, creation with the commit under **Advanced**, all six attestation checks passed, the server list showing the billed price (44.37 USD a month), **Find the current address**, the measured compose download, and deletion. The browser saw no page errors. Each time the workspace was empty afterwards.

Not covered: a Cloudflare domain on Phala, a Telegram account on the CVM, and the `dstack-0.6.1` image. The wizard's price before creation is Phala's compute price only, as its API names the disk price only for an existing CVM; the server list shows the billed price after creation. All prices in this section were observed on 4 November 2026.

## Live server journal inventory · 1 November 2026

At approximately 19:15 UTC, sequential read-only SQLite `query_only` queries counted events in ten workspace databases on a running server. Collection continued during the queries, so this is not an atomic snapshot across workspaces.

| Observed across the ten workspace databases | Count |
| --- | --- |
| Workspaces with at least one `message.new` event | 8 |
| Journal records, including operational events | 1,610,818 |
| `message.new` events | 900,232 |
| `message.edited` events | 80,683 |
| `message.deleted` events | 400,536 |
| `message.recovered` events | 53,403 |
| `message.snapshot` events | 2,183 |
| `media.saved` events | 50,418 |
| `media.failed` events | 5,575 |
| `coverage.gap` events | 11 |
| Database bytes, excluding WAL files and media | 7,909,003,264 |

These are **event counts**, not unique messages, people or media files. The same message can appear in more than one workspace, and an edit or deletion can produce another journal event. `media.failed` counts historical failure events, not the current media queue. The queries did **not** newly verify signatures and the full hash chain of all 1,610,818 records, measure delivery or throughput, or establish that every Telegram event was captured. No message content, account or chat IDs, workspace labels, credentials or private files are published here.

At this review, the separately configured public monitor covered three sources: one capture response was OK and two reported `no_response`. That monitor is not a ten-workspace health summary. `no_response` means the observer obtained no response at that check; message delivery is assessed from the journal. The smaller 29 October test below has its own explicit signature and hash-chain verification.

## Recent offline and fault-injection checks · 1 November 2026

These checks use fictional data and local files; they add regression evidence for the current source, not another live Telegram delivery measurement:

| Check | What it exercised |
| --- | --- |
| Source archive | Two builds with identical source bytes but changed file times produced the same ZIP and checksum. Exact archived bytes and file hashes matched; private test files stayed out. A simulated write failure left the previous archive and checksum intact. |
| Account export recovery | A simulated disk error on the second ZIP member left no published or partial ZIP. Resume kept every signed receipt byte-identical, produced an independently verifiable archive using the expected signing fingerprint, and retained the synthetic coverage gap as a partial result. Scale assertions guard against repeated task and receipt scans as histories grow. |
| Minimized disclosure privacy | Synthetic Unicode names and aliases, nested metadata, unknown senders and media fields were checked in all three identity modes and with text omission. A masking-limit failure did not silently weaken redaction; the original signed account export stayed unchanged. |
| Linux media decoder and OCR | A [Linux Docker CI check](https://github.com/phishdestroy/TgWitness/actions/runs/41341200341/job/124951395192) at commit `5a9e0814` passed generated Ogg/Opus, MP3, M4A, MP4 with video, WAV and FLAC inputs; forged-duration and secondary-file-reference controls; and 48 mutated parser fixtures (34 rejected, 14 decoded within bounds). Real Tesseract OCR ran with original bytes unchanged. The decoder and OCR ran in network-isolated containers without a Whisper model. |

The checks run in the [pull-request](https://github.com/phishdestroy/TgWitness/blob/main/.github/workflows/pr-check.yml) and [full Checks](https://github.com/phishdestroy/TgWitness/blob/main/.github/workflows/checks.yml) workflows. They test these failure boundaries, not every possible storage failure or whether a disclosure is automatically safe to publish. The media check does not measure speech recognition accuracy or Telegram delivery. Review each derivative before sharing.

## Live Telegram test, whole account · 29 October 2026

The same account kept running after the test below. On 29 October at 12:10 UTC its scope was switched to the whole account (channels, groups, private chats and bots); the panel asked for confirmation and the signed journal recorded it. By 17:59 UTC the journal held:

| Recorded, 27 October 21:36 to 29 October 17:59 UTC | Count |
| --- | --- |
| Journal records | 9,968 |
| New messages | 5,308 |
| of them after the switch to the whole account | 3,878 |
| Deletions | 784 |
| of them deleted after TgWitness had saved the message | 596 |
| Edits | 41 |
| Messages added by rechecks | 2,273 |
| Messages a recheck could no longer find | 18 |
| Messages saved by history snapshots | 242 |
| Saved attachments | 247 |
| Collector starts | 19 |
| Telegram FloodWait pauses | 26 |

- The panel's **Verify** check confirmed the signatures and hash chain of all 9,968 records.
- Messages came from 2 groups or channels and 3 private chats and bots. One message button was pressed from **Chats**; the request and Telegram's answer are in the journal.
- New chats were added when the collector rescanned the account at a restart. Saving the earlier messages of new chats was off (`history = 0`), so it is not covered here.

Not covered: reactions, membership changes, people profiles, sending messages, joining by invitation, proxies, notifications, the witness bot and several accounts at once. Public files contain only counts.

## Password-protected export opened with libarchive · 29 October 2026

An archive written by the export encryption code (WinZip AES-256, the format 7-Zip and WinRAR read) was opened with Windows `tar` (bsdtar 3.6.2, libarchive). With the right password the extracted file matched the original SHA-256; a wrong password was refused with "Incorrect passphrase". 7-Zip and WinRAR themselves were not run.

## Live Telegram test · 27 to 29 October 2026

One real Telegram account, signed in through the built-in API pair pool, recorded one supergroup from **27 October 21:36 UTC to 29 October 05:15 UTC**. The collector was started 12 times in that period.

| Recorded | Count |
| --- | --- |
| Journal records, including heartbeats and service records | 2,385 |
| New messages | 1,291 |
| Deletions | 472 |
| of them deleted after TgWitness had saved the text | 347 |
| Edits | 12 |
| Messages missed while stopped, added by the recheck | 12 |
| Saved attachments | 17 |
| Attachments not downloaded (downloads off, or no file to download) | 14 |
| Telegram FloodWait pauses, 3 seconds each, during rechecks | 8 |

- The signatures and the hash chain of all 2,385 records verify against the journal's public key.
- After a restart the collector resumes from the saved update state (`pts`) and Telethon catches up; the recheck then adds messages that are still missing.
- On every FloodWait the collector waited the requested time and continued.
- The account stayed signed in, and the same session reconnected on 29 October.

Not covered by this test: several accounts at once, administrator roles, proxies, outages longer than a few hours and the full [live acceptance procedure](https://github.com/phishdestroy/TgWitness/blob/main/docs/internal/LIVE_ACCEPTANCE.md). Public files contain only these counts: no account or chat identifiers, messages or credentials.

## Synthetic local capacity measurement

Private synthetic capacity run: 10,000 events across 100 chats.

| Metric | Observed |
| --- | --- |
| Input | 10,000 synthetic message events across 100 synthetic selected sources |
| Message text | 400 characters; no media or network |
| Ingestion | 53.273 seconds; 187.7 events/second |
| Handler latency | p50 4.162 ms; p95 7.784 ms; maximum 129.118 ms |
| ZIP export | 19.868 seconds with tracemalloc enabled |
| Peak traced Python export allocations | 0.51 MiB; not process RSS |
| ZIP size | 4,610,962 bytes |
| Independent verification | All 10,000 events verified |

This measures the local serial recording/export path on this Windows host; live delivery, concurrent accounts, Telegram throttling and media are measured in the live runs above. The measurement predates the final offline timestamp tooling, which leaves the measured message path unchanged.

## Full-text search on one million messages · 5 November 2026

Command: `python tools/bench_fulltext.py --messages 1000000 --work <empty directory>`

A deterministic signed journal of 1,000,000 synthetic messages in 200 chats (Russian word forms, Ukrainian, English, Chinese, Japanese, links and handles) was indexed in 71 seconds; the index took 58 MiB beside a 1.50 GiB journal. Best-match searches took 9 to 190 ms (median), against 2.4 to 3.2 seconds for the previous substring search. Full results, hardware and limits: [Full-text search](https://phishdestroy.io/tgwitness/docs/reference/search/#benchmark). Private compatibility tests also verify a frozen signed archive made before full-text search with the current verifier, and a current archive with the frozen archive's own `verify.py`.

## Parallel session and session integrity tests · 9 November 2026

Telethon clients on a RunPod Linux pod, dedicated test accounts, several clients sharing one session.

| Test | Result |
| --- | --- |
| History reads with two and three clients on one session | Every client read all messages: 10/10 and 15/15 |
| `auth.resetAuthorizations` while an observer is connected | Session left Telegram's list in 0.23 s; the observer's open connection kept working for over 60 s and completed 36 history requests |
| Enable, change and disable 2FA while an observer is connected | Observer kept working through the 10 s observed; Telegram 777000 posted each change |
| `auth.resetAuthorizations` from a session under 24 hours old | Rejected with `FreshResetAuthorisationForbiddenError` |
| Message to own Saved Messages as a push probe | 0 of 22 delivered as push; the probe was dropped |
| Two clients on one `auth_key` for 90 s, a request every 2 s | Both kept working; no `AUTH_KEY_DUPLICATED` |
| Import of Telethon SQLite, Pyrogram SQLite, TData ZIP and Account JSON | All four connected; Pyrogram files are converted by TgWitness |
| Requests after `auth.logOut` | `get_me()` returned `None`; reconnect reported not authorized; `GetDialogs` raised `AuthKeyUnregisteredError`. The collector takes the refusal cause from a real request |
| Telegram 777000 during login and 2FA changes | Posted new logins with client and IP, login codes, logins through external services and each 2FA change; no push update reached a parallel session |
| Telegram's session list for an imported session | Shows the client the session was created with, for example Telegram Widgets |
| Leaving a basic group | History stayed readable after leaving |

## Session key created inside the TDX VM · 9 November 2026

A separate run deployed commit `206fef76` to Phala Cloud `tdx.small`; all six attestation checks passed. The panel signed in by phone code inside the VM. A second session read the code from Telegram 777000 and passed it straight to the VM.

| Item | Result |
| --- | --- |
| Journal record 6 `account.authorized` | `key_origin: created` |
| Telegram's date for the new session | 5 min 9 s after the attested collector start, no restart in between |
| History of five sources | 131 messages, 25 attachment jobs |
| External time on two evidence passes | FreeTSA and Sectigo RFC 3161 replies; four OTS calendars pending at the end of the run |
| Second session on the same account | Worked at the same time as the VM session |
| Secret Chats in the VM | Unavailable: the Phala installer installs the hash lock without the `secret-chats` extra |

## Account security and interrupted exports · 10 November 2026

A RunPod Linux pod (8 vCPU, 16 GB) ran the panel on loopback with three workspaces and dedicated test accounts: one imported as an auth key and DC line, one from an Account JSON and one signed in from its own workspace. Revisions `6c7171228`, `4d6206a52` and `196a5bcd1`; the chat exports below were read with observation paused, as the panel requires.

| Check | Result |
| --- | --- |
| Auth key and DC line import | Recognised as **Auth key HEX and DC**; signed in within 4 s |
| Independent sign-in from a workspace | `key_origin: created` after 19 s |
| **Make this the only session** on a key minutes old | Telegram refused (`FRESH_RESET_AUTHORISATION_FORBIDDEN`); the panel named the 24-hour limit and the automatic option stayed on |
| Donor logged out 1–9 minutes after an independent sign-in (two accounts, November 2026) | The approved workspace session ended 0.1 s and under 60 s later with `AUTH_KEY_UNREGISTERED`; TgWitness now never ends the approving session |
| Witness probe | Answers arrived as live updates after 0.9–5.0 s. Workspaces with an imported key start in Read-only and record `probe.skipped` until it is turned off |
| A second client on the same session key for 5.5 min | 3 of 4 probe answers did not reach the collector live and were found in history (`coverage.gap`, `probe_ack_missed_live`); intake stayed marked as failed until the next live answer (`connection.restored`) |
| New sign-in with a login code | `account.sessions` listed it under `added` 65 s after Telegram's creation time; ended from **Sessions of this account** |
| Chat export stopped with `kill -9` after 80,000 messages | The signed partial copy survived; the next export resumed it with two sessions in `acquisition.sessions`; **Start over** discarded it |
| Crash in the files stage of a 106-message chat | Resumed from the signed checkpoint added in #193: 106 messages carried over, 19 files reused, 23 downloaded; `verify.py` exit 0 |
| Attachment reuse | 42 of 42 files reused from observation media and earlier exports |
| Set-aside cleanup | A copy set aside nine days earlier was removed at startup; newer copies stayed |
| TSA revocation data | OCSP for the Belgian and Sectigo signing certificates and the FreeTSA CRL stored next to the replies |
| Phala TDX deployment from the panel | Commit `6c7171228`, 2.5 min, all six checks passed, TCB UpToDate; `verify.py` checked the journal's TDX quote offline with Intel's collateral; CVM deleted afterwards |

The verifier summarised the probing workspace as "Witness probes: sent 32, answered live 28, answer missed live 4; slowest live answer 5.0 s", and the witness workspace as "seen 50, seen live 49, answered 50". Not checked in this run: alert delivery to an outside destination, which was not configured, and the automatic only-session run, which Telegram allows 24 hours after the key was created.

## Qualification status

The 9 November run added Phala Intel TDX deployment with a live Telegram account, drand checkpoint anchoring, session fingerprint logging, ID gap analysis, and verified export with HTML, JSON and PDF report seals.

The retained live journal establishes real collection, edits, deletion notices, attachments, rechecks and rate-limit handling. Docker/Compose and installation checks are recorded in CI; the real Phala deployment and external timestamp verification are described above. Saved MTProto state and catch-up are implemented. The newer backfill commits messages and cursors atomically and preserves available Recent Actions, with synthetic coverage today.

The following items remain unqualified:

- **Sealed export**: a key was created inside the VM (see above). Verifying an export from such a run at the `sealed` tier, and QR login from a phone, remain open.
- **Secret Chats in the Phala image**: the installer needs a hash-locked source for the `secret-chats` extra.
- **Message lifecycle matrix** (send, edit, delete, reactions, membership): the 9 November run collected channel history without exercising the write path. Steps 4–8 of the [live acceptance](https://github.com/phishdestroy/TgWitness/blob/main/docs/internal/LIVE_ACCEPTANCE.md) procedure remain unrun.
- **Proxy and MTProxy paths**, **Bitcoin-anchored OTS seal**, **witness tier**, and **external-key backup restore** were not tested in this run.

Further controlled runs for role visibility, reaction actors, prolonged FloodWait, account bans, network and process interruption recovery, mobile QR confirmation, sustained high load and long-term archival follow the [live acceptance](https://github.com/phishdestroy/TgWitness/blob/main/docs/internal/LIVE_ACCEPTANCE.md) matrix.

The [public audit directory](https://github.com/phishdestroy/TgWitness/blob/main/website/validation/audit/README.md) contains only the signed report and aggregate measurements. Detailed operational receipts, raw captures and account material remain private. Application screenshots in `docs/assets/` use fictional demonstration data.
