clembot

proof

Receipts, not testimonials

Everything below is exported from this platform's own operational data at checkpoint time. Where a number is unflattering, it is here anyway, the whole point of publishing a record is that you don't get to pick which parts.

Behind the headline: the evidence chain appending, with the last 13 of its 169 records drawn. Each one carries the hash of the record before it, so after every append a check walks back down the chain.

exhibit · audit-hardening · 2026-07-22

Recompute the chain yourself

Below are 5 consecutive evidence records exactly as they were written, plus the hash published for each one. Your browser recomputes every hash from the record bytes, no network call, no server. Then flip a record's text and watch the chain break from that point forward.

  1. buzz-hardening2026-07-22T16:52:27.996Z
    seq 73

    Tamper-evident chain over outputs/evidence/*.jsonl; anchor writes HEAD+chain-index.jsonl; 6/6 tests pass; verify clean on 73 real records.

    published
    4a03710c939af9dfa7ef48e86ec375ebd05d5c712e8f3478d36a0721b018188f
    computed
    n/a
  2. buzz-hardening2026-07-22T17:13:34.682Z
    seq 74

    logEvidence signs canonicalBody when AGENT_KEY_<ID> set; grandfathers unsigned. 7/7 UC1 tests + real-CLI verify/tamper e2e pass.

    published
    64d9ff1211695f95483bb03c81350a7210b13d0cd64df1861952a47488dd9d41
    computed
    n/a
  3. buzz-hardening2026-07-22T19:10:33.193Z
    seq 75

    applyTransition signs the transition tuple with the Publisher key; additive nullable cols; CONTENT_DB_PATH makes the sole-writer testable. 5/5 UC3 + UC1/UC2 regressions pass.

    published
    7140b32a8bc894b336cdb2e34d8c877c0a943572d7fb4a170e81bba50468f08c
    computed
    n/a
  4. buzz-hardening2026-07-22T19:52:11.694Z
    seq 76

    Pure read-only projection of evidence+transitions+slack into FTS5; deterministic event_ids; tenant-scoped; pipeline-activity reads projection w/ inline fallback. Fixed FTS5 special-char query bug. 6/6 UC4 tests.

    published
    f8034a425aa734c7bbb3d6b175a435ce915aad0878fa35e2f0c9eb32f903714a
    computed
    n/a
  5. buzz-hardening2026-07-22T19:56:53.334Z
    seq 77

    Every RAG answer appends verifiable event_ids from the UC4 projection; zero-receipt answers labelled unverified. 5/5 UC suites green. buzz-hardening COMPLETE.

    published
    3419953595ff0b4374bfa36cc5dce2f307d73822e0b3059615755334c4654685
    computed
    n/a
What this proves, and what it doesn't

What it proves

  • These 5 records are byte-for-byte what they were when the chain was anchored on 2026-07-22. Editing any covered field in any record changes that record's hash and every hash after it.
  • The segment descends from the published anchorb5b97423e7ecc593…, and its terminal hash is the committed head of the whole chain.
  • The verification runs entirely in your browser, over the same JSON that is printed on this page. There is nothing to trust in between.

What it does not prove

  • That the anchor is honest. A chain shows that history wasn't edited after anchoring. It cannot show that the anchor itself was truthful. That needs the baseline to be independently visible, and the repository holding it is private, so this is self-consistent and tamper-evident, not third-party attested. Saying otherwise would be the exact overclaim this site exists to avoid.
  • That an agent authored a record. That is what the Ed25519 signatures are for, and zero of the records are signed. The signing code is complete and tested; no keys have been provisioned. No public keys are published yet, so there is nothing here to check against.
  • That the work was good. A record proves an action was logged, not that it was correct. That is what the human gate is for.

evidence log

Every agent action, written down

169
evidence records
2026-03-19 → 2026-09-16
15
distinct agents
19
distinct commands
0
cryptographically signed
Signing code is complete and tested. No keys provisioned yet, so this is zero, and it says zero.

the ledger

138 commits, parent-linked

The build history of the engine itself, each commit shown with the parent it descends from. This is displayed linkage, not verified linkage: recomputing a git hash in your browser needs the raw commit object, which carries the author's email. Theevidence chain above is the part that is actually recomputed. Those two things are easy to blur and this site won't.

1,704
file changes
+148k
lines added
−15.4k removed
14
weeks with commits
2026-05-04 → 2026-09-14
  1. 7143ccd35911d12026-09-14

    chore(site): archive site/ — clembot.wanessalabs.com moved to ClemVault

    2 files · +69 −0

  2. 35911d196234752026-08-09

    checkpoint: De-hook layer: opt-in per-feed clickbait removal on the follows poller, local-Ollama-only, piloted on stratechery

    7 files · +159 −31

  3. 96234759909f592026-08-09

    feat(follows): de-hook pass — read article feeds as plain facts

    7 files · +826 −1

  4. 9909f5965da3972026-08-07

    checkpoint: Agent map: /system/ + per-tenant Your Team module, pushed and deployed

    8 files · +217 −38

  5. 65da39792771e82026-08-06

    content(site): what-it-replaces copy, and correct two stale claims

    4 files · +67 −59

  6. 92771e850596792026-08-06

    fix(test): the determinism check no longer dirties the tree or false-alarms

    1 files · +10 −2

  7. 5059679616474c2026-08-06

    feat(dashboard): "Your Team" — the per-tenant view of the agent map

    9 files · +788 −0

  8. 616474cc3a45fc2026-08-06

    feat(site): the agent map — an org chart generated from the repo

    39 files · +6368 −4

  9. c3a45fcf666c6e2026-08-06

    feat(slack): commit-specific marketing angles with a link to the diff

    4 files · +193 −14

  10. f666c6e7c34cb92026-08-04

    docs(newsroom): session record + correct the standup channel in the registry

    3 files · +15 −3

  11. 7c34cb9ecb2c7c2026-08-04

    merge: newsroom failure visibility + duplicate-poll guard

    0 files · +0 −0

  12. ecb2c7c424d1682026-08-04

    checkpoint: clembot-site phases 2b-5 + design/motion passes + 31-project audit; portable design panel v1.3.0 ported to 3 ClemVault projects

    9 files · +180 −52

  13. 3c1c29cd96893c2026-08-04

    fix(newsroom): make a failed standup announce itself instead of going quiet

    4 files · +276 −6

  14. 424d168bf4dd012026-08-04

    feat(design-panel): v1.3.0 — versioned installs, --check/--update, canvas colour resolution

    3 files · +141 −24

  15. bf4dd01ac787e12026-08-04

    feat(design-panel): grouped contrast sets, TS declarations; ported to clem-travels

    3 files · +115 −23

  16. ac787e135cb6a22026-08-04

    feat(design-panel): one-command installer, browser-native colour parsing, share prompts

    5 files · +401 −11

  17. 35cb6a21d2804b2026-08-04

    feat(site): portable design panel — live token editor with contrast readout

    8 files · +782 −2

  18. 1d2804b83547d02026-08-04

    feat(site): 31-project audit, cost section, og:image, 2 real Anatomy studies, screenshot policy

    35 files · +1194 −47

  19. 83547d09b248cd2026-08-03

    feat(site): Clembot mark in the masthead, favicon, and the canonical projects link

    11 files · +140 −5

  20. 9b248cd72e12f32026-08-03

    feat(site): rhythm and motion — scale contrast, full-bleed data, explanatory animation

    10 files · +575 −70

  21. 72e12f38e56e072026-08-03

    test(site): enforce the no-third-party-calls claim the footer makes

    2 files · +126 −1

  22. 8e56e0790ab5882026-08-03

    feat(site): design pass — Swiss/technical-document register, self-hosted type, enforced contrast

    28 files · +1101 −440

  23. 90ab5887e01ffd2026-08-03

    docs(site): keygen proposal — why Clemens must run it, and why it won't sign the exhibit

    1 files · +4 −1

  24. 7e01ffdbbebca32026-08-03

    fix(site): resolve the week-number dispute in the vault + wire the waitlist mailto

    6 files · +99 −32

Showing the 24 most recent of 138. 3 commit subjects are replaced with [redacted], they name client work, which never appears on this site. The commits themselves are still listed; only the message is withheld.

cost

What it took

The most-asked question about AI-built work is what it actually cost, and almost nobody publishes it. Here is what this system can measure about itself, and, immediately after, the number it cannot.

Elapsed
21 weeks
From 2026-04-24. Planned for 13; it ran long.
Commits
138
Across 5 months, 2026-05-04 → 2026-09-14.
Lines
+148k / −15.4k
Net 132k across 1,704 file changes. Generated code inflates this, it is a volume measure, not an effort one.
Agent actions logged
169
15 agents across 19 commands.
Projects audited
31
Status as of 2026-05-20. Not all of them are built, see the project list.
People
1
The gate has always been one person. That is the constraint the whole system is designed around.

[VERIFY]The money and the model spend are not published, because they are not recorded anywhere this site can read.

No token counter, no billing export and no cost log exists in the repo or the vault, it was never instrumented. Rather than estimate a figure from commit volume and present it with a currency symbol in front of it, the gap is left labelled. An invented cost on a page arguing against invented numbers would be the single worst sentence on this website.

To fill it: record the figure, then add it to the exporter, not to this page.

pace

Month by month

Per-project ship dates are not published because they do not exist in any source, the project audit records status without dates, and the vault's milestone log was abandoned after five days. What is real is the build itself, so that is what is shown.

  1. 2026-0547 commits+100.2k −9.6k
  2. 2026-0619 commits+12.4k −3.8k
  3. 2026-0744 commits+13.3k −1.1k
  4. 2026-0827 commits+21.6k −0.9k
  5. 2026-091 commits+0.1k −0.0k

activity

Shape, not transcript

Six agent personas talk to each other in a chat workspace all day. The activity log records volume and threading, it never recorded message bodies, so there are none to publish. What follows is honestly all that exists: 202 messages and 146 threaded replies over the last 30 days.

By persona

  • Clembot-Publisher136 · 20 threads
  • Clembot-Research64 · 21 threads
  • Clembot-Dev2 · 36 threads
  • Clembot-Marketing0 · 44 threads
  • Clembot-Brain0 · 21 threads
  • Clembot-Sales0 · 4 threads

By day

31 days with activity in the window. Gaps are real, nothing here runs on a cron that posts for the sake of posting.

The six personas
  • Clembot-Brain, Chief strategist. strategy and synthesis.
  • Clembot-Marketing, Content & growth lead. content angles, hooks, distribution, audience reach.
  • Clembot-Sales, Revenue lead. pipeline, conversion, client objections, demos.
  • Clembot-Research, Market intelligence. evidence, market patterns, competitors, precedent.
  • Clembot-Dev, Engineer. feasibility, automation leverage, what to build or instrument.
  • Clembot-Publisher, Operations & calendar. cadence, scheduling, the pipeline from draft to posted.