Six AI agents, one policy engine
Demo organization · controlled run
Jul 15, 2026 (five runs and the superseded watsonx receipt) · Jul 18, 2026 (watsonx receipt to cite)
Receipts: governed-run-v1
Shows. Agents running on Claude, GPT-5.6, Gemini, Grok and Llama each completed one low-budget research task and were allowed. An agent running on IBM watsonx declared a $500 high-risk operation, was stopped at a human-approval step (recorded verdict APPROVAL_REQUIRED) and halted. The first watsonx receipt (Jul 15) stored the wrong verdict, 'allowed'; it is marked superseded by the Jul 18 receipt.
Does not establish. That the six agents ran the same task: the sixth declared a different, deliberately high-cost operation. The vendors did not take part, and no audit, certification or endorsement is implied. Costs are agent-reported.
An approval, then a continuation under a second runtime
Demo organization · one controlled run
Aug 5–6, 2026
Receipts: governed-run-v1; obligation artifacts v1 and v2
Shows. A parent receipt (Aug 5, 22:54 UTC) names a Claude task runner. An approval request opened at 23:22 UTC names that receipt and a plan to continue on a watsonx leg; it was approved at 23:31 UTC. A receipt from the watsonx continuation agent was issued on Aug 6 at 15:14 UTC. Each artifact's hash recomputes in public, and the approval is bound to the parent receipt.
Does not establish. That the continuation is publicly bound to the approval or to the parent. The continuation receipt carries no reference to either, and PromptKing's own chain check labels the lineage legacy_unbound. The parent receipt records no policy verdict (no_policy_context) and no cost, so the reason for the approval request is not visible in it. One controlled run, not volume.
A receipt with an issuer signature and an external timestamp
Demo organization · one receipt
Signed Aug 11, 2026 · anchored 20:39:14 UTC
ES256 issuer key · RFC 3161 timestamp · Merkle batch
Shows. A receipt's content hash, signed with PromptKing's issuer key (held in a hardware key service) and verifiable against the published public key, bound into a Merkle batch, and timestamped by an independent timestamp authority. A clean client checked all four layers from the public bundle alone.
Does not establish. That the recorded action was correct or complete, or a legal guarantee that no party can dispute it. It is one receipt, not volume. The status the bundle reports is PromptKing's own check; a reader verifies the timestamp, certificate chain and signature independently.
Ten strangers, five seconds each
External readers · our own record
Aug 29, 2026
Article only; test records not reviewed here
Shows. We showed our published record to ten strangers for five seconds each. Some read the overall verdict as covering a line marked 'Authority: not established'. We logged it as a defect and changed the page the same day.
Does not establish. A statistical result. The article does not give how readers were recruited or the exact script, and we have not reviewed the test records for this page.
Intervention records
Fixture
Sep 15, 2026
Four fixture outcomes
Shows. The same record format shown for an intervention that was established, contradicted, not established, and still pending.
Does not establish. A live safety control. Fixture data only.
Decision threshold records
Fixture
Sep 19, 2026
Four fixture cases
Shows. Whether an act matched the confidence threshold its decision declared, including a case where the act was taken below threshold and the record flags the mismatch.
Does not establish. That a decision was correct or its confidence calibrated. Fixture only, no vendor call; the records are self-reported and unsigned.
ServiceNow fixtures
Fixture
Aug 15 – Oct 3, 2026
Work-record assessor, profile PK-WR-HOME-0-v1
Shows. The work-record assessor was run on synthetic incident and change fixtures. A read-only reader was provisioned on a ServiceNow developer instance.
Does not establish. Customer, production or enterprise ServiceNow data. The developer-instance reader's API authority is unverified, and the ServiceNow lane is not closed.
Source. Internal record. Not public today.
Salesforce Agentforce scenario fixtures
Fixture
Sep – Oct 2026
Work-record assessor, profile PK-WR-HOME-0-v1
Shows. Ten synthetic case-refund scenarios were run through the assessor to test how it reports a missing refund read, a differing amount, a removed approval and a late declaration. A supplier-onboarding scenario is at design stage.
Does not establish. A live Salesforce read. These are hand-authored fixture replays with no customer data, and no Salesforce relationship or endorsement is implied.
Source. Internal record. Not public today.
Salesforce controlled run sealed into a Work Record
Controlled run · rehearsal
Oct 24, 2026
PromptKing-owned developer org · portable attempted-vs-actual record
Shows. A synthetic supplier-onboarding run by an agent in a PromptKing-owned Salesforce developer org, reread afterwards and sealed into a portable record that compares what the agent claimed with what the reread showed.
Does not establish. The outcome. It is labelled a rehearsal because the reader was not yet independent of the agent; an independent reader has been built but its live capture has not run. No customer data, and no Salesforce relationship or endorsement is implied.
Source. Internal record. Not public today.
Computer-use run
Controlled run
Sep 2026
One bounded run
Shows. One bounded computer-use run by a founder-dispatched external AI runner on the founder's own Google Drive, recorded as a work record built around the agent's declared intent.
Does not establish. That the screen state was verified, or that the agent is safe. Nothing beyond that single run. The record shows what was declared and observed, not that intent was independently checked.
Source. Internal record. Not public today.
GitHub Copilot cloud agent on a throwaway repository
Controlled run · throwaway repository
Sep 2026
Capture format under test
Shows. A GitHub Copilot cloud agent was given tasks on a throwaway repository, and captures of the sessions were hash-manifested for the record format.
Does not establish. Independent confirmation. Each capture comes from one session, is unsigned, and has no independent witness; authority and event ordering are not established. Not a customer repository.
Source. Internal record. Not public today.
Transparency-log packs
Fixture
Sep – Oct 2026
Three synthetic packs
Shows. Three synthetic packs showed how a log's commitment can bound claims about the event fields it covers.
Does not establish. That any vendor's real log was checked. The verifier and inspection screen are not built, and a log is bounded evidence, not proof of the outcome.
Source. Internal record. Not public today.
Independent recompute plan and verifier kit
Built · not yet run outside PromptKing
Designed Sep 20 · rehearsal Sep 29 · kit re-run on Linux Oct 31, 2026
Plan, builder-run rehearsal, two separate verifiers
Shows. A plan for someone who did not build the system to recompute a record offline. A builder-run rehearsal on a fixture reproduced the result, flagged a one-byte edit, and reported 'could not check' when evidence was missing. Separate Node and Python verifiers of the receipt rule now agree on a 17-row synthetic dataset, on Windows and on Linux, with the expected matches and exceptions.
Does not establish. That anyone outside PromptKing has recomputed a record. All runs so far were by PromptKing and its own reviewers; no outside run is documented as of Nov 7, 2026. The kit is shared on request.
Source. Internal record. Not public today.