Anthropic Claude
ALLOWEDDeclared
$0.15
Actual
$0.12
Variance
-20%
Independent assurance for autonomous work
Autonomous work finished.
What can you actually prove?
PromptKing creates an independently checkable record of what was authorized, what happened across the run, and what outcome the evidence establishes.
Out-of-path
Independently checkable
Claim-bounded
The agent can make the claim. The evidence has to support it. A policy verdict in the record is not the same as an established outcome.
Canadian-built. Canadian-resident evidence.PromptKing is an Ontario company. Evidence data lives in Canada (ca-central-1). Sovereignty isn’t where your model runs — it’s whether you can prove what it did, and who holds that proof.
The category
Your runtime knows what it did.
That does not mean your enterprise knows what happened.
01
Execution systems
Run the work.
02
Logs & observability
Show activity.
03
PromptKing
Establishes what the evidence means for the run.
The questions that close the record
01Was this work authorized?
Not whether an agent started. Whether the declared permission and policy state were actually established before execution continued.
02Which agent and which governed run produced this outcome?
Identity and run identity are different facts. Knowing the agent does not identify the run, and neither fact is authority.
03Did the work actually complete across the handoffs?
A local completion signal is not whole-work closure. Handoffs, retries, and vendor calls have to remain visible in the record.
04What can the enterprise rely on now?
Only what the evidence supports. Missing, contradicted, or bounded facts stay bounded — they do not silently become proof.
Evidence lineage
Kept separate on purpose.
PromptKing keeps agent identity, run identity, authority, outcome, and evidence separate so the final record does not have to reconstruct them after the fact.
Authority
What permission was established?
Agent
Which agent was involved?
Run
Which governed run?
Handoff
What crossed systems?
Outcome
What was recorded as result?
Evidence record
What can be checked now?
The moat
The execution system does not get the final word.
Out of the execution path
PromptKing does not become the runtime.
Evidence across the run
Evidence is preserved across the work, not just at the final response.
Independently checkable
The resulting record can be examined separately from the system that performed the work.
PromptKing decides. Your stack executes. The evidence makes the recorded decision independently checkable.
Real evidence
Many facts. One decision-grade record.
One identical governed prompt. Six vendors. Every run got a receipt. Every receipt got a hash. In the recorded Client Zero run, watsonx was denied — and the agent halted immediately. Policy verdict is recorded. It is not, by itself, an established outcome.
OpenAI GPT-5.6
ALLOWEDDeclared
$0.10
Actual
$0.08
Variance
-20%
Google Gemini
ALLOWEDDeclared
$0.10
Actual
$0.09
Variance
-10%
xAI Grok
ALLOWEDDeclared
$0.10
Actual
$0.11
Variance
+10%
Meta Llama
ALLOWEDDeclared
$0.05
Actual
$0.04
Variance
-20%
IBM watsonx
APPROVAL REQUIREDDeclared
$500.00
Actual
$0.00
Variance
-100%
Policy halted the agent before the vendor was called
Every receipt is publicly checkable. Each SHA-256 hash is recomputed in your browser, not by our server.
Fleet lineage
One agent delegated to another — every hop got a receipt, and the forged parent was rejected fail-closed. Chain lineage is established from server-derived lineage evidence, rather than agent claims.
Integrity
SHA-256 verified
Vendor coverage
6 vendors
MCP tools
27 tools
Proof run
July 15, 2026
Stranger test
Passed July 31, 2026
Client Zero
June 30, 2026
Client Zero — validated on our own M365 Copilot tenant, completed June 30, 2026. No external logos. No fabricated case studies.
Who closes the record
When autonomous work matters, somebody has to close the record.
Security
Know whether execution stayed within established authority.
AI Platform
Give builders evidence without replacing their runtime.
Operations
Escalate contradiction, missing evidence, and incomplete outcomes.
Risk / Audit
Review a portable record instead of reconstructing a run from logs.
Claim discipline
Trusted because it knows when not to claim proof.
A hash can prove integrity without proving authenticity. A native event can prove the vendor reported an event without proving it was authorized. A completed workflow can still be unproven.
✗
No "the dashboard says so." The evidence must survive the interface.
✗
No inference into PROVEN. Weak correlation remains weak correlation.
✗
No completion = proof shortcut. Outcome claims stay bounded by the evidence actually captured.
✗
No compliance theatre. PromptKing reports evidence state, not universal legal or physical-world truth.
A link anyone can check.
PromptKing turns autonomous work into a record that can be inspected beyond the runtime that produced it.
See a real record →