Found is not understood
An agent may reach the page yet choose the wrong product, policy, contact, or next action.
What problem are we solving today?
AI can find your website. Can it do business with you—within your rules?
Most readiness scans stop when a machine can discover and read a page. Pointer tests what happens next: can an authorized agent find the right action, understand the rules, complete a bounded task, and return evidence another system can verify?
Pointer defines the measurement and proof requirements inside 3Dogs Nexus. It is not a certification, and no result is released without human approval.
The gap
01
02
03
A machine-readable site is not the same as a machine-operable business.
Understood is not authorized
An endpoint may exist without clear scopes, identity, consent, limits, or human escalation.
Completed is not proved
A successful response is weak evidence unless the inputs, authority, result, and recovery path can be confirmed in a separate evaluator run.
The Pointer evidence chain
Discovery is the starting line. Operational proof is the test.
Pointer evaluates five dependent layers. A failure in a critical layer caps the outcome; strengths elsewhere cannot average it away.
- DiscoverCan an agent locate the authoritative surface, identity, policies, and supported tasks?
- Execute L1Can it complete a bounded, non-destructive task against a declared contract?
- Govern L2Are identity, authority, consent, limits, audit, and human control explicit and enforced?
- Recover L3 gateDo denial, timeout, replay, revocation, and partial failure resolve within documented contract limits?
- Confirm L3Can a separate evaluator run repeat the task and confirm equivalent evidence?
The public ladder is cumulative: L1 Executable → L2 Governed → L3 Confirmed. L3 also requires every applicable recovery control to pass. A “separate evaluator” did not produce the original evidence, but may be operated by the same organization. Independent external confirmation is a different claim and must be evidenced separately.
What Pointer returns
Evidence states—not a flattering number.
A single score can hide the exact failure that prevents an agent from engaging or converting. Pointer keeps the evidence visible and the uncertainty intact.
The deliverable is a claim-by-claim evidence record containing, at minimum: claim ID, subject and environment, scope, source URLs and content fingerprints, observation and expiry dates, positive and negative results, status, limitations, and supersession history.
Confirmed Repeated in a separate evaluator run
Observed Captured, not yet confirmed
Claimed Declared without sufficient proof
Not evidenced Required proof was absent
Contradicted Evidence conflicts with the claim
Not applicable Excluded with a documented reason
One problem, one bounded proof
REVENUE PATH
SERVICE PATH
SUPPORT PATH
Test the engagement that matters—not a generic demo.
Can an agent become a qualified opportunity?
Test whether it can identify fit, obtain accurate commercial terms, submit a bounded inquiry, and receive a verifiable next step without losing attribution or consent.
Can an agent request the right work?
Test whether it can choose the correct service, supply required context, respect data boundaries, and route the request to an accountable human or system.
Can an agent resolve or safely escalate?
Test authenticated retrieval, policy accuracy, scope limits, recovery behavior, and a human escalation that preserves the audit trail.
From scan to evidence
A transparent assessment with human control at every consequential step.
Define
Name the business problem and bounded task.
Inventory
Map public surfaces, contracts, policies, and claims.
Challenge
Run normal, denied, malformed, replay, and failure cases.
Confirm
Repeat material claims with a separate evaluator run.
Prioritize
Rank fixes by engagement value, control risk, and effort.
Approve
Keep implementation and release explicitly human-gated.
Non-destructive by design. The assessment does not change production state. If a consequential action cannot be challenged within an authorized sandbox or other safe boundary, it is marked not tested this round—not forced. Implementation always requires separate authorization.
Subject-runnable, not pay-to-pass. Every eligible no-fee measurement includes complete findings, failure reasons, exact acceptance tests, method version, scope, configuration, test vectors, result derivation, rerun instructions, and evidence digests. The assessed organization can fix the problem and rerun the same measurement without paying 3Dogs. Paid work may cover additional interpretation, prioritization, remediation, or implementation, but cannot be required to understand or self-fix the result. Payment cannot improve the result, waive a gate, or control publication or appeal.
Useful baselines, harder questions
Pointer begins where public discovery scanners usually stop.
| Question | Discovery baseline | Pointer evidence requirement |
|---|---|---|
| Can AI find the site? | Robots, sitemap, metadata, machine-readable content | Preserved and separately rechecked |
| Can AI understand the offer? | Content and structured identity signals | Tested against actual task selection and claim consistency |
| Can AI safely act? | Often outside the scan boundary | Bounded execution with identity, authority, consent, and limits |
| Can failure be trusted? | Often outside the scan boundary | Denial, timeout, replay, revocation, soft-404, and recovery tests |
| Can the result be proved? | Point-in-time report | Signed or hash-bound evidence that a separate evaluator can confirm |
Public discovery scanners remain useful starting-point diagnostics. Pointer retains those checks without making claims about another product's complete current feature set.
We practice the harder test
3Dogs does not claim perfection.
Cloudflare's Agent Readiness web UI reported 100/100 and “Level 5 Agent-Native” for https://3dogs.ai in an artifact captured on August 5, 2026 at 23:46 PT; that artifact does not expose its collector version. Cloudflare's public MCP scanner, version 1.0.0, rechecked the same origin with profile all at 2026-08-08T21:08:51Z and returned Level 5/5 Agent-Native. The current programmatic result does not return an overall 100-point number, so Pointer does not translate one label into the other.
Cloudflare's result covers public discoverability, content accessibility, bot controls, and capability discovery. It does not establish Pointer's authorization-declaration, honest-error, bounded-execution, governance, recovery, or confirmation requirements.
The decisive self-test was not another document review. The 100/100 result coexisted with an advertised root commerce API, MCP endpoint, and documentation route that live requests did not support. Multiple reviews initially read the declarations and misclassified the missing commerce surface as an authorization defect. Calling the declared endpoints corrected the diagnosis: the advertised surface was not observed. Pointer now treats that sequence as a foundational control—documents can prove what was represented; availability, execution, and authorization require appropriate runtime evidence.
We are applying Pointer to our own A2A sales and marketing path first, recording each gap, intervention, and separately repeatable outcome. Public claims will follow the evidence, not precede it.
Review the attributed evidence, dates, digests, and limits →
What problem are we solving today?
Give Pointer one job that matters.
Choose one journey—a quote request, service request, or support case—and name the owner who can approve testing. The future no-fee measurement must return complete findings, exact acceptance tests, and everything needed to rerun after a self-fix.
Current release state: the subject-runnable measurement bundle is not yet released. Pointer will not issue a shareable result at any assurance level, publish a ranking, or sell remediation against method 0.2.1 until that bundle exists and the protocol-truth and soft-404 hard gates close on the live site. Internal diagnostics are not Pointer results. A qualified DEV pilot request identifies the business problem, bounded task, authorized actor, allowed inputs, prohibited actions, expected outcome, failure expectations, evidence requirements, human escalation, and evidence freshness window. No implementation or public release is automatic.