We don't grade our own homework and take our own word for it.

1. The bar is agreed before the work starts

Before we build or audit anything, we put the acceptance criteria in writing and agree them with you. The definition of "done" is fixed up front — it cannot quietly drift mid-project to match whatever the output turned out to be. If the work doesn't meet the bar we agreed, you'll be told it didn't, not sold a moved goalpost.

2. Nothing is self-graded

A result is never accepted just because the system — or the person — that produced it says it passed. Before a deliverable reaches you, a separate, independent check re-derives the result from the underlying source material, and a second reviewer whose explicit job is to find what's wrong has to clear it with zero open issues. Disagreement between passes is treated as a defect to resolve, not noise to average away.

3. "Unverified" never ships as "verified"

If a check can't run, an input is missing, or a status is ambiguous, the default answer is "not verified" — never a silent pass. You will sometimes see us say "we could not verify this" in a deliverable. That sentence is the guarantee working, not a gap in it.

4. What you approved stays what it is

Deliverables and run outputs are fingerprinted when they're produced, so a later edit to a "final" file is detectable rather than silent. This matters most on the things you'll rely on for a decision — a score, a report, a run summary. The version you signed off on is provably the version that exists.

5. Labels mean what they say

Every deliverable and every page on this site is labelled for what it actually is:

• DEMO — a labelled demonstration, run on sample data, not a live client engagement.
• Planning scenario — an illustrative example of how a service would apply to an industry, not a client case study or a performance claim.
• Unsolicited sample — real work we produced without being asked and without any client relationship, on real public data, clearly marked as such.

We do not upgrade a demo to a "case study" label, or a planning scenario to a "result," to make a page read better. If a claim isn't grounded in something we can point to, it doesn't go on the page.

What stays private — deliberately

The exact test batteries, scoring internals, review procedures, and tooling that keep these guarantees are Taz's working IP, and we don't publish them — here or in deliverables. What you get instead is stronger than a peek at the machinery: the agreed acceptance criteria, the evidence behind every finding, and deliverables you can independently re-check against the bar we agreed. Outcomes and proof are yours; the recipe is ours.

What we measured — July 2026

Every number below comes from a run we can re-execute, was scored against a bar we wrote down before the work started, and includes the results that went against us. All figures describe our own engineering benchmarks, at the scale stated — not client outcomes.

• Store readiness: 9 real storefronts audited on public pages only, robots.txt honored on every fetch (disallowed pages skipped and logged, never crawled). Scores ranged 12–19 out of 20. The two public rubrics we found in this field are covered check-for-check by our audit, each check executable rather than a checklist row.
• Knowledge formatting: 10,060 public-domain documents processed in one run, 100% without error, in just over two minutes; a stratified 205-document re-check agreed with the automated results on every sampled item.
• Evaluation calibration: two independent AI model families scored the same 30-example set; they agreed 79% of the time overall, above the pass bar we committed to in advance. One example produced an unusable verdict and is counted, not discarded.
• Output safety checks: four of five automated checks reached measured precision 1.00 and recall 1.00 on an adversarial test set built to fool them, and now block on failure. The fifth missed our recall bar and is reported as advisory only — not blocking — until it earns it.
• Cost audits: billing-export readers for four providers, each required to reject truncated, forged, mixed-currency, and duplicated files with a named reason. Savings projections are never produced without their assumptions attached. Only one provider's reader has been run on a real export so far; the other three are built from published formats and say so.
• Personal-data detection: three new detection classes each measured above 85% precision and 80% recall on a held-out split — synthetic training data, pilot scale, stated plainly.

What we did not achieve, published anyway: a small trained review model that failed to beat its baseline (kept as a negative result); a retrieval configuration that scored zero on one recall measurement (kept, and named the reason we're replacing that component); two personal-data classes with no usable public training data, left switched off rather than shipped weak.

What we found looking outward

In our own public research of competitor documentation for AI-eval and agent-testing services, we did not find published equivalents of these guarantees — a pre-agreed, unmovable bar; independent re-checking before delivery; or tamper-evident final artifacts. We searched; we didn't find it publicly documented. That's a fact about what's publicly written down, not a claim that no one else does this internally.