The artifact you buy.

One use case, three weeks, one fixed price. Everything below descends from this cover block.

As it arrived: "Can we use AI on our vibration data to predict failures on Line 3?" Nothing in that sentence can be tested. "Predict failures" has no metric, there is no number for what the team does today, and there is no owner — so the best vendor demo would have become the definition of success. That is the most-documented way these projects die.

As it left, after the rewrite:

Scope is this one workflow only. Line 2 (except CMP-3107, which feeds Line 3 actuators), other plants, non-rotating assets, and the order-entry initiative raised at kickoff are all out of scope, in writing. A second use case is a second audit, at the published price.

Every figure carries a tag: [counted] came out of a system of record or a countable sample; [estimated] is modeled. Counted numbers can gate a pass or fail. Estimated numbers are reported and never gate — a modeled number never decides whether money changes hands.

Two of the client's own records were wrong, and the baseline says so. A 47-minute filler fault was coded as "changeover," hiding it from the availability numbers; a gearbox failure record was keyed "fail to start" when the trend data and the record's own free text say vibration. Both were corrected against the physical record, and the corrected ledger governs everything downstream. Line speed was uprated in Q1 2026 — pre- and post-uprate windows are noted per asset.

Baseline signed as fair by the sponsor, 2026-07-15 (fictional) — before any threshold was negotiated. That order matters: nobody can move the goalposts later, including us.

Four thresholds, frozen in writing before the pilot starts, each set against the signed baseline by a margin the sponsor agreed was worth the pilot's cost. All four gate on [counted] metrics adjudicated against the frozen event ledger — never against the model's own confidence score.

The pass/fail rule, in one sentence: PASS means every threshold is met on the measurement sample, FAIL means any single one is missed — binary, no weighting, no averaging, no "substantially met," no Tektari discretion, and ties go to the client.

Your team runs it on your data — here, all 34 adjudicated events plus 220 stratified healthy asset-windows, frozen before any vendor tuning — and Tektari is not in the room. In Week 3 the client's own two people dry-ran the test on a six-event sub-sample with us observing and not assisting; the exit condition is that they can run it without us. A fifth line, projected downtime savings, is reported and never gates.

One boundary, stated in the test itself: on tier-zero assets the pull-or-defer call — the ~$214,000 kind of call — stays a human sign-off. The test measures whether the system flagged in time. It never treats the model as fit to make that decision.

The assessment: the problem is well-defined enough to pilot — after the rewrite, not before — and the acceptance test is runnable by the client alone. The recommendation, in writing: run the scoped pilot, with conditions.

• 1 The corrected ledger governs. A pilot scored against the raw keyed records inherits the human errors found in Week 2, and the result means nothing.
• 2 Tier-zero defer-vs-pull stays human. The pilot flags; it never decides.
• 3 The vendor accepts the frozen holdout in the contract. No access to the test sample or ledger for tuning, no post-freeze threshold changes, no per-asset hand-tuning during the run. A vendor who will not sign that clause is the finding.
• 4 Below 30 adjudicable events on a re-drawn window, stop. The population is then too event-thin for a defensible test, and the honest move is to say so — not to score a handful and call it validated.

The recommendation is written regardless of which way it cuts. When it cuts the other way — "do not run this pilot" — saying so costs Tektari the $2,500 tranche. That is deliberate.

What this assessment does not claim:

• — It does not predict the pilot will pass. It hands the client a test the client runs; the result is the client's, on the client's data.
• — It does not recommend or rank any vendor. The test is vendor-neutral — usable against any vendor's pilot, or an internal build.
• — It is not a certification, a warranty, or professional advice. It is an assessment against the published protocol, and nothing more.

$12,500 fixed. $10,000 invoiced at kickoff. The final $2,500 is invoiced only when the scoped pilot passes the acceptance test in section 04, as run by the client's own team on the client's data.

• pass $2,500 invoiced.
• fail Tranche waived — never invoiced.
• never run Tranche waived — never invoiced.
• dispute Unresolved after one re-run: tranche waived — never invoiced. Ties go to the client.

Per the charter: this fee credits toward no other Tektari offer, and no other offer is pitched inside the engagement. Tektari sold no implementation and took no vendor money in connection with this assessment.