Calibration addendum: how “evidenced” is read

BPEF v1.2-c · The framework's published reading of “evidenced”

The anchors say a capability must be "evidenced" without defining what satisfies it. This framework resolves that question as follows, with the worked examples below. They are the calibration, and they are written out so that the same question resolves the same way in every edition.

The evidence GRADES and CAPS are unchanged and still bind. A 5 still requires E1 or E2. Entirely-E3 evidence still caps at 3. Absence of evidence still scores 0 or 1. What follows governs only what counts as evidence of a capability, not how trustworthy a source is.

The general rule

A published, specific service description from the provider is evidence that the capability exists. It is not evidence that the capability is good, and it does not lift a score past 3 on its own.

Corroboration (a named client, an analyst evaluation, a review, a certification whose scope reaches the work) is what carries a sub-criterion to 4 or 5.

The strict reading treats a described-but-uncorroborated capability as absent and scores it 1. This framework treats it as present but unproven and scores it 3. Score it 3.

What this does not license: inferring a capability nobody claims, crediting a capability the provider's own materials contradict, or reading scale as capability. Those still score as the anchors direct.

Sub-criterion rules, with the worked example each rests on

B1.2 · Technical support depth

Worked example: Foundever and Konecta both scored 4 on a published tiering description with client evidence and no published SLA. Conectys, Mplus and Armatis scored 3 on the description alone.

B2.3 · Coverage model

Worked example: Teleperformance, Foundever, Konecta and Transcom all scored 5 without publishing shift-handoff practice. Armatis scored 3. No provider in the field scored below 3.

B3.1 · QA methodology

Worked example: Armatis scored 4 on ISO 18295-1 certification alone. Teleperformance and Foundever scored 4. Konecta, Transcom, Conectys and Mplus scored 3 on description alone.

B3.3 · Workforce management

Worked example: Teleperformance, Foundever, Konecta and Armatis scored 4 without naming WFM tooling.

B4.1 · Access control

Worked example: Teleperformance and Foundever scored 4 with scope-per-entity explicitly recorded as unverified.

B4.3 · Controls on sensitive workflows

Worked example: Foundever, Konecta, Transcom and Conectys all scored 4 without publishing a control mechanism. No provider serving regulated work scored below 3.

B6.2 · AI-human hybrid operations: how it was applied in this edition

The published anchor has two halves. Anchor 3 requires deployed AI in production described concretely (which workflows, which tools) and a stated human-in-the-loop quality control design, its own examples being review sampling of AI outputs, escalation triggers, and containment-versus-deflection honesty. Anchor 5 adds a stated position on what stays human.

In this edition the deployment half was evidenced systematically and the control half was collected only where it happened to surface.

B6.2 values should therefore be read as deployment-weighted. Across thirty-two providers, four publish a control over AI output that meets the anchor's own examples: TELUS Digital's confidence-threshold routing with an explicit statement that "a person, not a system, stays accountable"; HGS's human-in-the-loop oversight at every decision point carrying regulatory, financial or clinical consequence, with runbooks and rollback paths; Conectys's stated division of labour between the human and machine halves; and TaskUs's dated Global AI Policy with a risk-proportional human-in-the-loop model.

Simply Contact's Wizz Air case adds a fifth, found after scoring. Probe Group publishes a Responsible AI Policy that mandates human oversight but delineates no decisions as human-only and specifies no escalation procedure.

The scores are not being restated. The bias is uniform (every provider was scored the same incomplete way), so relative order is unaffected, and correcting the mapping would mean moving published numbers on a collection we know to be incomplete.

The next edition collects the control half as a category for every provider, and B6.2 is then scored against the full anchor for the first time.

The worked example, because it is the distinction that decides most cases. TDCX publishes a rich deployment: six named workflows, FLASH Suite, FastTrack, Acuity, and a stated QA mechanism to "automate the QA process and increase sampling volumes to build confidence in performance scoring". That sampling runs the wrong way for this anchor.

What is sampled is agent interactions, sampled by the AI, not AI outputs sampled by humans. And "humans in the loop", where it appears, attaches to a business-intelligence dashboard. An assist architecture is a delivery model, not a control, and the anchor says so itself: anchor 1 lists "agent assist" as an example of a tool merely mentioned.

B7.2 · Regulated-workflow capability

Worked example: field floor of 3 for every provider serving regulated work; Foundever and Konecta scored 4; Teleperformance 5.

Who authored it, not who hosts it

Two sources corroborate each other only if two different parties wrote them. The test is authorship, not the platform the text sits on, and the distinction decides several grades.

A G2 or Clutch seller profile is written by the provider: its description, its minimum project size, its rate band.

A provider's own site and its own seller profile are one author twice, and grading that pair E2 asserts an agreement that does not exist. The same applies to a company's second document: a key-figures page against its own CSR report is one author, not two.

A review on those same platforms is written by a customer. A rating is computed by the platform from those reviews. A LinkedIn headcount range is derived from individual employees' own profiles, and hiring volume from posted activity.

None of these is a company assertion, even though the company controls the page they appear beside. That is why A2.2 accepts "site + LinkedIn headcount range or hiring volume" as corroboration while E2 would reject "site + seller profile": the anchor points at figures the provider does not author.

It is the same distinction the live-posting convention rests on. A marketing page is the provider asserting a capability; a posting, a headcount range and a customer review are all artifacts the provider did not write, or did not write for that purpose.

A4.1 · Employee-side signal

The operative rule sits in the anchors appendix and is summarised here because it decides which number is read before any anchor is applied.

Score the employer profile with the largest review volume; treat a difference of 0.5 or more between sources as material, that being the width of one rating band and so the finest distinction this criterion resolves; and let a material divergence cost one interpolated anchor step rather than a whole band.

The rule is needed because an employer rating is not a single well-defined number. A platform may carry several unmerged entities for one company, two platforms may disagree, and both platforms offer city and country filters whose figures look identical in form to the company-wide one.

None of these is resolved by knowing the corporate history: Glassdoor merged Webhelp into Concentrix and Sitel into Foundever, and carries Konecta, Mplus and Transcom as separate unmerged entities. The profiles are therefore counted, not assumed.

Worked example: Transcom's employer footprint resolves to three unmerged Glassdoor entities, with 2,072, 217 and 131 reviews. The largest governs, giving 2.9 and anchor 0. Mplus resolves to three entities of 22, 17 and 5 reviews, where the largest is also the lowest; the rule is indifferent to which direction the sample points.

Live job postings as evidence

Eleven claims across both editions rest on live job postings, spanning eight providers. Because those claims are graded E2 rather than E3, the reasoning is set out here rather than left as an unstated exception to the evidence scale.

E1 and E2 describe evidence generated by someone other than the provider, and a provider writes its own job adverts.

The exception is deliberate and rests on a distinction the scale does not otherwise draw: a marketing page is the provider asserting a capability; a job posting is the provider acting on one. Reading a posting is not accepting the provider's account of itself; it is observing an operational artifact produced for a different purpose and drawing an inference from it. It is also independently checkable by any reader for as long as it is live.

What qualifies. A live posting, on the provider's own careers portal or on a third-party job platform, naming the site and the language or skill at issue. A posting that names neither does not support a footprint claim.

What it evidences. That the provider is staffing that capability at that location. It does not evidence that a marketed total is correct, and the method is as valuable for what it fails to corroborate as for what it confirms: where a provider markets thirty languages and postings evidence nine, both numbers are reported and the gap is the finding.

The limitation. Postings expire. The read date on the source line is what makes such a claim checkable afterwards, and a reader re-running the search later may find the posting gone without the claim having been wrong when it was made. This is the one evidence type on which the read date is load-bearing rather than administrative.

A4.2 · Hub and business continuity

Worked example: Teleperformance, Foundever, Konecta, Transcom, Conectys, Mplus and Armatis all scored 5. Simply Contact's 5 rests on a named disruption; the others' do not.

Unchanged: scored strictly in both readings

These are scored strictly. Do not loosen them.

Why this direction

The choice is not between a strict rubric and a lenient one. It is between two readings of an underspecified word, one of which is published here and applied consistently across both fields. Adopting the other would move scores to gain nothing that a written definition does not also gain.

The cost is real and belongs on the methodology page: this framework scores what a provider publishes about its capability, corroborated where corroboration exists. It does not audit delivery.

A provider that describes a control model it does not operate will score 3 here and fail in procurement. That is why every data sheet carries its E5 gaps as due-diligence questions, and why the limitations section says private delivery data must be validated in procurement.

That was always the honest description of what this instrument does. It is now written down.