QnAs 2-3-26.odt

2/3/2026 Q & As What's the corpora size for training and fine-tuning (number of documents)? WE EXPECT 300-400 DOCUMENTS PER YEAR. MOST ARE 1-3 PAGES LONG. A FEW COULD BE 20+ PAGES. IF THE QUESTION IS HOW MANY DOCUMENTS WOULD BE NEEDED TO CUSTOMIZE THE PRODUCT, PERHAPS 20 OR SO.

The RFP mentions subdomain engines, what's the number of domains envisioned, 1-4, 5-10 or more than that? IT COULD BE SEVERAL. CERTAINLY NO MORE THAN 10.

Does the security requirement allow for the use of LLMs, prompting, and RAG? The security requirements allow the use of Large Language Models (LLMs), prompting, and Retrieval-Augmented Generation (RAG), provided the solution complies with current federal mandates:

Federal Requirements (Agencies and Contractors)

M-25-22, Driving Efficient Acquisition of Artificial Intelligence in Government

M-26-04, Increasing Public Trust in Artificial Intelligence Through Unbiased AI Principles

Hosting: FedRAMP authorized or on premises only.

Data Protection: No sensitive data exposure to non-compliant systems.

Based on the Statement of Work a two-tier Machine Translation service designed to balance translation quality, cost predictability, and risk under a firm-fixed-price contract. THE TWO-TIER APPROACH IS ACCEPTABLE, PROVIDED THOSE USING THE GIST TRANSLATION SERVICE HAVE ACCESS TO A CUSTOMIZED MACHINE TRANSLATION DATABASE THAT RECOGNIZES OUR INDUSTRY-SPECIFIC TERMINOLOGY. HOWEVER, PLEASE NOTE THAT WHILE THERE WILL BE ONE PRIMARY TRANSLATOR USING THE MEMOQ SYSTEM, WE MAY HAVE ONE OR TWO OTHERS WHO ALSO NEED ACCESS TO MEMOQ INTEGRATION.

Tier 1 – Translator-Facing MT (memoQ Integration) For reviewed translations performed by the agency’s qualified human translator, we propose using Google Cloud Translation – Advanced via the memoQ-supported Google MT plugin. This tier supports:

Bidirectional Spanish–English translation within memoQ

Use of existing memoQ Translation Memory content and IBWC-provided texts

Initial customization through data cleaning, glossaries, and adaptive translation datasets for IBWC-specific terminology and subject matter

Formatting and document structure preserved by memoQ

This tier provides higher-quality, domain-aware translations suitable for formal workflows, while keeping the human translator fully in control of review and final output.

Tier 2 – Stand-Alone GIST MT For non-translator staff requiring rapid comprehension of short texts (e.g., emails or brief correspondence), we propose a separate, stand-alone GIST translation capability. This tier is:

Web-based

Intended strictly for ad-hoc, unreviewed “gist” translations

Implemented with fixed-capacity infrastructure to avoid per-character or per-request cost growth

Protected by access controls, input size limits, and rate limits consistent with the limited user population described in the SOW

This separation ensures that casual or intermittent usage does not drive unpredictable cloud API costs while still meeting the SOW requirement for a stand-alone translation method.

FFP Risk and Cost Guardrails As you are aware, Google Cloud Translation – Advanced is usage-metered by characters, pages, and model type. While this is appropriate for controlled translator workflows, it introduces risk under a firm-fixed-price contract if usage volumes cannot be reasonably forecast.

To mitigate this risk, our proposed approach:

Limits Cloud Translation – Advanced usage to memoQ-based translation by the designated human translator

Routes ad-hoc GIST translation traffic to a fixed-cost, locally hosted MT capability

Assumes usage consistent with one primary memoQ user and intermittent translation projects, as described in your responses (WHILE THERE WILL BE ONE PRIMARY MEMOQ USER, THERE MAY BE A COUPLE OF ADDITIONAL INTERMITTENT USERS)

We would appreciate confirmation on whether this two-tier structure is acceptable to IBWC as a means of delivering required functionality while maintaining predictable costs under FFP. SEE COMMENTS ABOVE.

Authentication Preferences (GIST MT) To prevent uncontrolled expansion of stand-alone GIST usage, authentication is required. Given the small user population (approximately six users), we would like to confirm IBWC’s preference for one of the following:

Microsoft Entra ID (Azure AD) authentication using IBWC-managed accounts

Federated login using Microsoft or Google email identities

CAC/PIV or another authentication mechanism preferred by IBWC

In alignment with federal cybersecurity mandates including Executive Order 14028, OMB Memorandum M‑22‑09 (Federal Zero Trust Strategy), and NIST SP 800‑63 guidelines, IBWC requires the use of phishing-resistant multifactor authentication (MFA) for all systems and services.

Based on these requirements, IBWC’s preferred authentication mechanism is:

CAC/PIV or other phishing-resistant MFA methods (e.g., such as FIDO2 and Web Authentication-based authenticators).

The selected authentication method will affect implementation complexity and cost, so guidance on this point would be appreciated.

FFP Flexibility Finally, given that human translator usage volumes cannot be precisely forecast and that Cloud Translation – Advanced pricing varies by model and API method, we request clarification on whether IBWC is open to:

An FFP structure based on stated usage assumptions and guardrails, or

An alternative structure (such as FFP with defined usage thresholds and bilateral modification if materially exceeded)

EITHER OF THESE APPROACHES IS ACCEPTABLE.