TLDR
Executive briefThree polished bids can answer every adjective and none of the work
- Write the RFP as an evaluation design: each material requirement names a claim, context, artefact, exercise, threshold, owner and consequence.
- Give eligible bidders equal opportunity to prove the same outcome through controlled cases, frozen configurations and recorded clarification.
- Escalate evidence proportionately from assertion to disclosure, inspectable artefact, reproducible witnessed test and operating receipt.
- Simulate a real operating day with identities, data, tools, people, queues, failure, recovery, load, cost and an unexpected change.
- Carry every winning claim, qualification, gap and threshold into delivery, acceptance, remedies, change control and exit without dilution.
Three bidders promise an accurate, explainable, responsible, secure and enterprise-grade assistant. Each response cites a benchmark, a policy, a security certificate and a polished demonstration. None states whose work must finish, which population matters, when the system must abstain, how retrieval entitlement is enforced, what a reviewer sees, what happens after a tool timeout or which external state proves completion. The tender has measured fluency in bid writing.
Abstract requirements survive because they feel flexible. Suppliers can interpret them generously, procurement panels can score them quickly and hard operating questions can be deferred to delivery. The ambiguity is commercially dangerous. One bidder means model accuracy, another means task completion and a third means reviewer agreement. All can claim compliance. After award, each side discovers that its private definition was never priced, tested or assigned.
The Reality-Test RFP turns the tender into a staged experiment. A Problem Truth Sheet establishes what is known and what the buyer owns. A Requirement-to-Receipt Matrix expresses every material need as a seven-field Reality Clause. A Bidder Evidence Ladder distinguishes assertions from operating receipts. An Equal-Case Tender Lab gives shortlisted bidders the same representative challenge. An Operating-Day Simulation tests the whole workflow under pressure. A Reality Gap Register owns what cannot yet be reproduced. A Bid-to-Contract Trace prevents winning evidence from evaporating after award.
The strongest rival is right about one risk: production-grade evidence demanded from every bidder can entrench incumbents, impose unpaid implementation, disclose intellectual property and reward large compliance teams. No tender can recreate every production dependency. The answer is a Progressive Disclosure Boundary. Begin with short comparable declarations. Release the same safe context to eligible bidders. Pay shortlisted suppliers when realistic work is material. Protect confidential artefacts through controlled inspection. Permit equivalent proof. Record irreducible gaps. The objective is fair reduction of uncertainty, not maximal paperwork.
Problem Truth SheetExpose the buyer's reality before asking suppliers to solve it
The truth sheet begins with the consequential decision or work outcome. Name the users and affected population, current baseline, volume, service window, constraints and unacceptable consequences. Describe the non-AI comparator. If process simplification, deterministic automation or additional operating capacity can meet the need, bidders should not be forced to wrap those answers in an AI label.
Buyer-owned dependencies receive equal prominence. These include data rights and quality, policy clarity, system access, reviewer capacity, adoption, language coverage, identity services and the ability to change an existing workflow. A supplier cannot price or accept an outcome obligation against conditions that remain hidden. An incumbent's private knowledge of undocumented defects is also an information advantage, so relevant known conditions are disclosed equally or isolated through a controlled test.
Unresolved assumptions are not written as facts. The sheet records who will resolve them, by which event and before which gate. If the buyer does not know whether required data can lawfully be shared, the tender cannot promise it. If the current baseline is unreliable, improvement thresholds wait for discovery. This protects suppliers from impossible promises and the buyer from optimistic bids that quietly assume clean data and unlimited human support.
The sheet does not freeze solution design prematurely. It establishes the problem physics and the boundaries of acceptable proof. Suppliers remain free to propose different models, rules, process changes, human services and integration designs. The more precisely the buyer states outcome, constraints and evidence need, the more genuinely solution-neutral the competition becomes.
Requirement-to-Receipt MatrixGive every material claim a discriminating test
The Reality Clause has seven fields. The claim states what the solution must accomplish or prevent. The context binds population, workflow, data, volume, authority and constraints. The artefact names what the bidder supplies before the exercise, such as a configuration record, data-flow map, control design or case log. The exercise describes what the buyer or independent witness will observe.
The threshold is declared before the test and separates pass, conditional pass and failure. The owner identifies who accepts remaining uncertainty and who must close it. The consequence is commercial: exclusion, clarification, paid discovery, priced obligation, acceptance hold, remedy or rejection. A requirement without consequence is often optional in practice, regardless of its wording.
Consider “the system must be accurate”. A useful clause might require the proposed configuration to complete a defined claims-triage task across specified language and documentation slices, abstain when evidence is insufficient and avoid a prohibited denial. The bidder supplies version and configuration artefacts. Every shortlisted bidder runs the same hidden cases. Thresholds cover task result, harmful error, abstention and review burden. The risk owner can accept only a named gap. Failure blocks award for that use.
The matrix keeps claims separate. A generic security certificate cannot compensate for failed recovery. Strong average quality cannot trade against a prohibited harmful effect. Attractive presentation cannot erase missing identity evidence. Weighted scoring can still compare negotiable dimensions, but non-negotiable gates remain conjunctive and visible.
| Reality Clause field | Question | Recorded example |
|---|---|---|
| Claim | What must the solution accomplish or prevent? | Complete eligible cases without prohibited commitment |
| Context | For which people, work, data, volume and authority? | Three languages, incomplete records, peak queue and bounded tools |
| Artefact | What is supplied before exercise? | Frozen configuration, data flow, tool contract and escalation design |
| Exercise | What will be directly witnessed? | Hidden cases plus timeout, load and change event |
| Threshold | What is pass, conditional pass or fail? | Declared quality, harm, abstention, latency and recovery gates |
| Owner | Who owns residual uncertainty and closure? | Named business risk owner and supplier delivery owner |
| Consequence | What happens commercially? | Clarify, discover, price, hold, remedy, reject or exit |
Bidder Evidence LadderAsk for the lowest burden that can answer the claim
Level one is an assertion: a bidder says a capability exists. It supports screening but rarely proves a material claim. Level two is structured disclosure: a completed response, model card, policy, certificate or FactSheet makes declared facts comparable. It can reveal a gap, but generic documentation does not establish fitness for this configuration and workflow.
Level three is an inspectable artefact: the panel examines a configuration, test record, architecture, data lineage, interface, control design or subcontractor list. Confidential inspection can protect legitimate intellectual property. The buyer asks for source code only when code access is necessary to answer the claim and a less intrusive form of evidence cannot do so.
Level four is a reproducible witnessed test. The supplier or buyer runs a frozen configuration against declared cases and conditions while the panel observes inputs, outputs, controls and end states. Level five is an operating receipt: evidence from a realistic or production-like resource confirms the external effect, intervention, recovery, cost and human work. Higher levels are not universally better. They answer narrower, more consequential claims with stronger context.
The Progressive Disclosure Boundary stages the ladder. All bidders receive concise requirements and submit comparable declarations. Eligible suppliers receive equal safe context. Shortlisted bidders obtain deeper case packs and controlled data. Material exercises are paid when they require meaningful configuration or integration. Confidential artefacts stay in clean rooms or witnessed sessions. Evidence is retained only to the degree necessary for award, challenge, acceptance and accountability.
| Level | Useful for | Cannot prove alone |
|---|---|---|
| 1 · Assertion | Initial market signal and clarification target | Any consequential operating claim |
| 2 · Structured disclosure | Comparable declared facts, policies, certifications and known limits | Buyer-specific task fitness or control operation |
| 3 · Inspectable artefact | Configuration, design, lineage, tests and provider-chain evidence | Actual behaviour under the buyer's cases |
| 4 · Reproducible witnessed test | Behaviour and control under frozen common conditions | Every unavailable production dependency |
| 5 · Operating receipt | External effect, human work, load, recovery and cost | Future unchanged performance without revalidation |
Equal-Case Tender LabMake competitive proof comparable without making solutions identical
The lab begins with a common case architecture. It includes representative ordinary cases, meaningful population slices, edge conditions, ambiguous records, prohibited actions, dependency failures and load. Some cases are disclosed so bidders can configure responsibly. Others remain hidden to reduce overfitting. Test contamination rules prohibit training on hidden answers, manual substitution that is not part of the proposed service and unrecorded configuration changes.
Every bidder receives the same material context at the same stage. Clarification questions and answers are recorded and shared when they affect the competition. An incumbent's knowledge is not erased, but known data defects, process exceptions and environment constraints are disclosed equally. When a fact cannot be published safely, the buyer can expose it inside a controlled environment or adjust the exercise so all suppliers face the same consequential condition.
Configurations are frozen and attributable. Record model and version, prompts, retrieval, tools, policy, identity, human support, infrastructure, region and subcontractors. A supplier-selected demonstration may still be useful for discovery, but award evidence comes from common cases. The panel preserves case-level results and receipts, not only a presentation and aggregate score.
Fairness is not a race to impose the same documentation. A smaller supplier may show a witnessed black-box test where an incumbent offers extensive compliance artefacts. If both answer the material claim, the proof can be equivalent. If a realistic integration exercise requires substantial work, shortlisted bidders are paid. Free implementation disguised as a proof shifts buyer cost to the market and filters for balance-sheet capacity rather than solution fitness.
| Case family | Equal condition | Evidence retained |
|---|---|---|
| Representative work | Same population, language, data quality and completion definition | Case result, abstention, escalation and end-state receipt |
| Edge and prohibited | Same ambiguity, missing evidence and unacceptable consequence | Control decision, rationale and blocked effect |
| Human interaction | Same role, information, authority, queue and time pressure | Reviewer action, override, delay and decision receipt |
| Failure and recovery | Same timeout, duplicate risk, unavailable dependency and restart | Retries, postcondition, reconciliation and recovery proof |
| Load and change | Same volume profile, tail condition and mid-session update | Latency, cost, capacity, revalidation and restriction |
Operating-Day SimulationTest the workflow that the contract will actually support
Isolated answer tests miss the system around the answer. The operating day begins with identity and authorised task origin. It retrieves data under actual entitlement rules, invokes the proposed model and tools, presents or executes work through the stated human authority and records the receiving-system end state. Queues, reviewer capacity, exceptions and escalation are part of the system, not environmental noise.
The simulation introduces realistic operating events. A retrieval source becomes stale. A tool times out after a possible commitment. Demand rises. A reviewer disagrees. A model or policy changes mid-session. A subcontractor is unavailable. The supplier must show abstention, bounded retry, independent postcondition, recovery, monitoring and change handling. Speed is measured with tail and exception conditions, not only a comfortable average.
Cost follows the same events. Record inference, retrieval, tools, human review, logging, support, incident work and recovery. The cheapest headline price may create the most expensive operating day when escalations and retries dominate. The exercise does not need to reproduce every production dependency, but it must reveal which cost and control assumptions depend on what is missing.
Human interaction is designed, not staged as a ceremonial approval. Reviewers see source evidence, uncertainty, proposed effect and alternatives. They have time, capacity and authority to reject. Override and escalation become receipts. A vendor that requires heroic human supervision has proposed a different service and cost structure from one whose difficult cases are bounded and legible.
Reality Gap RegisterCarry every untested production difference to an owned closing event
A test environment cannot reproduce every production identity, data volume, latency path, user behaviour, integration, jurisdiction or incident condition. The register names each difference, why it exists, which requirement it affects and which proxy evidence is used. It records uncertainty, compensating control, owner, closing event and latest date. A gap shared by all bidders is still a gap in the buyer's decision.
Proxy evidence must match the claim. Synthetic data can test flow and some privacy controls but may not represent language or rare-case quality. A black-box exercise can answer observable task and control claims even when source artefacts are unavailable, but it cannot prove hidden training-data facts. A certificate can show that an assurance process occurred, not that the proposed workflow recovers from a duplicate action.
Gaps change the commercial route. A bounded uncertainty may become a priced acceptance condition. A material integration unknown may require paid discovery. A missing buyer prerequisite can stop procurement. A non-negotiable harm or recovery gap blocks award. The register prevents “to be confirmed in implementation” from becoming an unlimited transfer of uncertainty to delivery.
| Production difference | Proxy and uncertainty | Control, owner and closing event |
|---|---|---|
| Live entitlement service unavailable | Recorded identity decisions; concurrent-load behaviour unknown | Restricted cohort; identity owner closes before acceptance |
| Sensitive multilingual records withheld | Clean-room sample; rare dialect coverage remains uncertain | Paid witnessed test; service owner closes before award |
| Foundation-model artefacts prohibited | Frozen black-box test answers behavioural claims only | Version pin and change notice; risk owner narrows intended use |
| Peak human queue not reproducible | Load model plus sampled reviewers; fatigue not observed | Capacity ceiling; operations owner closes during pilot |
| Migration path untested | Export schema inspected; timing and open-case transfer unknown | Exit rehearsal; procurement owner closes before renewal |
Bid-to-Contract TraceDo not let the winning configuration disappear after award
Winning evidence is attached to a versioned bid record. It names the proposed model, data, prompts, tools, humans, infrastructure, subcontractors, test cases, qualifications and thresholds. The award decision cites which receipts supported each material claim and which gaps were accepted. Clarifications that changed the offer become part of the same record.
The trace carries those facts into deliverables, acceptance criteria, service levels, change control, remedies and exit. A claimed witnessed recovery becomes an acceptance exercise. A promised language slice becomes a monitored service condition. A disclosed model dependency becomes a notice and revalidation obligation. A qualification becomes an explicit buyer dependency or a priced supplier task, not a footnote that delivery teams discover later.
Contract drafting cannot recover evidence the tender never asked for. Paper 128 will address durable change, incident and evidence clauses. The RFP's narrower job is to ensure the commercial schedule has something concrete to preserve. It also maintains defensibility: the selected bidder is awarded for evidence against published claims and thresholds, while non-negotiable failures cannot be disguised by a composite score.
No award is a valid result. If every bidder fails a prohibited harm or recovery threshold, the panel stops. It may repair the requirement, run discovery, buy a component under buyer integration, choose a non-AI service or return to the market. Selecting the least-bad bid converts a declared gate into theatre and teaches suppliers that critical requirements are negotiable after all.
Ten worked casesRun the claim, evidence, equality, gap and contract tests
1. Three claims of high accuracy
Only one bidder names the target population, abstention rule and buyer-relevant error cost. The Problem Truth Sheet fixes those facts for all. A Reality Clause separates completion, prohibited error and abstention. Assertions remain level one. Shortlisted bidders run equal hidden cases, and the winning thresholds travel into acceptance rather than becoming one weighted “accuracy” score.
2. Supplier demo meets the easy cases
A bidder performs well on its chosen set and poorly on unseen multilingual and low-documentation work. The equal-case lab records the frozen configuration and case-level result. The field receives a competitive paid discovery where material language work is compensated. The gap closes only when representative evidence supports a declared threshold; a polished demonstration cannot be substituted.
3. The assistant answers but cannot complete work
A conversational assistant passes isolated answer tests. During the operating day, retrieval ignores entitlements and escalation has no owner. The requirement is restated around authorised case completion. Identity, data, human authority and end-state receipts enter the matrix. Procurement stops until non-negotiable access and escalation prerequisites are evidenced.
4. Tool timeout causes a duplicate action
An agent completes quickly until a tool times out after an irreversible action may have committed. It retries and duplicates the effect. The recovery clause requires a stable transaction identity, resource-side postcondition, bounded retry and reconciliation. Failure is not offset by speed or presentation. Every bid failing the gate produces no award.
5. Certifications and policy without configuration proof
A supplier presents ISO certifications and a responsible-AI policy but no evidence for the proposed model, retrieval, data and workflow. The documents sit at disclosure level. The panel requests inspectable configuration artefacts and a witnessed buyer-case exercise. The bid may proceed to paid discovery, but it has not proven operating fitness.
6. A small supplier cannot fund an unpaid proof
The smaller firm has strong task evidence but cannot absorb production-like integration requested from all bidders. The Progressive Disclosure Boundary narrows early requirements, shortlists on proportionate proof and pays each finalist for the material exercise. Equal opportunity to demonstrate the claim protects competition better than identical unpaid implementation.
7. The incumbent knows hidden data defects
An incumbent has learned undocumented defects during an earlier contract. New entrants receive apparently equal files but face different information. The buyer records the defects on the truth sheet, shares relevant clarification and uses hidden cases that test handling rather than secret knowledge. If equality cannot be repaired, the procurement pauses and the decision record explains why.
8. Upstream artefacts are contractually unavailable
A foundation-model provider prohibits requested internal artefacts. The buyer asks which claim actually needs them. A frozen, witnessed black-box test can answer behaviour, control and recovery questions, while the visibility gap limits hidden-data claims. The component can remain eligible if the use is bounded, change is detectable and equivalent proof meets the published threshold.
9. Lowest bid, highest operating day
The lowest headline price drives many reviews, tool calls, escalations and support events. The simulation records cost at normal and tail load, including exceptions and recovery. The Reality Clause prices the complete service conditions. The apparent saving disappears in comparable receipts, and the award trace preserves operating-cost assumptions and change triggers.
10. The process alternative wins and the AI field fails
No AI proposal passes the non-negotiable recovery gate. A service redesign with deterministic routing meets the actual outcome, preserves reasons and costs less to operate. The panel refuses award and reframes the requirement around the non-AI service. A legitimate no-AI or no-winner outcome is evidence that the tender tested reality rather than managed a predetermined purchase.
Interactive decision artefactExecutable RFP Studio
The Studio selects one procurement route and generates the evidence boundary behind it. It does not produce a generic RFP, checklist or compensating vendor score.
Tender cadenceMove from problem truth to defensible award or deliberate stop
Before publication, a cross-functional team completes the Problem Truth Sheet, identifies consequential uses, defines non-negotiable controls and separates buyer-owned prerequisites from supplier obligations. It writes Reality Clauses and evaluation criteria before proposals arrive. Market engagement tests whether evidence burdens are feasible and whether alternative solutions should remain open.
At expression of interest, bidders receive short comparable questions. The buyer identifies eligible solutions without demanding deep confidential artefacts. Clarifications that affect material context are shared. Evidence burden rises with consequence, uncertainty and stage. The panel documents conflicts and assigns independent witnesses where needed.
At shortlist, the Equal-Case Tender Lab freezes configurations and runs common cases. Paid discovery covers material integration or data work. The operating-day simulation observes workflow, human authority, failure, load, cost, recovery and change. Case-level evidence stays inspectable, while protected bidder material remains within agreed confidentiality controls.
At award, the panel applies pre-declared gates. Non-negotiable failures cannot be compensated by price or presentation. Every accepted gap has an owner and closing event. The decision record cites the winning receipts and explains why alternative proof was equivalent. No award remains available when the evidence does not support a safe, accountable choice.
At contract and acceptance, the Bid-to-Contract Trace turns claims, configurations, qualifications and thresholds into deliverables, exercises, service conditions, change triggers, remedies and exit. Acceptance repeats the relevant operating evidence in the delivered environment. A material change expires the affected proof rather than inheriting the bid's confidence indefinitely.
After delivery or cancellation, the buyer records which assumptions, cases, thresholds, costs and contractual mechanisms proved useful. Lessons from weak and discontinued procurements are part of the acquisition asset. They improve the next truth sheet and tender lab instead of disappearing with the project team.
Primary evidenceReferences
Links and live status were checked on 2 September 2026. OMB M-25-22 dated April 2025 remains the current cited federal AI acquisition memorandum, and the GAO evidence was published in April 2026. NIST AI RMF 1.0 remains current while revision is under way. The Reality-Test RFP, five connected instruments, Reality Clause, Bidder Evidence Ladder, Reality Gap Register, Progressive Disclosure Boundary and Executable RFP Studio are this article's synthesis.
- UK Government, Guidelines for AI Procurement.
- U.S. General Services Administration, Buy AI.
- U.S. Government Accountability Office, Artificial Intelligence Acquisitions.
- U.S. Office of Management and Budget, M-25-22: Driving Efficient Acquisition of Artificial Intelligence in Government.
- NIST, AI RMF Playbook.
- NIST, The TEVV-Athlon Framework for Evaluating AI Systems.
- NIST, Artificial Intelligence Risk Management Framework 1.0.
- NIST AI 600-1, Generative Artificial Intelligence Profile.
- Treasury Board of Canada Secretariat, Directive on Automated Decision-Making.
- Innovation, Science and Economic Development Canada, Implementation Guide for Managers of Artificial Intelligence Systems.
- UK Government, The Digital, Data and Technology Playbook.
- CISA and partners, Choosing Secure and Verifiable Technologies.
- European Commission, AI Act regulatory framework.
- OECD, Governing with Artificial Intelligence.
- Johnson, R. et al., Disclosure or Marketing? Analyzing the Efficacy of Vendor Self-reports for Vetting Public-sector AI.
- Zick, T. et al., AI Procurement Checklists: Revisiting Implementation in the Age of AI Governance.
- Bähr, S. et al., Human-AI Interaction Requirements in Public Sector Procurements.
- Hickok, M., Public procurement of artificial intelligence systems: new risks and future proofing.
About this paperTurn tender language into observable commercial evidence
The Reality-Test RFP, Problem Truth Sheet, Requirement-to-Receipt Matrix, Equal-Case Tender Lab, Operating-Day Simulation, Bid-to-Contract Trace, seven-field Reality Clause, Bidder Evidence Ladder, Reality Gap Register, Progressive Disclosure Boundary and Executable RFP Studio are this article's synthesis. They adapt needs-first procurement, performance-based acquisition, contextual evaluation and fair competition into an executable tender method. The design does not treat adjectives, benchmarks, demonstrations, policies, certifications, source-code demands or composite scores as buyer-specific operating proof.