Named, inspectable tests
REPRODUCIBLE DECISION CONTRACT / CL-QB-1.0A cheap answer is useless
A cheap answer is useless
if it is wrong.
Across 32 synthetic adversarial cases: 0 false-qualified and 4/4 valid controls passed. CapacityLine qualifies only commitments that survive buyer policy, identity, transcript grounding, read-back, and consistency. It then stress-tests the fallback—and never releases an order.
In the published synthetic set
Prevents a trivial block-everything result
Applied to each commitment
SAME INPUTS · DECLARED EXPECTATIONS / CL-CB-1.0
What happens when you leave the evidence out?
168 synthetic cases. 6 policy families. Equal numbers of valid and invalid commitments. Every method sees the same inputs; refusing everything is measured as a failure to recover valid options.
| Method | Unsafe offers accepted Lower is better | Valid offers recovered Higher is better | Valid offers missed Lower is better |
|---|---|---|---|
| CapacityLineThe full production qualification evaluator. | 0 / 84 | 84 / 84 | 0 / 84 |
| Availability-only baselineAccepts a supplier's structured availability flag; no buyer-policy checks. | 78 / 84 | 84 / 84 | 0 / 84 |
| Structured-fields baselineChecks commercial fields, identity and validity; omits transcript grounding, final read-back, contradictions and voice quality. | 24 / 84 | 84 / 84 | 0 / 84 |
| Block-all controlRefuses everything, exposing the cost of a trivial zero-false-positive policy. | 0 / 84 | 0 / 84 | 84 / 84 |
Developer-authored, correlated regression fixtures—not an independent held-out study. No competing product, human team, phone-network performance or customer savings were measured.
QUALIFICATION GATE / RECOVERY FRONTIERQualify the fallback.
Qualify the fallback.
Then try to break it.
The same recorded commitments are replayed against harder operating conditions. In the deterministic reference case, every single shock remains covered and the first uncovered gap requires two concurrent conditions.
Inspect the comparison evidenceQuantity, deadline, price, exact-part, and primary-path tests remain covered.
Concurrent conditions—not a hidden confidence score—create the first uncovered gap.
The output is evidence for human review. CapacityLine creates no order, contract, or payment.
WHAT EARNS QUALIFICATION
Evidence is part of the decision—not decoration after it.
The evaluation is deterministic and fail-closed. Unknown stays unknown, missing data stays visible, and hard buyer rules cannot be traded for a higher aggregate score.
Supplier words, not generated prose
The evidence quote must occur in a supplier transcript turn. A plausible sentence invented after the call fails.
Named operating authority
An affirmative flag alone is insufficient. The respondent name and role must also be present.
Hard policy remains hard
Unapproved part, missing current certification, or disallowed origin cannot win on speed or price.
Final read-back and consistency
A commitment is not qualified without affirmative read-back and no material contradiction.
THE FAILURE SURFACE / ALL CASES
Inspect what the system refuses to overstate.
These are synthetic failure and uncertainty cases evaluated against the reference buyer policy. Each row names at least one check that prevented a false qualification.
Call result integrityMissing, malformed, or incomplete results never become supplier commitments.
3 CASESUnanswered callNo live commitment was obtained after the permitted call attempt.unreachable
Malformed provider resultNo live commitment was obtained after the permitted call attempt.unreachable
Incomplete structured resultRequired quantity · Ship-by date · Unit price ceiling · Approved part · Required certifications · Approved origin · Respondent authority · Quote validity · Transcript-grounded evidence · Final read-back confirmation · Statement consistencyineligible
Commercial commitmentQuantity, date, currency, price, availability, and validity must be explicit.
12 CASESUnknown availabilityFull availabilityreview
Explicitly unavailableFull availabilityreview
Partial availabilityFull availabilityreview
Missing quantityRequired quantityreview
Zero quantityRequired quantityreview
Insufficient quantityRequired quantityreview
Missing ship dateShip-by datereview
Late ship dateShip-by datereview
Missing priceUnit price ceilingreview
Non-positive priceUnit price ceilingreview
Wrong currencyUnit price ceilingreview
Expired quoteQuote validityreview
Buyer policyAn attractive offer cannot bypass part, certification, or origin policy.
4 CASESUnapproved partApproved partineligible
Ambiguous certificationRequired certificationsineligible
Expired certificationRequired certificationsineligible
Disallowed originApproved originineligible
Evidence & authorityNamed authority, grounded supplier words, confidence, read-back, and consistency are required.
10 CASESUnconfirmed authorityRespondent authorityreview
Missing respondent nameRespondent authorityreview
Missing respondent titleRespondent authorityreview
Missing transcript evidenceTranscript-grounded evidence · Final read-back confirmationreview
Missing evidence quoteTranscript-grounded evidencereview
Evidence quote absent from transcriptTranscript-grounded evidencereview
Low-confidence extractionTranscript-grounded evidencereview
Non-finite confidenceTranscript-grounded evidencereview
Negative final read-backFinal read-back confirmationreview
Conflicting statementsStatement consistencyreview
Conversation qualityStructurally complete output is still held for review when the live exchange is fragmented.
3 CASESLive result without quality assessmentNeeds buyer review for conversation quality.review
Repeated live openingNeeds buyer review for conversation quality.review
Fragmented live disclosureNeeds buyer review for conversation quality.review
HONEST LIMITS
This is a qualification test—not a claim of supplier truth or saved downtime.
- This benchmark tests deterministic qualification behavior, not phone-network reliability or supplier truthfulness.
- Synthetic fixtures are not customer outcomes and are not evidence of avoided downtime or savings.