Jev call copilot CurrencyTransfer research and architecture built 2026-09-27

Latest evaluation run

Run final-20260927, recorded 2026-09-27T07:19:28.104Z against Worker version 2f03062938a0, layers L4 and L6, replayed at 20× with fresh model answers. Rendered from eval/reports/final-20260927.l4l6.json by scripts/render-eval-report.mjs; the newest committed run report is the one published.

Summary

layer result checks passed
L4 replay against labels PASS 33/33
L6 cost and latency PASS 6/6
L1 customer_success cases 79/81 cases (97.5%) 164/166
L1 onboarding cases 91/91 cases (100.0%) 200/200

Policies

scenario policy version expected model
customer_success 4 jev-1.13.0
onboarding 4 jev-1.13.0

Replayed calls

role call scenario decision points requests errors cards shown completeness next step agreed
onboarding_strong 3339895706 onboarding 96 96 0 48 0.286 yes
onboarding_weak 3485591407 onboarding 50 50 0 32 0.143 no
cs_strong 3303259297 customer_success 131 131 0 111 – –
cs_weak 3347356034 customer_success 22 22 0 22 – –

Completeness is must-say items done over items applicable for onboarding; customer-success calls have no must-say list.

L4: replay against labels

role check result detail
onboarding_strong finished PASS every utterance answered
onboarding_strong zero_errors PASS errors 0
onboarding_strong one_model PASS models [jev-1.13.0]
onboarding_strong requests_eq_decision_points PASS requests 96, decision points 96
onboarding_strong card_shown_share PASS card shown on 48/96 (50.0%)
onboarding_strong recall PASS recall 2/2 (1.000, ≥ 1)
onboarding_strong precision PASS precision 1.000, no id outside said ∪ not_applicable persisted
onboarding_strong completeness PASS engine completeness 0.286 (done 2 / applicable 7 = 0.286), label completeness 0.286
onboarding_strong next_step_agreed PASS next_step_agreed persisted true, labelled true
onboarding_strong risk_flags PASS risk flags [], labelled []
onboarding_weak finished PASS every utterance answered
onboarding_weak zero_errors PASS errors 0
onboarding_weak one_model PASS models [jev-1.13.0]
onboarding_weak requests_eq_decision_points PASS requests 50, decision points 50
onboarding_weak card_shown_share PASS card shown on 32/50 (64.0%)
onboarding_weak recall PASS recall 1/1 (1.000, ≥ 1)
onboarding_weak precision PASS precision 1.000, no id outside said ∪ not_applicable persisted
onboarding_weak completeness PASS engine completeness 0.143 (done 1 / applicable 7 = 0.143), label completeness 0.143
onboarding_weak next_step_agreed PASS next_step_agreed persisted false, labelled false
onboarding_weak risk_flags PASS risk flags [], labelled []
onboarding_weak below_strong PASS completeness 0.143 (< strong's 0.286; labels 0.143 < 0.286)
cs_strong finished PASS every utterance answered
cs_strong zero_errors PASS errors 0
cs_strong one_model PASS models [jev-1.13.0]
cs_strong requests_eq_decision_points PASS requests 131, decision points 131
cs_strong card_shown_share PASS card shown on 111/131 (84.7%)
cs_strong resolved PASS resolution resolved, reached [owned, reported, resolved]
cs_weak finished PASS every utterance answered
cs_weak zero_errors PASS errors 0
cs_weak one_model PASS models [jev-1.13.0]
cs_weak requests_eq_decision_points PASS requests 22, decision points 22
cs_weak card_shown_share PASS card shown on 22/22 (100.0%)
cs_weak never_owned PASS reached [reported]

L6: cost and latency

measure value
requests 299 (0 without usage)
max input tokens 10,511
p95 input tokens 9,395
p95 provider latency 498 ms
p95 end-to-end latency 457 ms over 22 decision points of 3347356034 at 1×
highest cost per call $0.0396
check result detail
requests_measured PASS 299 requests with usage
max_input_tokens PASS max 10511 (≤ 12000)
p95_latency PASS p95 provider latency 498 ms (≤ 1500)
cost_per_call PASS 3303259297 $0.0396, 3339895706 $0.0345, 3347356034 $0.0068, 3485591407 $0.0172 (each ≤ $0.05)
model_expected PASS models [jev-1.13.0] (== jev-1.13.0)
e2e_p95 PASS e2e p95 457 ms (gate ≤ 3000, target ≤ 1500) over 22 decision points of 3347356034 at 1×

L1 customer_success: single-decision cases

Run final-20260927-cs, policy version 4, model jev-1.13.0: 79 of 81 cases pass (97.5%), 164 of 166 checks.

failing case failed checks
lfl_cs_payment_mentioned_no_amount fits::payment_mentioned_no_amount >= 0.6
lfl_cs_several_payments_at_once fits::several_payments_at_once >= 0.6

L1 onboarding: single-decision cases

Run final-20260927-onboarding, policy version 4, model jev-1.13.0: 91 of 91 cases pass (100.0%), 200 of 200 checks.

896 words · Internal working documents. Do not share outside CurrencyTransfer.