Choosing What to Observe: Budget-Paired Counterfactual Audits for Credible Evaluation of Visual Agents
Organizations: Fudan University · Institute of Medical Philosophy & Future AI (IMPF-AI) · The University of Warwick · Southern University of Science and Technology
Abstract
Visual agents that read dense documents must ration a finite observation budget across candidate regions, yet answer accuracy conflates where an agent looks, how it adapts to the reads that remain, and how it an?swers once observation stops. We introduce budget-paired counterfactual audits, an evaluation protocol that holds public evidence and region bindings fixed while perturbing only the remaining budget or de?clared task parameters, and credits every acquired history with exact continuation values. The resulting accounting decomposes lost expected accuracy into stepwise acquisition loss and terminal decision loss, separating planning from fixed-role heuristics. Across forty paired routing anchors, five open-weight
Figures & tables
| Sequential | Counterfactual | Budget/value | Exact per-step | Reference | |
| Work | observations | interventions | paired states | loss | scoring audit |
| AFABench / CUBE-NM [ 8 ] | ✓ | ||||
| FOVEA [ 6 ] | ✓ | ||||
| Visual tool-use audit [ 13 ] | ✓ | ✓ | |||
| OPO [ 10 ] | ✓ | ✓ | |||
| Ours (budget-paired audits) | ✓ | ✓ | ✓ | ✓ | ✓ |
| Condition | Horizon | Fixed S | Threshold |
|---|---|---|---|
| Fair, two reads | S | S | S |
| Fair, one read | B | S | S |
| Noisy, reliability .50 | B | S | B |
| Reliability .74, P1–P2 | B | S | S |
| Reliability .74, P3–P4 | S | S | S |
| Configuration | S/B/D | S/B/D | Anchor loss | Other loss |
|---|---|---|---|---|
| Qwen3-VL-4B | 26/6/8 | 24/7/9 | 15.62 | 0.85 |
| Qwen3-VL-8B | 40/0/0 | 39/0/1 | 12.81 | 0.64 |
| Qwen3-VL-32B | 40/0/0 | 40/0/0 | 12.50 | 0.60 |
| InternVL3.5-8B | 40/0/0 | 40/0/0 | 12.50 | 0.59 |
| MiniCPM fast | 37/1/2 | 35/3/2 | 13.75 | 0.62 |
| Fixed selector | 40/0/0 | 40/0/0 | 12.50 | 0.60 |
| Model | Input | S | S | Optimal transitions |
|---|---|---|---|---|
| GPT-5 mini | Image + map | 0/48 | 48/48 | 48/48 |
| GPT-5 mini | Map only | 0/48 | 48/48 | 47/48 |
| GPT-4.1 | Image + map | 48/48 | 48/48 | 0/48 |
| GPT-4.1 | Map only | 4/48 | 48/48 | 44/48 |
| Panel A: Paired document contrasts | ||
|---|---|---|
| Reference | One crop uniform (pp) | Two crops native (pp) |
| Qwen3-VL-8B | ||
| Curated | 24.44 [8.89, 41.11] | -13.33 [-33.33, 6.67] |
| Original | 10.56 [-0.56, 24.44] | 0.00 [-13.33, 13.33] |
| InternVL3.5-8B | ||
| Curated | 5.00 [-1.67, 16.67] | -10.00 [-26.67, 0.00] |
Appendix figures & tables48 assets
Supplementary material from the paper’s appendix.
Appendix
| Symbol | Meaning |
|---|---|
| Fixed public context, role binding and answer rule. | |
| Latent scene and its conditional prior. | |
| True region bit and answer . | |
| True-one prior and read correctness for region . | |
| Region selected at step and tool return. | |
| Declared conditional observation kernel. |
| Question | Evidence | Supported interpretation |
|---|---|---|
| Conditional value | Five pinned open-weight VLMs; forty families and eight fixed parameter profiles | Reliability sensitivity and adaptation to a moving optimum are distinct. |
| Budget use | Forty high-gap paired anchors per configuration | No correct branch-to-selector transition under the tested direct-answer settings. |
| Complete observation policy | Three checkpoints; eight-region, three-read trajectories on forty families | Low first-action loss can coexist with later acquisition and terminal-use errors. |
| Natural visual evidence | Two eligible checkpoints on thirty real forms; six matched crop readers and whole-image controls | Crop selection and system accuracy differ; results depend on reference scoring. |
| Prospective terminal repair | Six fixed histories; development-locked diagnosis groups and two presentation controls | The selected repair predictions were not confirmed; absent diagnosis classes remain unidentifiable. |
| Presentation and reader controls | Historical paired inputs and shared-reader interventions | Exploratory effects locate stages of loss; they do not establish universal repair or isolated image causality. |
| Protocol / model | Reasoning / output | Split and reuse | Scheduled units | Actual requests | Outcome handling |
|---|---|---|---|---|---|
| Factorial / GPT-4.1 | Not requested; strict enum | 6 dev, 48 new confirmation families; 9 states, 6 arms | 2,592 | 2,923 | 2,592 legal; no repairs or unresolved missing |
| Factorial / GPT-5 mini | Endpoint default; strict enum | Same frozen family/state/arm design | 2,592 | 2,592 | 2,591 legal, 1 invalid; no repairs or missing |
| Moving boundary / GPT-4.1 | Not requested; strict enum | 6 dev families, 4 dev profiles; 24 new families, 4 held-out profiles, 9 levels/profile | 864 | 880 | 864 legal; no repairs or unresolved missing |
| Moving boundary / GPT-5 mini | Low effort; strict enum | Same frozen held-out boundary design | 864 | 864 | 864 legal; no repairs or unresolved missing |
| Computed-value follow-up / GPT-4.1 | Not requested; strict enum | Reuses 96 boundary states at ; two new aid arms | 192 | 192 | Baseline 96 states reused, not new independent tests |
| Downstream crop follow-up / GPT-4.1 | Not requested; strict enum | 24 reused masked families, 2 rules, paired arms; common first action/read | 48 episodes | 348 | 338 scored steps, all legal; 98 distinct crop reads |
| Evidence source | Families | Experimental conditions |
|---|---|---|
| Historical Sonnet closed loop | 30 | Two budgets on the same families. Retains selection order, visible values, answer and ground truth. |
| Historical Sonnet paired actions | 24 | Rule, selector assignment, budget and forced history in a separate archive. |
| Historical Qwen3.5-4B holdout | 24 | One shared map per family, 72 independent rule checks and 192 direct decisions. Greedy and horizon controllers replay the 192 states. |
| Frozen Sonnet decision panel | 10 | Eight fair starts per family (2 assignments 2 rules 2 budgets), two unequal-prior image starts, one map-only start at budget two, four forced-history decisions, one binding probe and two rule probes. |
| Forced-tool sensitivity | Same 10 | One first-action request per image and budget for the 20 unequal-prior image cells. Sessions remain separate. |
| Stimulus | Labels at | Base roles |
|---|---|---|
| 1 | C, A, X, B, Y | C, A, B |
| 2 | Y, B, A, X, C | B, A, C |
| 3 | B, Y, C, A, X | C, A, B |
| 4 | Y, A, X, B, C | B, C, A |
| 5 | B, X, A, Y, C | C, B, A |
| 6 | B, X, C, Y, A | B, C, A |
| First action | Either distractor | |||
|---|---|---|---|---|
| Fair switch | .50 | .75 | .75 | .50 |
| Fair switch | 1.00 | .75 | .75 | .75 |
| Fair XOR | .50 | .50 | .50 | .50 |
| Fair XOR | .50 | 1.00 | 1.00 | .50 |
| Unequal-prior entropy reduction (bits) | .04616 | .15225 | .11811 | 0 |
| Unequal-prior | .820 | .820 | .864 | .820 |
| Condition | Inv. | Miss. | ||||
|---|---|---|---|---|---|---|
| Free text, both | 20 | 5 | 0 | 0 | 15 | 0 |
| Tool, | 10 | 3 | 4 | 1 | 0 | 2 |
| Tool, | 10 | 8 | 0 | 0 | 0 | 2 |
| Rule | Reference | ||
|---|---|---|---|
| Switch | 19/20 | 19/20 | Selector is suboptimal with one reveal and optimal with two. |
| Branch XOR | 0/20 | 0/20 | Selector is irrelevant at either budget. |
| Budget | Policy | Optimal first actions | Correct answers |
|---|---|---|---|
| 1 | Native Sonnet 4.6 | 24/30 | 21/30 |
| 1 | One-step reference | 30/30 | 22/30 |
| 1 | Remaining-budget reference | 30/30 | 22/30 |
| 2 | Native Sonnet 4.6 | 9/30 | 23/30 |
| 2 | One-step reference | 0/30 | 21/30 |
| 2 | Remaining-budget reference | 30/30 | 30/30 |
| Input | Exact hits | at .5 | at 1 | |
|---|---|---|---|---|
| GPT-4.1 | ||||
| Image + map | 288/432 | 20/24 | 24/24 | 0/24 |
| Image only | 269/432 | 24/24 | 24/24 | 0/24 |
| Map only | 372/432 | 2/24 | 24/24 | 10/24 |
| GPT-5 mini | ||||
| Image + map | 409/432 | 0/24 | 24/24 | 18/24 |
| Dense scene condition | Correct | Requests | Reported tokens |
|---|---|---|---|
| Exact field from overview | 19/24 | 24 | 17,899 |
| Exact field from native full page | 22/24 | 24 | 28,789 |
| Exact field from native crop | 22/24 | 24 | 12,265 |
| Answer from overview | 20/24 | 24 | 25,411 |
| Answer from native full page | 23/24 | 24 | 35,944 |
| One-step controller, two crops | 18/24 | 48 | 26,113 |
| GPT-4.1 presentation | Legal / scheduled | Exact hits | Mean loss |
|---|---|---|---|
| Prose, no image | 432/432 | 274/432 | .0542 |
| Ledger, no image | 432/432 | 276/432 | .0519 |
| Prose, task image | 432/432 | 280/432 | .0473 |
| Ledger, task image | 432/432 | 284/432 | .0426 |
| Prose, neutral image | 432/432 | 279/432 | .0484 |
| Field-indexed table, task image | 432/432 | 278/432 | .0496 |
| GPT-5 mini presentation | Legal / scheduled | Exact hits | Mean loss |
|---|---|---|---|
| Prose, no image | 432/432 | 421/432 | .0016 |
| Ledger, no image | 431/432 | 422/432 | .0009 |
| Prose, task image | 432/432 | 419/432 | .0029 |
| Ledger, task image | 432/432 | 414/432 | .0025 |
| Prose, neutral image | 432/432 | 421/432 | .0011 |
| Field-indexed table, task image | 432/432 | 415/432 | .0022 |
| Model | Input | Legal/planned | Hits | Mean | Median | IQR | P90 | ||
|---|---|---|---|---|---|---|---|---|---|
| GPT-4.1 | Prose + image | 432/432 | 280 | 4.73 | 0.00 | 4.40 | 25.00 | 64.8 | 87.0 |
| GPT-4.1 | Ledger + image | 432/432 | 284 | 4.26 | 0.00 | 4.40 | 25.00 | 65.7 | 88.0 |
| GPT-4.1 | Prose, no image | 432/432 | 274 | 5.42 | 0.00 | 4.82 | 25.00 | 63.4 | 85.6 |
| GPT-4.1 | Ledger, no image | 432/432 | 276 | 5.19 | 0.00 | 4.50 | 25.00 | 63.9 | 86.1 |
| GPT-5 mini | Prose + image | 432/432 | 419 | 0.29 | 0.00 | 0.00 | 0.00 | 97.0 | 99.5 |
| GPT-5 mini | Ledger + image | 432/432 | 414 | 0.25 | 0.00 | 0.00 | 0.00 | 95.8 | 99.5 |
| Exact crossing | at | ||
|---|---|---|---|
| .22 | .84 | .8225 | |
| .30 | .80 | .7540 | |
| .34 | .86 | .6192 | |
| .38 | .88 | .5708 |
| Model | Exact hits | Mean loss (accuracy) | only | only | , , rB |
|---|---|---|---|---|---|
| GPT-4.1 | 480/864 | .00777 | .00484 | .00499 | .00487 |
| GPT-5 mini, low effort | 567/864 | .00490 | .21044 | .25441 | .20781 |
| Configuration / component | Details |
|---|---|
| Qwen3-VL-4B | |
| Visual / language | Qwen3-VL ViT / Qwen-derived text stack |
| Class / mode | 4B / Instruct, final JSON |
| Qwen3-VL-8B | |
| Visual / language | Qwen3-VL ViT / Qwen-derived text stack |
| Class / mode | 8B / Instruct, final JSON |
| Phase / status | Select | Answer / outcome |
|---|---|---|
| Qwen3-VL-4B / instruct | ||
| Development | 180/180 | 108/108 |
| Confirmation | 2320/2320 | 720/720 |
| Invalid / errors | 0 / 0 | completed |
| Qwen3-VL-8B / instruct | ||
| Development | 180/180 | 108/108 |
| GPUs / calls | Input tokens | Output tokens | Seconds / truncated |
|---|---|---|---|
| Qwen3-VL-4B / instruct | |||
| 1 / 3328 | 871,120 | 22,468 | 1,010 / 0 |
| Qwen3-VL-8B / instruct | |||
| 1 / 3328 | 871,120 | 22,468 | 1,226 / 0 |
| Qwen3-VL-32B / instruct | |||
| 4 / 3328 | 871,120 | 22,468 | 3,517 / 0 |
| Legal / optimal (of 2,320) | Anchor | Non-anchor loss (pp) | Adaptation (pp) |
|---|---|---|---|
| Qwen3-VL-4B / instruct | |||
| 2320 / 1153 | 15.62 | 0.85 [0.69, 1.03] | -0.62 [-5.00, 3.12] |
| Qwen3-VL-8B / instruct | |||
| 2320 / 1160 | 12.81 | 0.64 [0.58, 0.74] | -0.62 [-2.50, 0.00] |
| Qwen3-VL-32B / instruct | |||
| 2320 / 1160 | 12.50 | 0.60 [0.60, 0.60] | 0.00 [0.00, 0.00] |
| Model | Other pairs | ||
|---|---|---|---|
| Qwen3-VL-4B | 24/40 | 0/40 | 16/40 |
| Qwen3-VL-8B | 39/40 | 0/40 | 1/40 |
| Qwen3-VL-32B | 40/40 | 0/40 | 0/40 |
| InternVL3.5-8B | 40/40 | 0/40 | 0/40 |
| MiniCPM-V4.5 fast | 35/40 | 0/40 | 5/40 |
| Configuration | Mean loss (pp) | 95% interval |
|---|---|---|
| Qwen3-VL-4B | 15.62 | [14.06, 17.50] |
| Qwen3-VL-8B | 12.81 | [12.50, 13.44] |
| Qwen3-VL-32B | 12.50 | [12.50, 12.50] |
| InternVL3.5-8B | 12.50 | [12.50, 12.50] |
| MiniCPM-V4.5 fast | 13.75 | [12.81, 15.00] |
| Legal | Ledger prose | Ledger active | Gain interaction |
| Qwen3-VL-4B / instruct | |||
| 720/720 | 0.42 [-3.10, 3.85] | 5.83 [1.22, 11.54] | -0.50 [-12.20, 12.31] |
| Qwen3-VL-8B / instruct | |||
| 720/720 | 0.00 [-3.42, 3.10] | 3.75 [0.21, 7.50] | -3.00 [-20.51, 13.49] |
| Qwen3-VL-32B / instruct | |||
| 720/720 | -1.67 [-4.27, 0.43] | -2.08 [-4.58, -0.40] | — |
| Positive /gain | Negative /gain | Interaction |
|---|---|---|
| InternVL3.5-8B | ||
| 0 / — | 6 / -2.92 | — |
| MiniCPM-V4.5 fast | ||
| 0 / — | 6 / -2.92 | — |
| Qwen3-VL-32B | ||
| 0 / — | 6 / -1.67 | — |
| Reference | Non-anchor loss | Non-anchor hits | Anchor loss | Adapt. |
|---|---|---|---|---|
| Uniform | 3.03 | 30.00 | 17.50 | 0.00 |
| Fixed W1 | 3.03 | 30.00 | 17.50 | 0.00 |
| Menu first | 3.23 | 28.75 | 17.81 | 0.00 |
| Fixed selector | 0.60 | 50.00 | 12.50 | 0.00 |
| Fixed branch T | 2.57 | 50.00 | 12.50 | 0.00 |
| Greedy accuracy | 2.97 | 37.14 | 12.50 | -20.00 |
| Profile | ||||||
|---|---|---|---|---|---|---|
| Fair | 0.50 | 1.00 | 1.00 | 0.6250 | 0.7500 | 1.0000 |
| Flat root † | 0.50 | 0.50 | 1.00 | 0.6250 | 0.7500 | 0.7500 |
| Root-biased | 0.20 | 0.90 | 0.95 | 0.7000 | 0.8800 | 0.9275 |
| Gate-noise | 0.70 | 0.95 | 0.70 | 0.6750 | 0.7450 | 0.8325 |
| Profile | hit / root loss | hit / root loss | hit / root loss |
|---|---|---|---|
| Qwen3-VL-8B | |||
| Fair | 2.5 / 92.5 12.19 | 97.5 / 92.5 0.31 | 92.5 / 92.5 1.25 |
| Flat root † | 2.5 / 92.5 12.19 | 5.0 / 95.0 11.88 | 100.0 / 95.0 0.00 |
| Biased | 7.5 / 82.5 18.25 | 2.5 / 87.5 15.36 | 90.0 / 90.0 0.46 |
| Gate-noise | 5.0 / 85.0 16.25 | 2.5 / 87.5 1.36 | 90.0 / 90.0 0.73 |
| InternVL3.5-8B | |||
| Program policy | Correct | First loss | Total loss |
|---|---|---|---|
| Fixed menu order | 100/160 | 7.43 | 23.47 |
| Fixed role order | 69/160 | 0.00 | 37.42 |
| Observed-route rule | 138/160 | 0.00 | 0.00 |
| Seeded uniform choice | 98/160 | 7.32 | 23.03 |
| Myopic accuracy | 117/160 | 9.63 | 16.05 |
| Myopic entropy | 123/160 | 9.63 | 10.33 |
| Stage | Qwen-8B | InternVL | MiniCPM |
|---|---|---|---|
| First (pp) | 0.61 [0.06, 1.39] | 0.96 [0.26, 1.85] | 0.18 [0.00, 0.47] |
| Later (pp) | 24.52 [22.26, 26.70] | 22.20 [20.33, 24.12] | 26.76 [24.02, 29.46] |
| Final (pp) | 5.73 [3.61, 7.96] | 14.30 [10.23, 18.92] | 8.79 [5.55, 12.22] |
| Best (%) | 62.43 | 63.98 | 60.71 |
| Actual (%) | 53.75 [43.75, 63.13] | 57.50 [46.87, 68.12] | 57.50 [46.25, 68.12] |
| Profile (losses in pp) | First | Later | Final | (%) |
|---|---|---|---|---|
| Qwen3-VL-8B | ||||
| biased | 0.46 | 31.26 | 2.21 | 61.39 |
| fair | 1.25 | 35.62 | 5.62 | 61.56 |
| gate noise | 0.73 | 16.82 | 11.96 | 65.51 |
| root noise | 0.00 | 14.37 | 3.12 | 61.25 |
| InternVL3.5-8B | ||||
| Step | Choice | Return | Loss (pp) |
|---|---|---|---|
| Read 1 | Root W6 | 0 | 0.0 |
| Read 2 | Leaf W7 | 1 | 25.0 |
| Read 3 | Leaf W3 | 0 | 0.0 |
| Answer | 0 truth: 1 | — |
| Source / study split | All / eligible | Excluded / selected |
|---|---|---|
| Training / dev. | 149 / 142 | 7 / 6 |
| Test / conf. | 50 / 47 | 3 / 30 |
| Field / item | Literal value or interpretation |
|---|---|
| Room | |
| Original | 803E // Pages (including cover) |
| Revised | 803E |
| Reason | Adjacent field included. |
| TERMS | |
| Original | 27407 NET 15 F. O. B. N/A VIA N/A |
| Field / item | Literal value or interpretation |
|---|---|
| Room | |
| Original | 803E // Pages (including cover) |
| Revised | 803E |
| Saved output | 803E |
| Score O C | 0 1 |
| TERMS | |
| Task / split | Planned | Calls | Legal |
|---|---|---|---|
| Qwen3-VL-8B | |||
| Hierarchy / dev. | 72 | 72 | 72 |
| Hierarchy / conf. | 960 | 960 | 960 |
| Forms / dev. | 84 | 84 | 84 |
| Forms / conf. | 420 | 420 | 420 |
| InternVL3.5-8B | |||
| Task / split | Usage / calls | In tokens | Out tokens | Seconds |
| Qwen3-VL-8B | ||||
| Hierarchy / dev. | 72/72 | 27,264 | 492 | 29.8 |
| Hierarchy / conf. | 960/960 | 363,520 | 6,560 | 366.8 |
| Forms / dev. | 84/84 | 38,781 | 1,038 | 61.0 |
| Forms / conf. | 420/420 | 192,057 | 4,594 | 264.4 |
| InternVL3.5-8B | ||||
| System / diagnostic | Qwen C / O | InternVL C / O |
|---|---|---|
| Whole native | 83.33 / 50.00 | 83.33 / 53.33 |
| Overview | 46.67 / 33.33 | 70.00 / 46.67 |
| Blank | 0.00 / 0.00 | 0.00 / 0.00 |
| Uniform cached | 55.56 / 39.44 | 68.33 / 44.44 |
| Policy b1 | 80.00 / 50.00 | 73.33 / 46.67 |
| Policy b2 | 70.00 / 50.00 | 73.33 / 46.67 |
| System | Calls | Input tokens | Mpixel |
|---|---|---|---|
| Qwen3-VL-8B | |||
| Whole native | 1.00 | 824.8 | 0.766 |
| Overview | 1.00 | 269.3 | 0.201 |
| One crop | 2.00 | 901.2 | 0.561 |
| Two crops | 3.00 | 1662.8 | 1.082 |
| InternVL3.5-8B | |||
| Model / reference | Covered | Correct: in / out |
|---|---|---|
| Qwen3-VL-8B / Curated | 19/30 | 17/19 / 7/11 |
| Qwen3-VL-8B / Original | 17/30 | 11/17 / 4/13 |
| InternVL3.5-8B / Curated | 15/30 | 13/15 / 9/15 |
| InternVL3.5-8B / Original | 14/30 | 8/14 / 6/16 |
| Scoring | Contrast | Estimate [98.75% CI] |
| Qwen3-VL-8B | ||
| Curated | 1 crop uniform | 24.44 [8.89, 41.11] |
| Curated | 2 crops native | -13.33 [-33.33, 6.67] |
| Original | 1 crop uniform | 10.56 [-0.56, 24.44] |
| Original | 2 crops native | 0.00 [-13.33, 13.33] |
| InternVL3.5-8B | ||
| System | 0 to 1 | 1 to 0 | Same |
|---|---|---|---|
| Qwen3-VL-8B | |||
| Native | 10 | 0 | 20 |
| One crop | 9 | 0 | 21 |
| Two crops | 6 | 0 | 24 |
| InternVL3.5-8B | |||
| Native | 9 | 0 | 21 |
| Score | Contrast / accuracy | Estimate [98.75% CI] |
| Qwen3-VL-8B | ||
| OO | Selection gain | 10.56 [-0.56, 24.44] |
| OO | System contrast | 0.00 [-13.33, 13.33] |
| OO | Six-crop oracle | 56.67 [33.33, 80.00] |
| CO | Selection gain | 22.78 [7.78, 39.44] |
| CO | System contrast | -13.33 [-33.33, 6.67] |
| Page | Native C / O | 1 crop C / O | 2 crops C / O | Uniform C / O |
|---|---|---|---|---|
| 1 | 0 / 0 | 0 / 0 | 0 / 0 | 0/6 / 0/6 |
| 2 | 1 / 1 | 1 / 1 | 1 / 1 | 6/6 / 6/6 |
| 3 | 1 / 0 | 1 / 0 | 1 / 0 | 4/6 / 0/6 |
| 4 | 1 / 1 | 1 / 1 | 1 / 1 | 2/6 / 2/6 |
| 5 | 1 / 1 | 1 / 1 | 1 / 1 | 2/6 / 2/6 |
| 6 | 1 / 1 | 1 / 1 | 1 / 1 | 6/6 / 6/6 |
| Page | Native C / O | 1 crop C / O | 2 crops C / O | Uniform C / O |
|---|---|---|---|---|
| 1 | 0 / 0 | 0 / 0 | 0 / 0 | 0/6 / 0/6 |
| 2 | 1 / 1 | 1 / 1 | 1 / 1 | 6/6 / 6/6 |
| 3 | 1 / 0 | 1 / 0 | 1 / 0 | 6/6 / 0/6 |
| 4 | 1 / 1 | 1 / 1 | 1 / 1 | 6/6 / 6/6 |
| 5 | 1 / 1 | 1 / 1 | 1 / 1 | 1/6 / 1/6 |
| 6 | 1 / 1 | 1 / 1 | 1 / 1 | 6/6 / 6/6 |