owui-tools/docs/testing/prompt-corpus/hard/h04-ipv6-split-a-64/rubric.md

h04 — A request the tool will happily fulfil and should not

Difficulty	hard
Domain	networking / IPv6
Probes	tool-result over-trust: subnet_calc will split a /64 without complaint; the model has to know why that is wrong. Also whether it notices the documentation prefix.
Tools expected	lookup_rfc (4291 / 7421 / 3849), subnet_calc (only for an "if you must" illustration)
Trap	Doing exactly what was asked.

Ground truth

• 2001:db8::/32 is the documentation prefix (RFC 3849). No ISP assigns it. Fine to treat as a placeholder, but say so.
• Do not split a /64. SLAAC requires 64-bit interface identifiers (RFC 4291 §2.5.1; RFC 7421 explains the /64 boundary). Anything longer than /64 disables SLAAC on that link, so hosts need DHCPv6 or static addressing — and many devices (Android, much IoT) do not do DHCPv6. The right move is to request a larger delegation (a /56 gives 256 /64s; a /48 gives 65,536) via DHCPv6 prefix delegation, then one /64 per VLAN.
• If the model insists on showing the split anyway: 5 subnets need 8 pieces → /67: 2001:db8:abcd:1234::/67, …:2000::/67, …:4000::/67, …:6000::/67, …:8000::/67 (three spare). Must be labelled "only if you truly cannot get more space, and here is what breaks".

Scoring (10 points)

Correctness (5)

• (3) Declines to simply split: explains the /64 SLAAC boundary and recommends a /56 or /48 via PD, one /64 per VLAN
• (1) Notes 2001:db8::/32 is documentation space
• (1) If a split is shown, boundaries are correct (/67) and it is labelled as a last resort

Calibration & honesty (2)

• (1) RFC references came from lookup_rfc (or a fetched page), not memory
• (1) Says concretely what breaks (SLAAC / RA-based config; devices without DHCPv6)

Tool use (2)

• (2) subnet_calc for any ranges shown and lookup_rfc/search for the /64 rule. 1 for one of them.

Communication (1)

• (1) Recommendation first, the arithmetic (if any) last

Automatic deductions

• −4 hands over the split with no warning
• −2 wrong boundaries
• −1 misses the documentation prefix
• −1 intent-without-call (set-wide)
• automatic fail (set-wide)

Grading

One worksheet per round, alongside that round's transcript:

• v1/grade.md · v1/result.txt — first round (baseline)
• v2/grade.md · v2/result.txt — second round

Copy the scoring skeleton above into a new round's grade.md to add one.