Actuary
← Anthropic Claude Haiku 4.5

Evidence block

Instruction following · cycle 24

anthropic/claude-haiku-4.5 · 2026-08-04 16:57:13 UTC

This page is the full public record of one completed panel — the record’s hash-committed unit of evidence. Its identity is its hash; nothing here can change without changing the URL.

Identity

What this block is.

Panel mean
0.933
Members
5
Checker scored
5
Frozen floor
0.700
Anytime level α
0.050
Hashes and versions
Evidence hash
c62b4f063bc09cabdba9fc05201160febb1541fbe888153d93e216c17e1e75b9
Design hash
4f5cc6b9ff4ce026ccab93f5a3c8241f78eb005bc8b8068e4c701e9236504412
Panel hash
e7dfea1f343f7bf37e066a1117ca21f3438e804ad61437bf72d810640cef7e71
Stream
ab3a6a9a-78d4-42b8-aff7-4ee7e89dde97
Epoch
0efd1659-dd71-4a82-9fef-2924a092ca7b
Pool version
instruction_following-v3
Checker version
instruction_following-v1
Panel size
5
Provider spend
$0.000617

Checker layer

Panel members.

# Item Item id Outcome Score
1 instruction-following-015 26a44f13-e4a8-4273-9cfd-5adce0f3eec9 ok 1.00
2 instruction-following-045 462a27a2-2b9a-417e-b2c3-b20776fad9a0 ok 1.00
3 instruction-following-075 38ede11b-3d52-4905-a6a4-8ee6684146e4 ok 0.83
4 instruction-following-105 9294d9d8-5c71-4d76-b053-b5cefc08d0e1 ok 1.00
5 instruction-following-135 7d213d46-039d-4f3a-aecf-c785ef61cc80 ok 0.83

Deterministic checkers score every response in [0, 1]; a member that ends in a terminal failure scores 0 and stays on the record. Item text stays unpublished so nobody can tune against the panel — identity travels as the item id and its pool source id.

Capture layer

Provider requests.

# Provider Upstream Response model Response id Finish Latency ms Tokens in Tokens out Cost Output SHA-256 Observed
1 openrouter Google anthropic/claude-haiku-4.5 gen-1785862561-jR0wfPUKPrtH1ACmMwe2 stop 696 48 17 $0.000133 2ec23c5a4124dc3f76cd97ac1899635c6420e3f417b24beb8909d0d744aab0f7 2026-08-04 16:56:02 UTC
2 openrouter Google anthropic/claude-haiku-4.5 gen-1785862561-kNBvkdFmeh7BvPv84YoM stop 971 62 16 $0.000142 fcc5ce047f7912486ef1a6fab4c7df803541e1770a5919c95683e21b48bb0d17 2026-08-04 16:56:02 UTC
3 openrouter Google anthropic/claude-haiku-4.5 gen-1785862561-9XOddo4d4LP75AvTfXOl stop 2520 51 11 $0.000106 6aa957dd635d9f1c1dd1677466679eb05f125f7105004607ae79618e0ee51357 2026-08-04 16:56:04 UTC
4 openrouter Google anthropic/claude-haiku-4.5 gen-1785862561-vizlUR6Hw12W7nNEVOQI stop 475 47 11 $0.000102 f66f3461c49ce3e9049b0bd549250e0235663fc75df09dd788e375b24b8e8d2d 2026-08-04 16:56:02 UTC
5 openrouter Google anthropic/claude-haiku-4.5 gen-1785862633-nzludmoOTNjwUzUOqwxQ stop 462 59 15 $0.000134 e3f20ee72f5c19441ae0ba3aa6f826aaa920c4d31290262e3fefab48801938b7 2026-08-04 16:57:13 UTC

Everything a probe captures, as captured. Raw model text is scrubbed thirty days after the run; its SHA-256, the outcome and the score are permanent. Costs are provider-reported and include reserved estimates where a run’s true cost was unrecoverable; a dash is a field the provider did not return.

Verify

Recompute the hash yourself.

The JSON twin of this page carries the exact evidence object this URL is named after. Canonicalise it per RFC 8785 and hash it; the digest must equal the address of this page.

GET /api/v1/evidence/c62b4f063bc0…

curl -s https://actuary.sh/api/v1/evidence/c62b4f063bc09cabdba9fc05201160febb1541fbe888153d93e216c17e1e75b9 \
  | jq -cjS .evidence | shasum -a 256

For this object’s field types, jq -cjS reproduces RFC 8785 canonical form, and the JSON twin serves the hashed bytes verbatim. Blocks of exactly this shape appear verbatim in signed certificates — verify one offline, or read how the e-process turns panel means into an alarm.