Live demo · sample:agent_traces
Run it twice. The hash does not move.
800 agent runs, ranked by likelihood of failure. Free sample dataset, no signup, no quota. Every figure on this page came out of two runs made 33 seconds apart — and you can make a third.
# no Authorization header — returns a test key, sample datasets only
$ curl -X POST https://hunter-seeker.io/api/v1/agents/register
$ pip install hunter-seeker
$ hs rank --dataset sample:agent_traces \
--entity run_id --outcome failed --reading trace@1
A test key resolves the free sample datasets only — enforced before any upload URL is minted or any byte is read. The blast radius of that key leaking is zero by construction, which is why this page can hand you one.
The determinism proof
Same decision, fresh receipt
Two runs on the same dataset, 33 seconds apart. Four hashes identical; the verdict id and the signature differ, because each receipt is issued at its own moment.
score_set_hash 1190bef2922e750a0ac8d77169c58a72 identical
spec_hash sp1_3043035b0b858e2fb352d76e72a71e14 identical
dataset_content_hash 001cf48de0c0c2edd3a4985fefb83612 identical
model_ref mr1_d48a6f743bd4d967da1167141316ec76 identical
core_hash 4d24e9fcfe4d6d42a4ece61ac02e72ef identical
differs — each receipt is issued at its own moment:
verdict_id 01M21ZJHSM8AVX3FEQVX4CDKT0
01M21ZKJ8BY5PF2DDZ2S2YQHEN
signature 805UGROle-Hpqiiiu10wx3Wu…
h5IYl3Ij-RMs6uNFC2qcfZA-…
The score set is the claim; the signature is the receipt. A matching score_set_hash across two runs is the whole determinism argument, and it is checkable by anyone who runs the same command.
Left channel · process
What the engine is doing, while it does it
A pending poll carries a stage, a completion fraction, and an append-only facts_so_far. These are counts and identities only — never a lift, a direction, or a rank order a later stage could reverse.
stage: "profiling" frac: 0.05
row_count: 800
entity_identity: column:run_id
columns_quarantined: []
stage_history: ["profiling"]
reading trace@1 rows_in 800 → rows_out 800
groups_bound agent 4 · task 5 · tool 4
time_axis_parsed_as iso_offset
Nothing with a sign is streamed. Someone who sees a lift at six seconds and a refusal at fourteen will remember the lift — so the conclusion arrives once, whole, and only at the end.
Right channel · verdict
The conclusion, revealed once and whole
usability: actionable. The engine certified this run. Cleared the published lift bar at 4.102564 realised on every analysed row, with 4.285714 top-decile lift.
usability · actionable
gate · cleared
holdout · 480 / 160 / 160
Engine version 0.3.6
Core hash 4d24e9fcfe4d6d42a4ece61ac02e72ef
Input hash 001cf48de0c0c2edd3a4985fefb83612
Spec hash sp1_3043035b0b858e2fb352d76e72a71e14
Analysis id sk1_8ee820f66815903fdf3c56a8fae50f97
Signing key 2026-q3
Public ledger hunter-seeker.io/.well-known/jwks.json
Verdict id 01M21ZJHSM8AVX3FEQVX4CDKT0
$ hs-verify verdict.json signature.json
→ valid
Verification is credential-free — Ed25519 over RFC 8785 canonical JSON, checked against the public key set. No account, and no call to us if you cache the keys.
The band boundary
The top of a list is not uniform
Scores on this run spread cleanly from 0.6947 down to 0.2651 — and the band changes partway down. Branch on the band, never on the score.
rows 1–46 · act · L3 · certified
row 47 down · escalate · L1 · uncertain
run_0287 0.694746 act L3 certified
…
run_0530 0.531973 act L3 certified
run_0049 0.515386 escalate L1 uncertain
…
run_0507 0.265143 escalate L1 uncertain
Sixteen thousandths of a point separate the last act row from the first escalate row. Anyone acting on "the top 100" of this ranking would action 54 rows the engine asked them to escalate.
Top factors, engine-authored and in engine order: different tool, higher agent_prior_n, lower input_tokens, higher tool_prior_outcome_rate. Three of the four are derived by the trace reading — they are not columns in the raw table.
What this run refuses
The same 800 rows, ranked raw, return nothing
Drop the trace reading and the engine declines: two viable features where it needs four. No ranking, no model_ref, no Verdict — and retry: "unproductive", which tells your retry loop to stop rather than spin.
The leak guard on the cleared run removed exactly one column — run_id, excluded as an identifier rather than flagged as leakage. An empty leak guard is not a clean bill of health, and the engine says so rather than implying otherwise.
Four more decisions, and what each one won't sayThe pattern
Four conditions that only mean anything together
Pulled from the same analysis as the run above, at no extra cost. Read them as one combination — they were found together and only mean what they mean together.
tool is one of "browser"
agent_prior_n higher
input_tokens lower
tool_prior_outcome_rate higher
calibration error 0.054589
validation holdout · 480 train / 160 held out / 160 reserved
leak guard run_id [excluded] — identifier, not a predictor
usability actionable
Three of these four are derived by the trace reading rather than read from the table — agent_prior_n, tool_prior_outcome_rate and the prior rates behind them are computed over strictly earlier runs, so nothing from the future of a run informs its own score.
Association, not causation. Do not re-order them, do not rank one above another, and do not report any one of them on its own — the engine's own instruction, passed through verbatim.
Frequently asked questions
