Use cases
Four decisions, and what the engine won't say about them
Each of these runs on a free sample dataset in about thirty seconds. Each one also names its own limit.
Engine 0.3.6, core hash 4d24e9fc… Every figure on this page came out of a live call. Nothing is estimated.
01 · sample:agent_traces
Which of your agent runs will fail
800 runs across 4 agents, 5 tasks and 4 tools. Ranked raw, the engine refuses.
{ "result": "none",
"reasons": ["Insufficient features: found 2 viable arms, need 4. …"],
"retry": "unproductive",
"guidance": "Retrying the identical call will return the identical result." }
No ranking, no model_ref, no Verdict. retry: "unproductive" is the engine telling your retry loop to stop. Add reading: {"kind": "trace", "version": 1} and the same 800 rows clear at 4.285714 top-decile lift. The reduction derives each run's prior failure rate over strictly earlier runs — agent_prior_outcome_rate, agent_prior_n, tool_prior_outcome_rate — features that are not columns in the raw table.
What it refuses: the raw table. Two viable features where it needs four.
02 · sample:saas_churn
Which accounts are about to churn
800 accounts. 4.444444 top-decile lift, holdout validation on 480 train / 160 holdout / 160 reserved.
support_tickets > 3
logins_last_30d <= 14
tenure_months <= 13
monthly_charges > $157.41
→ 1% of accounts match all four, and they are
4.477611940298507× more likely to churn
rows 1–56 · act · L3
row 57 down · escalate · L1
The top of the list is not uniform. Acting on "the top 100" would action 44 rows the engine asked you to escalate.
What it refuses: below the published floor a partitioned run cannot judge either half. See case 5.
03 · sample:b2b_leads
Which leads are worth working
800 leads. 5.263158 top-decile lift, calibration error 0.044679.
demo_requested = true
first_touch_days <= 7
page_view_tier = high
industry is_not_one_of [manufacturing, retail]
→ 1.25% of leads match the full pattern,
at 7.346938775510204× lift
What it refuses: to tell you what to change. On this dataset hs_explain_levers returns an empty set — the engine found no minimal change it has evidence for, and says so rather than inventing a next-best-action. principal_reasons comes back empty for the same reason.
04 · sample:sensor_stream
Which machines are about to fail
4 signals from 40 machines, every 6 hours for 30 days. 4,815 raw readings and 15 fault intervals reduce to 1,160 daily snapshots with trailing 1/3/7-day windows. 7.518539 top-decile lift.
9 snapshots · act · L3
~68 · escalate · L1
24 · refuse · L0 · low_likelihood
This is the run where all three bands appear at once. refuse is not an error — it is the engine declining to vouch, and it caps autonomy at observe-only.
vibration_slope_3 high single-feature AUC (0.97) on held-out data;
possible leakage — verify this value is historically
known BEFORE the outcome. NOT removed. [advisory]
vibration_std_7 AUC 0.95 [advisory]
vibration_slope_7 AUC 0.93 [advisory]
vibration_std_3 AUC 0.91 [advisory]
temp_c_slope_7 AUC 0.89 [advisory]
temp_c_slope_3 AUC 0.86 [advisory]
What it refuses — and this is the part worth reading: a feature that predicts the outcome almost perfectly is usually a mistake in the data, not a discovery. The engine cannot tell which from the values alone, so it hands you the number and the question instead of a clean-looking model.
With this reduction the engine scores snapshots, not machines — n_entities: 40, median_rows_per_entity: 29, and the ranked ids read row 1127.
05 · sample:saas_churn with partitions: 2
When it can't judge, it says so in the handoff
800 rows, below the 1,250-row floor a partitioned run needs. The run comes back usability: "unjudged", with model_ref: null, verdict: null, and no band on any row.
"pooled: the POOLED corpus cleared the published lift >= 1.5 bar…
It is NOT a verdict on this run: the partitioned draws were not
judgeable, so no draw-level clearance was established.
Read usability and partitions.clearance_frequency."
clearance_frequency: { cleared: 0, of: 2, judgeable: 0 }
Then hand the analysis to another agent with hs_context_brief, and the brief refuses too:
"USABILITY: unjudged. The engine did not certify this run, so nothing
below is a certified finding… The drivers are reported because the
engine reported them; do not act on this ranking, and do not relay
them as a finding."
What it refuses: to let an uncertified result travel as a finding. The caveat is attached to the artifact, not to the conversation the artifact came from.
Regulated decisions
A decision an examiner can follow
Lending, insurance and merchant portfolios need three things a model ordinarily can't give an examiner: a decision traceable to its inputs, the same answer on a rerun, and a named human who owns it.
Hunter-Seeker returns all three — a signed Verdict carrying analysis_id, spec_hash and dataset_content_hash; a deterministic engine; and on_behalf_of naming the authenticated human, with person-level decisions capped at L2, and at L0 when no human is on the credential.
There is no free sample for this one yet.
Frequently asked questions
