You signed in with another tab or window. Reload to refresh your session.You signed out in another tab or window. Reload to refresh your session.You switched accounts on another tab or window. Reload to refresh your session.Dismiss alert
{{ message }}
Repository navigation
JIT admission: openIMIS claim program refused under clinical-write (risk override, idempotency key, low confidence) #1329
The launch demo's openIMIS claim-entry program, compiled through the JIT wrapper from one Claude run, cannot be admitted. The standard run gate refuses its attended canary under clinical-write, so nothing runs. The demo and the scoreboard therefore run it on the demo profile, labelled "not admitted".
Evidence: demo/cua_to_program/results/launch-4/admission.json (the gate's full output) and packages/flow/benchmark/scoreboard/raw/openimis/first_run/launch-4/items.csv (on #1303). Every other gate check passes: identity coverage 8/8, effect coverage 1/1, approval fallback, sealed manifest.
The full list of certification violations, from evaluate_policy(workflow, clinical-write, require_current_risk_certification=True) on that bundle:
Step
Action
Rule
Reason
step_002
press Tab (after typing the insurance number, to trigger the lookup)
require_reviewed_action_risk
special key may navigate, edit, or invoke a shortcut; no bound operator classification
step_012
click on an unlabelled visual target (service picker)
require_reviewed_action_risk
unlabelled pointer control; no bound operator classification
step_012
same
require_human_approval_below_confidence
compile-time confidence 0.70 < 0.80
step_015
click the icon-only Save
require_idempotency_key_for
no declared system-of-record effect carries an idempotency key
step_015
same
require_human_approval_below_confidence
compile-time confidence 0.70 < 0.80
Gaps
The JIT admission path has no way to certify a risk override.review and canary don't carry a qualification step, so these forms can't be admitted through compile_agent, however good the run is. Admission needs a reviewed risk classification for each step, bound to the bundle and policy, recorded as the gate expects.
No idempotency key on a mined effect. The mined record check already selects the claim by the confirmed claim_no parameter. Effect mining could propose that bound field as the idempotency key, for the reviewer to confirm. Check first whether openIMIS enforces unique claim numbers.
I rebuilt the recorded launch-4 run offline on #1303 + #1336 + #1339, with no model calls, and evaluated it under clinical-write with current risk certification.
Step
Rule
Verdict
step_015
require_idempotency_key_for
Missing evidence, fixed in #1339. The mined check already selected the claim by the confirmed claim_no with expected_count=1, so it already refuted a second claim with the same number. It just didn't declare the key. A person now names it: confirm_params(idempotency_key="claim_no") or jit review --idempotency-key claim_no. After that, the violation is gone.
step_002 (Tab after Insurance No.)
require_reviewed_action_risk
Real. A special key can edit or invoke a shortcut, and only a person can classify it.
step_012 (click)
require_reviewed_action_risk
Real. It's an unlabelled pointer control. In this recording, the DOM label at the click is "Explanation" (the text box), not the service picker.
step_012, step_015
require_human_approval_below_confidence (0.70)
Real. Confidence is template 0.5 + hashed identity 0.2. The capture recorded no structural locator for either click, and OCR found no label in either crop. The run's read-back doesn't change that. It confirms the record after the save. It doesn't help find the control before the click. A misplaced focus click on step_012 would put the explanation in another field, and step_015 would save it before the read-back refutes it.
I couldn't check whether openIMIS enforces unique claim codes: Docker is down here, and the load is about 220. The key detects a duplicate (more than one record with the run's code refutes the run); it doesn't prevent one. The person naming the key asserts that the app keeps claim codes unique.
What a person would approve
There's no admission path for these yet. review records "reviewed every step" and nothing per step, and require_human_approval_below_confidence has no sign-off record at all: only higher confidence clears it. Proposal: the JIT review takes a per-step classification, signed and bound to the bundle digest and policy, which the gate accepts as the operator override it asks for:
step_002 Tab: "moves focus from Insurance No. and triggers the insuree lookup; reads only, writes nothing; reversible."
step_012 click: "focuses the Explanation text box; writes nothing by itself; reversible." This also covers the low-confidence sign-off after the reviewer looks at templates/step_012.png.
step_015 click: "the icon-only Save of the claim form", signed off after looking at templates/step_015.png and the submit binding (icon hash, div>div>span>button, bottom right). fix(flow): run gate counts a submit binding as an icon-only Save's identity #1274's submit-binding evidence is the candidate replacement for this sign-off once it passes its safety review.
Re-run result
Not admitted. Certification still refuses on the four violations above, so the canary's run gate refuses before anything runs. Confirm (with the key) and review succeed. No placeholder effects.
Activity
abrichr commentedon Oct 19, 2026
Which refusals are real
I rebuilt the recorded
launch-4run offline on #1303 + #1336 + #1339, with no model calls, and evaluated it underclinical-writewith current risk certification.require_idempotency_key_forclaim_nowithexpected_count=1, so it already refuted a second claim with the same number. It just didn't declare the key. A person now names it:confirm_params(idempotency_key="claim_no")orjit review --idempotency-key claim_no. After that, the violation is gone.require_reviewed_action_riskrequire_reviewed_action_riskrequire_human_approval_below_confidence(0.70)I couldn't check whether openIMIS enforces unique claim codes: Docker is down here, and the load is about 220. The key detects a duplicate (more than one record with the run's code refutes the run); it doesn't prevent one. The person naming the key asserts that the app keeps claim codes unique.
What a person would approve
There's no admission path for these yet.
reviewrecords "reviewed every step" and nothing per step, andrequire_human_approval_below_confidencehas no sign-off record at all: only higher confidence clears it. Proposal: the JIT review takes a per-step classification, signed and bound to the bundle digest and policy, which the gate accepts as the operator override it asks for:templates/step_012.png.templates/step_015.pngand the submit binding (icon hash,div>div>span>button, bottom right). fix(flow): run gate counts a submit binding as an icon-only Save's identity #1274's submit-binding evidence is the candidate replacement for this sign-off once it passes its safety review.Re-run result
Not admitted. Certification still refuses on the four violations above, so the canary's run gate refuses before anything runs. Confirm (with the key) and review succeed. No placeholder effects.