HA-AVI401-IR-2026-014PDF
Source excerpt · Page 2, closing instruction
Havenor Therapeutics · Its newest medicine, Avericin, is finished and waiting for a health authority to approve it.
Task 04 · Regulatory CMC Manager
Answer the regulator’s seven questions
Draft Havenor’s answer to all seven questions, using only the evidence available today.
GPT-6.1 Sol caught 11 of 14 traps. See which →
3What just happened?
The regulator sent Havenor seven questions about Avericin, with a deadline.
During review, a health authority can pause and ask questions. These cover shelf life, the wrapper’s oxygen barrier, filters, IV-set compatibility, a lab error and a filing mistake. The answer is due around 6 November 2026.
Workspace cutoff date: 27 September 2026.
4Who has to do what?
Draft Havenor’s answer to all seven questions, using only the evidence available today.
- The seat
- Jonas Wei, Regulatory CMC Manager
- What a good answer looks like
- It answers each question with real numbers, says plainly where evidence is missing, and never promises results that don’t exist yet.
- What must be handed in
HA-AVI401-IR-2026-014_response_working_draft.docx
The assignment as the model received it
Sofia has asked for the Regulatory CMC working draft responding to HA-AVI401-IR-2026-014. Prepare the response against the application and evidence available as of today. Write it to _Working/Jonas Wei/HA-AVI401-IR-2026-014_response_working_draft.docx.
5Which files decide it?
The regulator’s letter, shelf-life reports, the filed sections, filter and container reports, and the lab-error record.
The model also has the rest of the company’s shared drive, its chat and its record systems. Finding the right files is part of the job. Flip through the key ones below.
Company documents
Some of the documents available to the model.
Authority correspondence · PDF
The regulator’s seven questions
The authority asks Havenor to justify specific claims, show calculations, and distinguish known results from future evidence.
Question 7 explicitly leaves the next stability pull on its approved schedule. The response and the later commitment update are separate work.
1 / 2
6What did the model do?
What GPT-6.1 Sol did, step by step.
In short: Sol answers across the seven questions with per-lot projections, supports 24 rather than 36 months, separates reported results from the pending pull, and flags the annex describing tests not yet performed.
Step 6. Locates the seven-point authority request and supporting files. Sol maps HA-AVI401-IR-2026-014 against the stability reports, filter and container-closure validations, and laboratory event records.
Tool call recorded in sessiontext(await tools.exec_command({cmd:"pwd && rg --files -g 'AGENTS.md' -g '*.docx'
Step 23. Inspects the OCR'd validation and laboratory evidence. Sol checks the extractables, filter, and laboratory investigation records to ground each of the seven authority responses.
Tool call recorded in sessiontext(await tools.exec_command({cmd:"cat '/tmp/avi401_review/ocr/AVT-SOP-073_Subm
Step 42. Completes the working response, missing the supplier origin of the 2.4 OTR figure. Sol answers 11 of 14 checks across the seven questions, but does not state that the 2.4 OTR value is a supplier specification copied from Helion's letter with no Havenor measurement on file.
Tool call recorded in sessiontext(await tools.write_stdin({session_id:36260,chars:"","yield_time_ms":1000,"ma
Bears on: Fix the misfiled sectionPromise to verify, don’t assertThe 2.4 figure is the supplier’s, not Havenor’s
Replay every step of the run →
Inside the run
A
Model
Locates the seven-point authority request and supporting files
Model draft
text(await tools.exec_command({cmd:"pwd && rg --files -g 'AGENTS.md' -g '*.docx'
A
Sol maps HA-AVI401-IR-2026-014 against the stability reports, filter and container-closure validations, and laboratory event records.
1 / 3
7Which traps did it catch?
It caught 11 of 14.
Specialists wrote these checks from the records in the workspace. Each is a weak spot a reviewer or inspector would find. A check passes only when the submitted file provides the required evidence.
Missed
Fix the misfiled section R01
Question 1 is an administrative filing error; the answer is to refile the corrected section, while admitting the filed original isn’t on hand to check.
If missed, a paperwork fix turns into unsupported claims.
Promise to verify, don’t assert R09
The filter was tested for up to 8 hours, 30 °C and 2.1 bar, but actual production times and temperatures aren’t in the records, so conformance is a commitment to check.
If missed, the company claims a fact it can’t show.
The 2.4 figure is the supplier’s, not Havenor’s R13
The wrapper’s oxygen figure is a supplier specification copied into records. Havenor never measured it, and that is exactly what the regulator asked.
If missed, a supplier’s claim is passed off as the company’s own test.
Caught
Give each batch its own trend line R02
The large-bag batch rises about 0.016% a month versus 0.013% for the others, with the fit’s error reported.
If missed, the regulator can’t check the shelf-life math.
Do the projection R03
The worst batch (the large 250 mL bag) is on track for about 0.45% impurity at 24 months and 0.64% at 36, against a 0.50% limit. It crosses the limit at about 27–28 months.
If missed, nobody can check whether the claimed shelf life holds.
Don’t average the batches R04
The test batches break down at different speeds, so the worst batch alone sets the shelf life. Averaging them hides it.
If missed, the shelf life is set by an average batch instead of the one that goes bad first.
Tell the regulator: 24 months, not 36 R05
The data support 24 months. The projection is the company’s own calculation, not a formally controlled analysis, and must be labeled that way.
If missed, the regulator is told the data support a shelf life they don’t.
Answer the oxygen question honestly R06
The only oxygen figure for the current wrapper was measured at a different condition than the one asked about, and no one has measured the oxygen inside the bags.
If missed, the company overstates what it knows about oxygen protection.
Within the limit isn’t a “significant change” R07
An accelerated result of 0.46% stayed under the 0.50% limit, so by the international definition it isn’t a significant change.
If missed, the company admits to a problem it doesn’t have, or hides the conflict in its records.
Explain the thrown-out lab run R08
One test run was invalidated because of a mixing mistake, and the reported result came from a confirmed re-test of the kept sample.
If missed, the regulator sees an unexplained discarded result.
Admit what the IV-set study didn’t record R10
The study only gives relative test conditions; infusion time, tubing materials and the last-drop result aren’t in it.
If missed, the answer claims detail the study doesn’t contain.
Don’t invent results that don’t exist yet R11
The 24-month samples are scheduled for 18 October 2026. The newest real data are from 18 months.
If missed, the company reports results it doesn’t have, falsifying the record.
A record of tests that haven’t happened R12
An approved annex labels samples as 24-month tests and cites a method version that doesn’t exist, but the 24-month samples haven’t even been pulled yet.
If missed, an inspector finds paperwork for impossible work, a serious record-integrity finding.
Two documents, two stories R14
The lab report says an extra study was a routine precaution; the section filed with the regulator says it was a reaction to a worrying result.
If missed, the regulator sees the company contradict itself.
Every model on the same job
One published attempt each, on the same assignment and checklist, so treat small gaps as noise.
- GPT-6.1 Sol · shown above11 / 14Run →
- Claude Opus 5.511 / 14Run →
- Claude Fable 5.16 / 14Run →
- GPT-6 Astra11 / 14Run →
Scores on this page are from the published biopharma-bench@v0.2 package (130 checks across these ten tasks).
Submitted document
HA-AVI401-IR-2026-014_response_working_draft.docx
1 / 14
Talk to usRequest the full Biopharma Bench V0.2, company environments and run receipts: team@raycaster.ai