[SakarZaidan]

A shape-and-chemistry matching exercise, shipped together with the clinical attrition that kills the candidate afterwards. The second half is the design, not an afterthought.

candidate-A:
  Puzzle complementarity 100% (4 matched, 0 clashing, 0 unfilled)
  Preclinical pass chance adjusted to 90% by the fit
  Passed preclinical
  Phase I pass rate 52.0%, unchanged by the fit
  Passed Phase I
  Phase II pass rate 28.9%, unchanged by the fit
  Failed Phase II

Stopped at Phase II: no measurable benefit to patients
  The biggest single reason drugs fail. Over half of late-stage failures are lack
  of efficacy. The target was hit and the patients did not get better, which
  usually means the target was the wrong one.

A perfect solution to the puzzle still usually fails. The fit changes the preclinical odds, because chemistry is what preclinical work tests. It does not touch Phase II at all, because whether a target is the right target is not a question about the molecule. test_a_perfect_fit_does_not_rescue_the_clinical_odds holds that asymmetry across 500 seeds.

Subplan 33 was explicit that a student told "you cured it" has learned something false. Nothing here congratulates anyone who reaches approval; the text says what it is, and a test asserts the wording.

The numbers

The phase rates are the only measured numbers in this sandbox. 52 percent through Phase I, 28.9 through Phase II, 58 through Phase III, 8.1 percent reaching approval, twelve or more years and over 2.5 billion dollars per approved drug. Everything else in the sandbox is ILLUSTRATIVE, and the contrast is deliberate because the failure rates are the curriculum.

test_the_phase_rates_multiply_to_roughly_the_published_overall_rate multiplies the three and checks the product against the published figure, so an edit to one of them cannot drift the lesson quietly.

Refusing to look like a prediction

Verdict is an enum with exactly one member, so there is nowhere to put "calibrated". A test asserts the member count, which makes "this is never a prediction" a fact about the type rather than a convention.

No field is named affinity, potency, IC50 or anything else that reads as a physical quantity, and test_no_result_uses_the_language_of_efficacy checks none appears in a result's attribute surface or its rendered text. The completion measure is called complementarity and its docstring says it is a puzzle fraction.

Safety: curation, and now tested

Subplan 33 argued an intent filter is not implementable. The edit that disables a harmful gene is the same edit reversed, and the software sees identical letters. Any filter would be trivially bypassable or restrictive enough to block real learning, and both failures leave someone believing a control is in place.

What is enforceable is what exists, so four properties now have tests rather than assurances:

1. The sandbox imports nothing from the design tool. Checked by parsing every module with ast, not by trying an import, because an import inside a function would not show up at runtime and is exactly how this property gets lost. It therefore has no route to synthesis, and integrase-screen's gate is not weakened by the sandbox existing.
2. Every shipped gene is declared with its origin, and the test fails if a reachable gene is missing from that list.
3. Exactly one piece of content is not invented: thirteen codons of human haemoglobin beta, needed because the sickle cell variant has to sit at its real position. The test fails if that count rises, which makes adding real sequence a visible decision.
4. Nothing imports a module that writes files or opens connections.

I broke both structural guards to check they work: added import integrase and import socket to a sandbox module, confirmed each test caught it at the right line, reverted. A guard that cannot fail is worse than no guard, because it reads as assurance.

Self-review

• 20 new tests, 93 Python tests total. mypy --strict, ruff and the voice check clean.
• Two bugs found and fixed while writing: str.title() turned "phase II" into "Phase Ii", and a docstring claimed the phase rates multiply to roughly 5 percent when the arithmetic gives 8.7.
• The documented example output was generated by running the documented code, not written by hand.
• COMPLEMENTS is tested for symmetry, since a non-symmetric pairing table makes the puzzle unsolvable in one direction and it would not be obvious from reading it.

Not in this PR

Huntington's repeat-count dose response and cystic fibrosis dF508, which need repeat expansion and deletion in Sequence. The zoom navigation subplan 33 describes: the containment hierarchy exists and nothing walks it. Diploid genomes, without which sickle cell recessiveness cannot be taught, which the trait layer currently states in its output rather than modelling. All logged.

A biology educator to review the content remains the largest open item and is not something I can close alone.