# R12 — Technical precision and assessment consistency

## Cold understanding — AI role simulation / content review

I understand this as a fictional research hearing about a precise historical question: how an advantage affected another political actor’s options. I choose one country, keep an initial idea, select two supporting cards and one complication, and write an actor–action–changed-option connection. Two supplied source-use questions test particular inference errors; they do not diagnose my paragraph. I decide whether to keep or change my reasoning, identify another explanation and needed evidence, and test the main explanation against a second country. My final writing is one current 60–85-word explanation and one shared 30–45-word comparison. A person judges their meaning. Software records inclusion, lengths, IDs, and confirmations. Historical outcomes remain fixed.

This is an **AI simulation of a second technical editor**, not a real human editor or browser trial. I read only `Player_Screens.txt` and `How_to_Play.txt`. I did not inspect source code, earlier reviews, original sources, downloads, or browser behavior. References below identify the frozen text, with representative line numbers. Joined words in the extraction are not reported as visual spacing defects.

## 1. The main paragraph receives an additional criterion only at sign-off

**Evidence:** The recurring paragraph instruction asks for an action, changed options, both supporting IDs, and a source limit (for example, Player Screens 361–366). At Finish, the second self-check adds that the paragraph “deals with the complication” (789). The opening final-work summary also omits that requirement (85; repeated in all four cases).

**Why it matters:** Addressing a complication is central intellectual work. A student can follow the explicit drafting instruction and still discover a new paragraph requirement at submission. A complication note on the board is not necessarily the same as addressing it in the explanation.

**Smallest replacement:** Replace the recurring paragraph instruction with: “In your own words, explain how an action changed another actor’s options. Cite both supporting card IDs, address the complication, and state what at least one source cannot show. Use 60–85 words.” In each opening summary, replace “using both supporting card IDs and a source limit” with “using both supporting card IDs, addressing the complication, and stating a source limit.” Keep the later self-check.

## 2. The reading duty and the recorded minimum are different

**Evidence:** Choose evidence says, “Before you explain: open the full text and Source details of the cards you use” (145; recurring). Inclusion checks require only “At least one main-case source record opened” (821; repeated across cases). Counts distinguish cards from source records, and explicitly say opening does not prove reading (183).

**Why it matters:** The minimum recorded event can look like satisfaction of the larger reading instruction. “Source record” also changes the visible name from “Source details”; it could be mistaken for opening the original website.

**Smallest replacement:** Change that inclusion label to “At least one main-case Source details section opened — this records a minimum, not completion of all required reading.” Add to the earlier instruction: “Read these for every card you use; the inclusion list records only a minimum opening check.” This preserves the stronger reading duty while describing the visible threshold honestly. If the intended check is every used source instead, that is a specification change requiring separate implementation verification.

## 3. The changed-option prompt can demand an unsupported baseline

**Evidence:** Step 3 asks, “What could they do before, and what could they do after the action?” in every case (for example, 337–338). Yet the Serbia starters include “Opponents could organize but faced unequal visibility,” while RS3 reports monitored coverage during a defined period. Those observations need not establish an earlier period with different visibility.

**Why it matters:** A literal before/after answer can invite an invented baseline. It also works poorly for continuing constraints or an opportunity whose availability is uncertain. The surrounding material carefully separates observations and inferences; this prompt should apply the same discipline.

**Smallest replacement:** “Name the affected actor and the option the action opened, narrowed, or left uncertain. Describe a before-and-after change only where the evidence supports it.” This still requires a mechanism and named actor, while allowing an explicit evidential limit.

## 4. “Every stage” overstates where the current paragraph appears

**Evidence:** The recurring paragraph instruction ends, “This same paragraph appears in every stage” (364 and later occurrences). The frozen Opening view and Choose evidence screens do not contain that paragraph editor. Compare visibly has its own comparison field.

**Why it matters:** This sounds like a navigation guarantee. It also blurs the distinction between one shared main explanation and a separate shared comparison. The useful promise is that students are maintaining one current paragraph, rather than writing a new paragraph at each appearance.

**Smallest replacement:** “This is your one current explanation, shared wherever this field appears.” Keep the separate comparison message, “This same comparison appears in the final record.” Neither instruction needs to promise an editor on every screen.

## 5. Two question-level directions point at one response pair

**Evidence:** Each of the two hearing questions repeats, “Record one other possible explanation and one evidence request in the short response fields below the questions” (414 and 446; equivalent wording in all cases). After both questions there is only one visible alternative/step field and one evidence-request field, headed “Two short responses for a human reader” (471–477).

**Why it matters:** A careful student may think each question requires its own alternative and evidence request, producing four responses in two shared fields. The final list requires one alternative/step and one evidence request, which supports a combined task.

**Smallest replacement:** At both question locations use: “Across these two questions, record one alternative explanation or different step and one evidence request in the shared fields below.” Change the later heading to “One alternative and one evidence request.” This makes the response count exact without reducing discussion.

## 6. “Evidence ID” is less precise than the citation rules

**Evidence:** A keep-decision reason asks for “an evidence ID” (436; recurring), while the final explanation requires supporting **card** IDs and the comparison requires a second-case **card** ID. The game displays table IDs such as RU0, card IDs such as RU2, and underlying source IDs such as RU-B. The glossary explains the latter distinction only near the end (3507).

**Why it matters:** These identifiers serve different jobs. A student can reasonably cite the original source ID in a reason, then assume the same syntax satisfies a card-ID inclusion check. Two cards can also share one original source, so card count does not imply independent corroboration.

**Smallest replacement:** Change the keep-reason label to “Why can you keep your explanation? Short phrase, with a card ID such as RU2.” Add beside the recurring main citation instruction: “Use card IDs exactly as shown, such as RU1 and RU2. Source IDs such as RU-B identify the underlying documents.” Adapt examples to the current country. Retain the glossary’s source-independence warning.

## 7. “Practice readings” is an imprecise inclusion-check noun

**Evidence:** The checklist says “Two practice readings and board decisions recorded” (823; recurring). The hearing displays answer choices, “Consider this reading,” a decision, and “Apply to my board.” Elsewhere, opening records explicitly does not establish that a student read them.

**Why it matters:** “Readings” can mean documents read, interpretations selected, or interpretations judged correct. The task repeatedly and appropriately separates software inclusion from human judgment. This one checklist label should name the recorded student actions with equal precision.

**Smallest replacement:** “Two practice answer choices and two hearing decisions recorded.” Keep the introductory statement that inclusion cannot judge writing quality. Do not use “correct readings” or imply that attempting the practice demonstrates sound historical reasoning. This is a label recommendation; the packet does not establish the actual software threshold.

## 8. “Apply to my board” leaves the scope of a decision unclear

**Evidence:** Each hearing offers “Keep it, with a reason,” “Make a step more precise,” and “Change supporting evidence,” followed by “Apply to my board.” The output heading is “What you applied,” while the current-explanation placeholder promises the app will never write or rewrite the paragraph.

**Why it matters:** A connection step, evidence selection, hearing decision, and paragraph are different objects. “Apply” does not tell students which object changes or whether they must still revise their own writing. This is an ambiguity in the stated contract, not an observed automation failure.

**Smallest replacement:** Use “Record my hearing decision” and “Recorded hearing decisions.” Add: “Review any changes to your connection or evidence, then update your paragraph yourself.” If the action actually performs a specific board mutation, name that mutation alongside the recording label after checking the interaction separately.

## 9. Finishing and exporting need a clearer handoff

**Evidence:** The final screen offers “Finish this hearing record,” “Preview current record,” and several export formats. Its explanation says, “The HTML report is a readable record with citations. The JSON save reopens the editable workspace” (841–848). Earlier work and refreshed questions are also said to remain “in your record,” without specifying export coverage.

**Why it matters:** Students need to distinguish readiness for human review, a readable hand-in, and an editable backup. They also need a reason to inspect the preview before assuming that a report includes every earlier attempt. The existing warning that “Included” never means correct is strong and should remain.

**Smallest replacement:** Add beneath the finish control: “Finishing records your readiness for human review; it does not approve your reasoning.” Replace the export explanation with: “Preview the current record before sharing to check which current and earlier work it includes. Download HTML for a readable record with citations. Keep the JSON save, then use Import a JSON save to resume editing.” The finish-status sentence is a specification to verify against the action, not a claim that I tested it. I did not validate any generated report or download.
