A criterion stricter than its source fails correct runs: it pins a figure the source leaves open, picks one answer of several, insists on one of two sources, or scopes to a step that lacks the values. A looser one passes wrong runs.
Do this instead
Grade the shape of an answer the source leaves open, accept either source, scope to where the values appear, and add bounds: exactly, only, between.
Where it happened
3 real tasks. Each shows the spot and what it should have been; click any line to open the task there.
the words at faultthe row or field at faultmissing from the taskwhat is righta task file
Example 23Visual Learning · 3 Leg A turns
What is wrongCriterion 23 requires PCR-2026-0688 and its 10 July 2026 acceptance in the change board step. The golden's change board scene shows neither; both sit in an earlier scene.
Leg A turn 3 asks for one more decision step in source_check.mp4, to show that “finding a discrepancy, accepting a change request, and releasing a changed figure are different decisions”.
The change-board step keeps the documented scopes separate: PCR-2026-0688 is a later record against 32-21-07 Figure 2 that was accepted on 10 July 2026…
source_check.mp4 presents acceptance and release as separate states: the July acceptance of the 32-21-07 Figure 2 change request did not put the change in the manual, which happens only at the 30 September release.
How it breaks the standardThe PCR number and the 10 July date appear in source_check.mp4, but only in the earlier record scene, 26 to 39 seconds, and Leg A turn 3 never asks for them in the new step. So the golden is Not Present on criterion 23 at weight 5. This is Overfitting, which costs points.
What is wrongCriteria 13 and 28 require Aunt Rosalind as the possible guest. The voice note asks about family members, plural and unnamed, and the golden names Tamara Brown-Ellis instead, failing both.
Leg A turn 3 hands over turn-3-voice-note.mp3, which asks about “family members who might stay with us” and names no one. The records support more than one.
combined-home.html identifies a family member from the records as a possible overnight guest and states the recorded need that shapes where they sleep.
How it breaks the standardThe golden's combined-home.html answers the voice note with Tamara Brown-Ellis, from Contacts, Calendar and Email, and never mentions Rosalind, so the golden is Not Present on both criteria. Pinning one person where the voice note asks for any family member the records support is Overfitting, which costs points.
What is wrongCriterion 26 accepts the register prices only from the 5 May calendar event. The same list is attached to the 4 May mail, and the golden took it from there.
The nine price changes sit in two records, the 5 May calendar event and price list 5 may.pdf attached to the 4 May mail. Both carry the held prices too.
…The list you sent Marguerite the night before ("New price list, into the register tomorrow at seven", attachment price list 5 may.pdf) is the source of truth…
The trajectory grounds the nine old and new register prices in the 5 May register list, from either the calendar event "New prices in the register" or the price list attached to the 4 May mail.
How it breaks the standardLeg A turn 1 names neither record, and either one gives every price. Criterion 29 rewards reading the mail, and a run that takes the whole list from there, as the golden did, fails criterion 26 at weight 5. Pinning one of two valid sources is Overfitting, which costs points.
Overfitting: Criteria that are overly specific, inflexible or too rigid - they incorrectly reject a subset of valid implementations. Underfitting: Criteria that are overly broad, permissive or loose - they accept valid implementations, but also incorrectly accept invalid implementations too.
Both are Moderate.
Why it matters
The rubric exists to separate a correct run from an incorrect one. A criterion a correct run cannot satisfy marks down the runs that did the work, and it usually catches your own golden, because the golden followed the evidence while the criterion followed your memory. A criterion with no ceiling lets a run that also took the distractors pass, so the traps the task was built around cost nothing.
Example 23Criteria a correct run cannot pass, or a wrong one canCosts points
Visual Learning3 Leg A turnstask 6aa34dd1625bb5b220c6b1c7
From the taskObjective criterion 23
23
The change-board step keeps the documented scopes separate: PCR-2026-0688 is a later record against 32-21-07 Figure 2 that was accepted on 10 July 2026, while the group of three 26 May corrections was separately scheduled for the 30 September 2026 release, so acceptance and release remain distinct states.
From the taskFrom the Leg B conversation, with Model B
Model B · Leg B turn 4
Updated `source_check.mp4` in `/app/outputs` (now **2:52**, under three minutes), with the transcript refreshed to match.
## What I added: the change-board decision step (scene 14)
A new scene, **"Three decisions, not one,"** shows that **finding a discrepancy, accepting a change request, and releasing a changed figure are three separate decisions** — grounded in the status history around the same three cases:
- **Decision 1 — Find the discrepancy** *("Is there a discrepancy?")* → 26 May flight line: **3 found**. Raise it, don't quietly fix.
- **Decision 2 — Accept the change request** *("Do we make this change?")* → at the board: **accepted / deferred / returned** (e.g. **32-21-07 accepted (Jul); a 29-11-06 Fig 12 request deferred**).
- **Decision 3 — Release the changed figure** *("Does it ship — and when?")* → **figure freeze 25 Sep → quarterly revision to Meridian 30 Sep**. Accepted is not released.
Banner + narration drive the point home: *finding ≠ accepting ≠ releasing — three decisions, three records.* Each data point is drawn straight from the Slack/email record, and I kept the framing within what the evidence establishes (I don't claim a specific PCR *is* one of the May corrections — just that the status history *around* these figures shows all three decision types).
## Everything preserved
- **Both learning activities** intact — the predict-and-check on view reading (scenes 10–12) and the closing find-the-problem exercise (scenes 15–17).
- The **three flight-line cases** (scene 3), the **separate work-email record** with its evidence-bounded framing (scene 4), and the **full FAA leader definition / both panel headers** all unchanged.
- **Under three minutes** — verified against the encoded file at **2:52** (gTTS is deterministic, so the runtime is stable), achieved by tightening narration to offset the added scene.
Both deliverables are in `/app/outputs`: `source_check.mp4` and the updated `source_check_transcript.md`. Build script: `/app/work/build.py`.
This conversation is closed.
Task Runner
From the taskLeg A prompt, turn 3, to Model A
Leg A turn 3
For the change-board version, add one more decision step to source_check.mp4. Use the status history around the same three cases to show that finding a discrepancy, accepting a change request, and releasing a changed figure are different decisions. Keep both learning activities and the three-minute limit.
Task filesource_check.mp4
Look at The golden video. Play both scenes: the PCR number and the July date appear only in the earlier one, so a run that did the work still fails the criterion as scoped.
source_check.mp4golden artifact
Play the scene
Example 24Criteria a correct run cannot pass, or a wrong one canCosts points
Property & Space5 Leg A turnstask 6aa88e8600a9cc42267c9867
From the taskObjective criteria 13, 28
13
combined-home.html identifies Aunt Rosalind as a possible guest.
The trajectory's connected-record result identifies that Aunt Rosalind cannot walk far.
+1targetTrajectorycategoryAgent Behavior
From the taskLeg A prompt, turn 3, to Model A
Leg A turn 3
I recorded my reaction in turn-3-voice-note.mp3. Please check it and revise the same combined-home.html. In the page, make it clear which details came from my records and which are assumptions.
Task filecombined-home.html
Look at The golden page. It names Tamara as the guest the records support, and Rosalind appears nowhere on it.
combined-home.htmlgolden artifact
Task fileturn-3-voice-note.mp3
turn-3-voice-note.mp3input file
Example 25Criteria a correct run cannot pass, or a wrong one canCosts points
Creative & Media5 Leg A turnstask 6aa34dd1625bb5b220c6b1a8
From the taskObjective criteria 26, 29
26
The trajectory grounds the nine old and new register prices in the 5 May 2026 07:00 PT calendar event "New prices in the register".
+5targetTrajectorycategoryTool Use
29
The trajectory retrieves the held prices `$5.00` and `$2.00` from the 4 May mail after identifying "the bag and the kids one" from `IMG_3118.jpg`.
+5targetTrajectorycategoryTool Use
From the taskFrom the Leg B conversation, with Model B
Model B · Leg B turn 1
Done. `storvik_price_cards.html` is in `/app/outputs`.
## What I pulled it together from
- **The photo** → the voice and format: short, plain, first person, the closing line *"Ask me, not the till,"* and your habit of putting the source account in the corner (that card's **GG-1184**).
- **The clip** → the leftover pack on the pass is stamped **5 × 7**, kraft. Every card is built to print at exactly that size on that stock.
- **The register list** → I didn't retype it. The list you sent Marguerite the night before ("*New price list, into the register tomorrow at seven*", attachment **price list 5 may.pdf**) is the source of truth — the nine that moved, and the two you held.
- **The two letters** → the *why*: **Golden Glen Creamery** (GG-1184, butter $4.68 → $5.94/lb) and **Samish Bay Mill** (ref SBM/HAU/2026-04, flour $41.50 → $46.25/sack).
## The set (one 5×7 card each, kraft, dark-pen handwriting)
Butter cards carry **GG-1184** in the corner; flour loaves carry **SBM/HAU/2026-04**; coffee carries the roaster/dairy — same corner-source habit as the card that's up.
1. Kanelsnurr **$4.75** (was 4.25) — *"Butter went up. You've bought this one since we opened, so you'll feel it."*
2. Kardemommebolle **$5.00** (was 4.50)
3. Skolebrød **$5.25** (was 4.75) — *"Same creamery as March."*
4. Wienerbrød **$5.75** (was 5.25) — *"The one that's all butter."*
5. Croissant **$5.00** (was 4.50)
6. Rugbrød loaf **$9.50** (was 8.50) — *"The mill in Bow, first of April."*
7. Solsikkebrød loaf **$9.00** (was 8.00)
8. Drip coffee **$3.25** (was 3.00)
9. Latte **$5.25** (was 4.75)
I also added a tenth companion card — *"These didn't move. Day-old bag still $5. Kids' bolle still $2."* — since you were firm those two stay put. Drop it if you only want the nine.
## Printing
Open it, `File → Print`, paper **5×7** at **100% / actual size** (no scaling), margins **None**, and turn on **Background graphics** so the kraft colour prints. One card per sheet; the grey instruction strip doesn't print.
Two calls I made that you can veto: the coffee/latte reasons ("beans went up," "beans and milk both up") aren't backed by a letter the way butter and flour are — I kept them plainly honest rather than inventing a supplier notice. And I used proper ø spelling (Skolebrød, Wienerbrød, Rugbrød, Solsikkebrød) to match your hand, not the ASCII on the register PDF. Say the word if you'd rather they match the till exactly.
From the taskLeg A prompt, turn 1, to Model A
Leg A turn 1
My manager told me not to put a card in the window. I put one up anyway, in my own handwriting, and it did what I needed. I still don't have a proper set of window cards for the ones that moved, written in the same voice and sized to the stock I actually buy. The photograph is the card that's up. The short clip is the leftover pack on the pass. The list went into the register that morning. The letters from the creamery and the mill are already in my mail. I'm not typing the list out again. Can you build the set in a file called storvik_price_cards.html?