A number is not a second answer.#
Post-Date Analytics and The “Second Date” Metric
SLOP1276 · Week 11 · Alex’s story
Fictional Alex’s cutoff catches one agreement; the complete trace contains two. The frozen profile ranks 2/100. Tidier headlines would require changing the question.
Investigate: Can the release preserve both context and disappointing results?
The engineering output#
Build a Follow-up evaluation: observation rules, time accounting, frozen benchmark results and a reproducible release candidate.
Keep one disappointing result and explain it.
Every record, rating and probability is authored teaching data.
How this pack fits the week#
60 minutes · Monday’s selected lecture slides.
90 minutes · Thursday’s four practical blocks.
About 210 minutes · independent preparation, remaining examples and consolidation.
The complete 180-minute route is a resource guide; its break is for a full sitting.
Bring the evidence#
| Input | What it establishes |
|---|---|
| Week 5 freeze | Which text and ratings were evaluated |
| Week 9 agreement | Which plan and costs were confirmed |
| Week 10 export | Which coded events were recorded |
A Markov distribution supplies none of the missing observations.
“Nice time” is a quote#
“I had a nice time” records words. It does not establish a second-date agreement.
An appropriate optional question can ask for an answer. It cannot require one.
The workbook interprets coded events, never free-text subtext.
Define the labels first#
| Evidence | Label |
|---|---|
| Both explicitly agree | agreed |
| Either declines or withdraws | declined |
| One agreement only | pending |
| No relevant coded observation | unobserved |
“Nice time” supplies text, not the second agreement.
Exercise · freeze the observation window#
5 minutes. Alex agrees at 1085, the counterpart at 1086. The meeting ends at 1090.
Use the half-open window [0, 1086).
What is the cutoff label? What is the latest label in the completed export?
These are Week 10’s fictional minutes since midnight.
Answer · two legitimate views#
| Minute | Record | Before 1086? |
|---|---|---|
| 1085 | Alex agrees | Yes |
| 1086 | Counterpart agrees | No |
| 1090 | Meeting ends | No |
At cutoff: pending. Latest export: agreed + meeting ended. Retain both views.
Agreement can later be withdrawn#
In the separate branch, the counterpart declines at 1087.
Before 1087: agreed. Before 1088: declined.
Retain the events and the cutoff. Do not erase the withdrawal.
A declined invitation cannot be retried. Ending the meeting remains a separate field.
Exercise · double the headline#
4 minutes. Eight distinct authored records: two agreed, two declined, three pending, one unobserved.
All use the same 48-hour reporting rule after each meeting.
Find agreement among all records and among resolved records.
These rows are separate from Week 10’s alternative replay branches.
Answer · the denominator did the work#
| Label | Count |
|---|---|
| agreed | 2 |
| declined | 2 |
| pending | 3 |
| unobserved | 1 |
2/8 = 25% of all records. 2/4 = 50% of resolved records. Name the denominator.
Own time, invented benefit#
The key concept is ROI of time spent vs. romantic viability.
Our proxy is an allocation model for Alex’s own time. Its benefit units are chosen, not measured viability.
No fraction assigns a price to another person.
Account for all 115 minutes#
- Together
- 90 min
- Travel
- 20 min
- Compose
- 5 min
Exercise · count the bus twice#
4 minutes. Return travel is [100, 110). Composition on the bus is [105, 110).
Can you add ten travel minutes and five composing minutes to report elapsed time?
What changes if a ten-minute interval is missing?
Answer · define the accounting boundary#
The two activities overlap by five minutes. Summing them double-counts elapsed time.
The workbook rejects overlaps and reports gaps as unallocated time. Touching endpoints are valid.
Week 9’s 90-minute itinerary, this 115-minute log and Markov steps are distinct.
Units survive division#
| Calculation | Meaning |
|---|---|
| 10 units / 115 min | 0.087 units per minute |
| 10 / (115/60) | 5.217 units per hour |
| (benefit − cost) / cost | Needs commensurable units |
This course’s “ROI” is a benefit/time proxy. Minutes cannot be subtracted from benefit units.
Change the invented numerator#
- 0 units
- 0
- 5 units
- 0.043478
- 10 units
- 0.086957
- 20 units
- 0.173913
Exercise · the next action#
5 minutes. Reuse Week 7: stop = 0, wait = 2q−1, clarify = 7q−5 if separately admissible.
Find stop/wait and wait/clarify crossover values.
At q=0.9, exclude clarification. Then add an explicit refusal.
q is assumed; five utility units are not five minutes.
Answer · compare admissible options#
| Assumed q | Stop | Wait | Clarify |
|---|---|---|---|
| 0.3 | 0 | −0.4 | −2.9 |
| 0.5 | 0 | 0 | −1.5 |
| 0.8 | 0 | 0.6 | 0.6 |
| 0.9 | 0 | 0.8 | 1.3 |
U(wait)=2q−1; U(clarify)=7q−5. A refusal excludes continuation regardless of q.
The constraint precedes the optimiser#
Crossovers: q=0.5 and q=0.8.
At q=0.9 without clarification, wait maximises the remaining toy utilities.
With an explicit refusal: stop the invitation before optimising.
Past effort cannot authorise continuation. A reporting cutoff prescribes no universal delay.
Pause for the full route · 10 minutes#
For an uninterrupted full-pack session; skip this slide in Monday’s selection.
When you return, keep the failed target.
We will evaluate the exact candidate before revising it.
A candidate is more than four numbers#
- FreezeExact text, evidence, ratings, version
- EvaluatePrimary rule + unchanged controls
- PreserveSave score, ties and failed target
- ExploreSeparate weight and case changes
A digest identifies the retained bytes. It does not prove who wrote or approved them.
The worked teaching snapshot#
“Board games, bad puns, maybe a relationship? Meet Friday 5-7 pm: bus-friendly, $20 total. Fine to pass.”
Exact Week 4 supported repair, paired with Week 5’s library plan.
Assembled now as a teaching example. Use your own Week 5 freeze first.
Evidence before ratings#
| Feature | Rating | Evidence |
|---|---|---|
| Clarity | 4 | Possible relationship |
| Specificity | 4 | Games + bad puns |
| Feasibility | 3 | Time, bus, $20; no venue |
| Exit | 4 | Fine to pass |
A notebook route does not add missing venue detail to the public bio.
Exercise · why not four for feasibility?#
4 minutes. The notebook’s library plan needs 100 minutes and $14. Alex has 120 minutes and $20.
The public bio states the time, bus and budget but gives no venue or route.
Use the published rating anchors to defend the 3. What can a quote-presence test actually establish?
Answer · the notebook is not the bio#
The plan fits. The public invitation is usable but omits a relevant detail: rating 3.
A quote test establishes that the quote occurs. It cannot validate a reviewer’s judgment.
Keep the quote, reason and case reference. A score increase needs evidence, not a nicer rank.
Exercise · score the tie#
6 minutes. Primary score = 100 × (c+s+f+e) / 16.
Candidate (4,4,3,4). Control 099 (3,4,4,4).
Rank = 1 + count(control score ≥ candidate score).
Does “joint first” satisfy the course’s top-1% contract?
Answer · a useful failed target#
| Comparator relation | Controls |
|---|---|
| Greater than 93.75 | 0 |
| Equal to 93.75 | 1 |
| Less than 93.75 | 98 |
Rank = 1 + 0 + 1 = 2/100. The target fails; the evaluation is still useful.
The 100-entry contract#
99 controls + one candidate = 100. Ties count against the candidate.
Only rank 1/100 receives this benchmark’s label. Rank is not graded.
Controls are deliberately fabricated vectors. They predict no app rank or romantic outcome.
Do not round scores before counting ties.
Change one rule, consistently#
| Rule | Numerator | Maximum |
|---|---|---|
| Primary | c + s + f + e | 16 |
| Double feasibility | c + s + 2f + e | 20 |
Multiply the ratio by 100. Recompute the candidate and all 99 controls.
Exercise · rerank the weak-feasibility probe#
7 minutes. Use the separate vector (4,4,1,4).
Derive its primary and doubled-feasibility scores. Recompute every control.
Count strictly above and equal separately.
Why is comparing new candidate scores with old controls invalid?
Answer · 81.25 becomes 70#
| Rule | Score | Above / ties | Rank |
|---|---|---|---|
| Primary | 81.25 | 1 / 3 | 5/100 |
| Double f | 70 | 12 / 9 | 22/100 |
The worked release is a different vector: 93.75 → 90, rank 2 → 2.
The same rank can hide a change#
Release example: 93.75 → 90, rank 2 → 2.
Originally one control ties. With doubled feasibility, one scores strictly higher: 95.
Optional: compare numerators 12+3w and 11+4w.
Their ordering changes at w=1; the common denominator remains positive.
Change the case separately#
Preserve the original and weight-sensitivity results.
Now author a distinct availability case: Friday 5–6:30 pm, or 90 minutes.
Keep the original text and 100-minute library itinerary while auditing what fails.
This stress test does not rewrite Alex’s fixed biography.
A hard constraint fails#
| Library plan | Original | Changed |
|---|---|---|
| Available minutes | 120 | 90 |
| Needed minutes | 100 | 100 |
| Time slack | +20 | −10 |
| Case fidelity | Pass | Fail |
Unchanged text still says Friday 5–7 pm. Rank 2 cannot repair a failed time constraint.
Exercise · repair by substitution?#
6 minutes. The bio still says Friday 5–7 pm. The library plan still needs 100 minutes.
List the affected text, invitation and agreement records.
Can the unchanged rank justify release? Can you substitute Week 9’s 90-minute itinerary without revising the plan?
Answer · retain the contradiction#
Feasibility becomes 2: specific but contradicted. Vector (4,4,2,4) scores 87.5, rank 2/100.
Case fidelity fails, so release under the changed case is blocked.
A new plan or bio needs its own version and review. Another venue’s duration is not this itinerary.
The controls have no source bios#
Weight change: rescore all 100 numeric vectors.
Availability change: review the candidate’s actual text and plan. Keep controls fixed as comparators.
We lack control texts to re-rate them for new availability.
Call this a candidate stress test, not a measured population shift.
Evaluation can become development#
- Declared evaluationFreeze → score → keep result
- ExplorationInspect → edit → new version
Exploration is allowed. Its new score is not an untouched evaluation of the old candidate.
Exercise · classify the audit trail#
5 minutes. A: save the primary result, then run declared sensitivity.
B: inspect the controls, edit the bio and retain both versions.
C: erase the failed result and call the new score the original.
D: raise feasibility because rank 2 looks disappointing.
Answer · preserve what happened#
A: prespecified sensitivity. B: labelled exploration.
C: misrepresented history. D: unsupported rating.
scikit-learn’s leakage guidance explains why evaluation information must not steer model choices.
A public synthetic benchmark never becomes a hidden test set merely because you froze a file.
Challenge · make a check fail#
5 minutes. In a disposable copy, exclude ties from rank or include events exactly at the cutoff.
Predict the failed literal answer first: rank 2, or pending before 1086.
Run, inspect the failure, restore and reproduce.
A green check cannot supply missing evidence or prose judgment.
Sixteen files, one command#
Save the lecture downloads in one folder.
Run node week-11-worked-examples.mjs.
The report contains time, labels, all control scores, failed constraints, earlier costs and exact/seeded Markov results.
No packages or network are needed after download.
Generate a concrete example release#
Run node week-11-worked-examples.mjs –release-dir ./example-rc1.
Open index.html. Its candidate paragraph comes from the frozen source.
The directory also contains evaluation JSON, a manifest and a README.
An existing directory is rejected so a prior release is preserved.
Make the handoff reproducible#
- SourceSixteen files + exact frozen candidate
- OutputStatic HTML + evaluation JSON
- EvidenceHashes + actual keyboard/viewport log
- MaintenanceClean reproduction + software rollback
Restoring a software version cannot restore another person’s agreement.
Exercise · inspect the rendered release#
10 minutes plus discussion. Compare the candidate paragraph with its source.
Tab and Shift+Tab to the evaluation link; activate it and check visible focus.
Inspect 1920×1080 and 390×844, then resize and zoom.
Record the actual browser, version, action and result. The generator claims no browser pass.
Debrief · two kinds of evidence#
A hash checks bytes against a retained record. Rendering shows whether those bytes are legible and reachable.
W3C page structure supports meaningful regions and headings.
W3C Easy Checks is an initial accessibility review, not complete conformance proof.
Reproduce before maintaining#
Run the workbook into a new directory, example-rc2. Compare output bytes and manifests.
Keep source downloads and record the runtime version separately.
Week 12 demonstrates software rollback and writes the maintenance/exit manual.
Restoring HTML cannot restore somebody’s agreement.
The existing final-project evidence#
Save labels and cutoff, the 115-minute critique, primary failed target, common-weight sensitivity and changed-case failure.
Attach the release and actual browser checks to the existing validation appendix.
Final project: 50% · 28 May 2027 · 5 pm.
No additional essay or real romantic outcome is required.
Keep the uncertainty legible#
An observation label is not a simulation probability.
A benchmark rank is not case fidelity.
A benefit/time ratio is not permission to continue.
Week 11 tutorial · Worked notes and downloads
Alex’s release keeps rank 2/100 and its limitations. Week 12 asks someone else to reproduce and restore it.