/score-audit, expected score and the next cheapest point

The point-budget equivalent of a weight budget. Adapted from SumoX-26's /bom-builder, which did the same job for grams.

Answers one question: given where the project actually is, what is the expected qualification score, and what is the cheapest point still unclaimed?

The model

Qualification %  =  0.50 × (trial points / 19)
                 +  0.20 × code
                 +  0.10 × video
                 +  0.20 × report

Full breakdown in docs/SCORING.md.

Steps

1. Read results/trials.csv. Take the best trial score, since the committee takes the best of three attempts. If the file is empty or missing, say so plainly: trial points are 0 banked, and everything else is projection.

2. Score the trial component honestly, per item, marking each banked or projected:

   Item	Points	Status
   Column publish	+1	banked / projected / not started
   Column annotated image	+2
   Navigate to column (needs the grasp)	+3
   Row publish	+1
   Row annotated image	+2
   Grasp	+3
   Navigate to bin (needs delivery)	+3
   Place, dropped +2 or gentle +4
   Collisions observed	-0.5 each

   Apply the two contingencies. Navigation to the column scores nothing without a successful grasp, and navigation to the bin scores nothing without delivery. Do not count them as banked while the grasp is unproven.

3. Assess the three deliverables against their stated criteria, not against a feeling:

   Component	Weight	Assess by
   Code	20%	Are perception, planning and execution in distinct nodes? Does every stage have a timeout and a retry budget? Are recovery paths present and tested? Are comments explanatory?
   Video	10%	Does a take exist that satisfies all six constraints? Has a full run been timed against the 5-minute ceiling?
   Report	20%	Do the six sections exist? Do results/trials.csv and the RViz screenshots contain the required numbers and figures?

   The video and report should be scored as near-zero until the artifacts exist. A plan to write a report is not 20%.

4. Compute the expected score on the 100-point scale, showing the arithmetic, and separate what is banked from what is projected.

5. Find the cheapest unclaimed point. Rank remaining work by qualification points per hour, and prefer certainty. Reference values:

   Work	Qualification points	Certainty
   Each vision publish	2.6	High. Pure OpenCV
   Each vision annotated image	5.3	High
   Gentle place instead of drop	1.05	High once the grasp works. Hours of Cartesian pathing
   The grasp, with its withheld navigation points	15.8	Low. The keystone
   Delivery chain after the grasp	18.4	Medium
   A recovery path	Part of the 20% code component	High
   The report's six sections	up to 20	High, if the data exists
   The video	up to 10	High. A checklist
   Speed optimisation	0 outside a tie	n/a

6. State the one recommendation, with the arithmetic behind it.

Output format

## Score audit, <date>, <days> days to 2026-09-15

### Trial component (50%)
Best logged trial: <n>/19  →  <n × 0.50/19 × 100> qualification points
Banked: <list>
Projected: <list, with what each depends on>

### Code (20%)
<assessment against modularity, error handling, comments>  →  <estimate>

### Video (10%)
<does a compliant take exist>  →  <estimate>

### Report (20%)
<do the sections and the data exist>  →  <estimate>

### Expected qualification score
Banked:    <n> / 100
Projected: <n> / 100

### Cheapest unclaimed point
**Do next:** <one thing>
**Worth:** <qualification points>
**Cost:** <rough hours>
**Certainty:** high / medium / low
**Instead of:** <what this displaces, and why that's right>

Rules

• Banked means a logged trial recorded it. Everything else is projected and must be labelled so. This distinction is the entire point of the skill.
• Apply the contingencies. The two navigation +3s are the most commonly over-counted points in this challenge, because they look earned when the robot drives correctly and they are not.
• Score the deliverables as near-zero until the artifacts exist. An intention is not a mark.
• Prefer certain points over uncertain ones when the hours are comparable.
• Never recommend work that isn't on the rubric. No dashboards, no backup perception pipelines, no visualisations. There are 29 days.
• Recompute rather than reusing a previous audit's numbers. The point of this skill is to reflect the current state.
• If the 2026-09-06 grasp decision is due or overdue, say so and give the arithmetic for both paths.