Sample · Week 2
One extract from the Week 2 lab build: Meridian Textiles
Week 2 rebuilds the Scope 1 and 2 inventory you calculated by hand in Week 1, this time with Claude working from a skill file you write. Below are three documents from that build. Meridian Textiles is a fictional company, the teaching case for all six weeks; the lines below come from its teaching pack, not from a real plant, and quantities are left out on purpose. On the Thursday the mentors build it live on Meridian. On the Sunday you build your own, on an unfamiliar company, and hand in the same kind of trail and iteration record.
1. Ledger lines, classified
The agent maps each line of the utility and procurement ledger to scope and category, with a confidence column. Anything it cannot place with confidence carries a mandatory “needs human review” flag.
| Line | Item | Scope | Category | Confidence |
|---|---|---|---|---|
| 0112 | Diesel, DG sets, Unit 1 | 1 | Stationary combustion | High |
| 0127 | Coal, boiler house | 1 | Stationary combustion | High |
| 0203 | Grid electricity, Unit 1 | 2 | Purchased electricity | High |
| 0204 | Grid electricity, Unit 2 | 2 | Purchased electricity | High |
| 0311 | Natural gas, dyeing line | 1 | Stationary combustion | High |
| 0417 | Refrigerant top-up, chiller | 1 | Fugitive | Medium |
| 0418 | LPG, canteen | 1 | Stationary combustion | Medium |
| 0472 | Effluent treatment | 1 or 3 | Depends on the boundary memo (the 40% effluent JV) | Needs human review |
Scope and category only. The boundary decision behind line 0472 stays with you; the agent flags it and does not decide it.
2. Calculation trail (extract)
The arithmetic runs in executed code, not in prose, and the agent may only use the factor library supplied to it. Every line cites the factor row it used. Figures are hidden here; the point is the trail.
| Line | Activity data | Factor row cited | Result |
|---|---|---|---|
| 0112 | IPCC 2006 Guidelines, Vol 2, Table 1.4 CH₄ and N₂O: Vol 2, Table 2.3 · IPCC AR6 GWP-100 | ||
| 0127 | IPCC 2006 Guidelines, Vol 2, Table 1.4 CH₄ and N₂O: Vol 2, Table 2.3 · IPCC AR6 GWP-100 | ||
| 0203 | CEA CO₂ Baseline Database, v20 India grid, location-based | ||
| 0204 | CEA CO₂ Baseline Database, v20 India grid, location-based | ||
| 0418 | None | Stopped · no factor in the loaded library |
Line 0418 is the lesson. With no factor in the library, the agent stops and asks. It does not supply one from memory.
3. Iteration record
Each lab deliverable gets two marks, recorded separately: one for technical correctness, from the sustainability mentor, and one for iteration quality, from the AI mentor. This record is what the second mark reads.
| # | Prompt | What came back | Decision, and why |
|---|---|---|---|
| 1 | Calculate Scope 1 and 2 for the classified ledger. | Totals written in prose. No code was run, so there was no trail to check. | Rejected. Arithmetic in prose cannot be audited. Rerun with an instruction to write and run code. |
| 2 | Rerun as executed code, citing a factor row for every line. | A full trail. Line 0418 carried a factor that looked right, but no row in the library matched it: it came from the model’s memory. | Rejected. An invented factor. Refusal rule added to the skill file: if the factor is not in the supplied table, stop and ask. |
| 3 | Rerun with the updated skill file. | The same trail, except that the agent stopped at line 0418 and asked for the factor. | Accepted. The gap is now visible instead of silently filled, and a person decides what goes in. |