← Week 6 portalMaster file in the engagement folder:
my-work/week-6/fund-level/process-log/process-log-v2.md

Week 6 process log (v2): what worked, what did not, and what it took

Record ids. D-6-n is a Week 6 decision, J-n a journal entry, F-0n a review finding, DR6-n a data request, A-n an assumption; all are in the engagement record.

D6, part 1 of 3. Final version. Rulings are by Daniel Vetsko (engagement owner). Facts come from the Week 6 record (my-work/week-6/state-live.md; look up any id with kit_tools.py query). Where a number is a count from the record, it says so. Version 2 drops when each stage happened (D-6-63), because it does not bear on how the work is done again, and adds the close-out (section 7). Effort hours were not measured, so this log reports counts and order, never effort.

1. The shape of the week

PhaseWhat happenedWhere to look
Scoping and scaffold22 scoping decisions (D-6-1 to D-6-22), the week scaffolded, a blank-shell scout sandbox set up (D-6-9, D-6-15 to D-6-18)J-001
Scout runA sandbox build with no client data: a 32-concept register, four framework files, an engine, 21 QA checks, a ten-company synthetic portfolio, a query pack. Its proposals became the owner's agendamy-work/experiments/blank-shell-disclosure-v1/final-report.md; J-002, J-003
G1 specRead the scout, the data room (250 files), the protocol and the frozen weeks; wrote the charter (15 rules), data map, 20 assumptions and the spec; ten owner rulings; signed (J-021)J-004 to J-021
BuildIntake, engine copy, framework verification, D2 to D5, two separate-session tests of the query packJ-023 to J-044
G2 resultsGate checks, reviewer packets, two reviewers, one confirmation pass, signed (J-056)J-045 to J-056
G3 handoffProcess log, Week 1 reconciliation, pricing reflection, hand-off notes; frozen (J-059)J-057 to J-060
Close-outEvery open item closed or handed to Northwood; Scope 3, PCAF and the exposure by fund built; all outputs reissued as v2 under a final gateSection 7; D-6-60 to D-6-67

2. Counts from the record

3. What worked

  1. Scout first, gated build second. The blank-shell scout produced a complete, runnable library with no client data in its head. The gated build then spent its time on differences from the scout, as the Week 4 lesson said.
  2. Running real data through synthetic-tested code. The scout's synthetic tests passed. Real data exposed two defects at once: the engine added the fund roll-up row to the company rows (Scope 1 would have read 23,899 t against 11,599 t) and the LP report printed company slugs and internal ids that the scout's name check missed. Both were fixed and each now has a check and a failing-case test.
  3. A clean re-run check that copies only the pinned inputs. It found two unpinned inputs before G2 and proves the leave-behind rebuilds byte for byte.
  4. Reading the output as the audience. Two separate Claude sessions given only the query pack found nine gaps and one real inconsistency. A reviewer reading as Sarah found the intensity figure did not recompute from the numbers printed beside it. A confirmation pass caught a misattribution (the executive assignment was the engagement owner's decision, not Northwood's).
  5. Questionnaire reads with quotes checked by script. Thirty answers were entered with a verbatim quote; the intake stops if a quote is not in the file.
  6. Decisions in the owner's words, as they were made. Every ruling went to the decision log the same turn, so packets could point to ids instead of re-explaining.
  7. Verifying frameworks against official text, and saying what was not read. The verification log found three errors in the scout's library (ESRS 2 governance numbering, the EDCI metric list, and the claim that UK SRS use is voluntary).

4. What did not work, or cost more than it should have

  1. The scout assumed an EU-facing LP base. Northwood's data room names no EU or UK LP, so the Tier 1 EU frameworks are library coverage, not what Northwood is asked for (D-6-39).
  2. Synthetic data hid the double count. A check on the headline figures (Q22) was added only after real data broke it. Put a fund-level roll-up row in every synthetic case, and a recompute check in the first version of any engine.
  3. Several sources could not be read the normal way. The IFRS S2 text is behind a login (AASB and UK SRS texts were used as the proxy). PDFs needed a workaround and each read needed the owner's yes. Plan for this in the next reference build.
  4. A gate step ran out of order. At G1 the sign-off was recorded after a check showed a stale-state false alarm. The sequence is now: regenerate the state, run the check, read the verdict, then record and freeze.
  5. Inputs and ids collided with the kit's rules late. Derived row ids read as decision ids, file locators needed full paths, two inputs were unpinned. Run lint after the first deliverable, not before the first packet.
  6. Week 4's "carbon" column was the wrong thing to show a board. It is the customer carbon-data line plus a small carbon-price piece. Before quoting a frozen column, recompute one example by hand.
  7. Peer benchmarks are thin. Public data supports a comparison for only two companies.
  8. The re-run needs a technical person. The nine steps are commands; the re-run note says who runs them and what Sarah reviews.
  9. A builder wrote into a frozen folder at close-out. The first v2 run of the dashboard builder wrote the query pack into the v1 folder, overwriting one frozen file. It was restored from the archived v1 builder and hash-checked against the freeze (J-072). Lesson: give every versioned output its own folder name, and hash-check the freezes after the first run of a changed builder.

5. Carry into the next engagement's reference build

6. The scout's 23 guesses: what is known now

GuessResult
1 One row per company, concept and periodHolds after the malformed-row repair and the slug join
2 Scope 2 location-based and market-based kept apartHolds; market-based is not calculated (DR6-05)
3 One dollar loss per company for a named scenarioBroken: Week 5's effect is a ranged forward figure; the range sits on the data row and a range token prints it
4 Targets carry a baseline year and methodBroken: no company has a target
6 Governance names one responsible bodyBroken: no body exists at any company
7 One sign-off contact per companyPartly: roles range from CEO to VP of Risk; two companies name two people
8 Sector plus size band does not single a company outBroken: four sectors hold one company; descriptors split by fund (D-6-57)
9 One fixed qualification sentenceConfirmed by the protocol
10 Revenue per company in USD millionsHolds; the period (twelve months to January 2026) differs from the emissions year
11 and 12 EU LP requests and EU factsUnsupported: no EU LP is named
13 Narrative figures are tokensHolds; count and range tokens were added
14 Stable company idsBroken in two source files; a join map resolves it
15 Earlier disclosures existConfirmed absent; Year 1
16 Board-page effect wording is a fixed lineReplaced by real effects from Weeks 4 and 5
20 Public peer figures per peer and metricBroken for company-level peers; ENERGY STAR site benchmarks for two companies (D-6-58)
22 California business test answerable yes or noNot answerable from the data room; a legal question

The remaining guesses (5, 17, 18, 19, 21 and 23) stay untested or are legal: they are on the assumptions register (A-14, A-20).

7. Close-out

The owner ruled on every open item (D-6-60 to D-6-67):

Portfolio edition: Kith Climate cohort case, fictional client.