|Kith Consulting
← All documentsPortfolio edition · fictional client
01Guide

How to re-run and update this work

This guide is for whoever at Northwood updates the work next quarter. The short version: replace the data, set a new report date, run the numbered scripts in order, and rebuild the pages. Every script refuses to overwrite a file, so each run goes in its own new folder. The full run order, with what each script reads and writes, follows.

How to re-run and update the Week 4 materiality work (re-run note v18)

What this covers. The five deliverables: the connected dataset (D1), the financial exposure model (D2), the extended data (D3), the materiality matrix (D4) and the initiative ranking with the one-page view (D5), plus the explorer: one page that puts the exposures and the initiatives in one ranked value-creation backlog, with the cost, effort and payback of acting beside the exposure. Version 12 replaced the earlier client deck and portal with the explorer; version 18 adds the revision layer and rebuilds the deck, explorer, exposure model, matrix and initiatives page, data book, re-run guide and portal as one delivery set. Versions 8 and 9 add the ten-company rebuild: all ten portfolio companies run through one protocol (documented inputs for the six companies outside the priority four, an evidence sweep of the data room, the customer carbon-data requirement as rule R12, and the cost of responding to it as a labelled estimate). Version 9 applied the findings of its review and version 10 settles the findings left after it: the company inputs behind the exposure model sit in two tables you edit, not in code; the physical-weather lines follow one set of rules (a loss someone else bore is not the company's, a company that reports no loss is shown as reporting none, an unreported loss is derived only from the company's own files); the sweep reads its recorded judgments and its percentage ranges as ranges; every benchmark behind the cost estimate has its address and the date it was checked; and a held-out check of the sweep's reading is shipped with its labels. Everything is rebuilt by 24 Python scripts run in a fixed order. Nothing needs a Kith tool, a Kith account or a Kith person. You need Python 3 (the delivered run used 3.9.6), the two add-on packages in requirements-v1.txt, openpyxl and xlrd (python3 -m pip install -r requirements-v1.txt; they read Excel files) and your own data room.

The revision layer (steps 23 and 24). After the explorer, the work was revised and finished as one delivery set for the reader at Northwood: all ten companies, the priority four marked, no exposure line left without a figure or a stated reason, one exit timing per company, and one portal. Step 23 writes a new version of the ten-company data layer (connected dataset v4, exposure lines v6, matrix v5, ranking v6) in which gaps are estimated from published benchmarks or from the portfolio's own observed figures, marked grade C, and each estimate is cross-checked in revision-validation-v1.csv. Step 24 builds the client deck, the explorer, the exposure model, the matrix and initiatives page, the data book, the re-run guide and the portal, with copies of the data files. The method for the revision is in charter-v7-addendum-v1.md; every outside figure used is in revision-benchmarks-v1.csv with its source, date and sentence.

What you receive (the full list, with a fingerprint for each file, is handover-list-v17.csv in this folder):

  • requirements-v1.txt, the exact package versions used, and setup-rerun-folder-v3.py, which sets up a fresh folder for a run;
  • the 24 scripts, in my-work/week-4/pipeline/, the two helper modules the evidence sweep imports (sweep_logic_v6.py, sweep_logic_v8.py), the five helper modules the page builders import (w4_model_v1.py, w4_pages_v1.py, w4_content_v1.py, w4_deck_v1.py, w4_process_v1.py), and kit.py, the small helper that stamps each run into a log and refuses to overwrite an existing file;
  • the reference data in my-work/week-4/reference-data/ (EIA prices, FRED rates, FEMA disaster declarations, the IAC energy-saving database), each with its source and download date in the assumptions register;
  • the assumptions in three tables you can open in Excel: assumptions-register-v3.csv (A-01 onward: value, unit, source, date, who owns it, which deliverable uses it, grade A to C) g2-assumptions-v1.csv (the inputs behind the Clearview fleet estimate, labelled as an estimate) and assumptions-register-additions-v7.csv (A-44 and A-45, the day rate and effort days behind the cost-to-respond estimate; step 21 appends them to the register from step 19 to make assumptions-register-v10.csv);
  • the method: charter-v7.md (the rules for probability, dollar impact, bands and ranking, and for the customer carbon-data requirement, the one protocol for all ten, the evidence sweep, the cost to respond and the explorer), data-map-v1.md (which data-room file supplies each input), g2-contract-v1.md (the shape of every output file), rule-r12-v2.md (the customer carbon-data requirement rule) and d5-response-cost-method-v2.md with its inputs response-cost-inputs-v3.csv and response-cost-components-v1.csv (the components, effort ranges and day rate behind the cost estimate), and the two tables behind the exposure model: d2-source-inputs-v1.csv (every company input with its file, line and expected text) and d2-company-rubric-v1.csv (the customer-line judgments per company);
  • the evidence-sweep record the sweep script reads: sweep-commissioning-v7.csv (the tests the sweep logic passed and the fingerprints of that logic) and sweep-judgments-v4.csv (the 42 paragraphs a reviewer has judged; the sweep leaves them out of its queue). The harness that recommissions the sweep logic if you change it is test-sweep-v7.py (with test-sweep-v1.py). The held-out check of the sweep's reading is sweep-holdout-sheet-v1.csv (12 paragraphs), sweep-holdout-key-v1.csv (the script's reading), sweep-holdout-agent-v1.csv (the review agent's blind reading), sweep-holdout-labelled-v1.csv (the labels) and sweep-holdout-score-v1.csv (the comparison) and sweep-qualexist-queue-v1.csv (the twelve paragraphs the qualified-existence rule queued), with build-holdout-batch-v1.py and score-holdout-v1.py to repeat it;
  • the revision inputs: revision-benchmarks-v1.csv (every published figure used for an estimate, with its source, vintage and quotation), charter-v7-addendum-v1.md (the rules the revision rests on), explorer-owners-v6.csv (owners and first moves, reworded for three newly sized measures) and process-snapshot-v3.json (a reading of how the work was gated and reviewed, shown in the back-end note; it describes this delivery and is not recomputed);
  • the page assets in my-work/week-4/leave-behind/ui/ (the look of the explorer, from the Week 3 client deck);
  • the Week 3 carbon files that D1 and D2 build on, and the delivered outputs themselves, so you can compare a new run against them.

Set up for each run

  1. Keep the received files unchanged in one folder (the received folder): my-work/... at the paths in handover-list-v17.csv, and your data room at data/northwood-capital/.
  1. Run, with your own report date: python3 my-work/week-4/leave-behind/setup-rerun-folder-v3.py <received-folder> <new-folder> --report-date YYYY-MM-DD. Use a new folder for every run (it refuses a non-empty one, and the scripts never overwrite). The setup script copies every file the scripts need, but not the delivered outputs, which stay in the received folder to compare against. It creates the empty output folders, adds the marker file CLAUDE.md (the scripts look for it to find the project root), links your data room, puts the decisions behind the delivered files (id, date, summary, decision and rationale; 129 rows, including the two pre-registration decisions the extended-data script checks) in record/decision-log-v1.csv, and writes the settings file as my-work/week-4/manifest-v1.json.
  1. manifest-v1.json is the only place the report date lives; no script has it typed in. It also holds the input pins: the fingerprint (SHA-256) of every data-room, Week 3 and reference file the work was built on. Step 2 of the run order stops if a pinned file no longer matches. Version 4 of the settings file also pins the data-room files the evidence sweep and the explorer read (including Northwood's firm overview). Use the delivered date only to reproduce the delivered numbers exactly.

Run order

Run each from the new folder: python3 my-work/week-4/pipeline/<script>. Each script stops with a plain message if a table header, a period label or a cross-check no longer matches, and none will overwrite a file that exists (see "Updating" below).

#ScriptReadsWrites
1profile-data-room-v1.pydata roommethodology/data-room-profile-v1.csv
2extract-register-values-v5.pyreference data, data room, Week 3 ledgermethodology/register-extraction-v5.csv (register values re-read from source, for comparison)
3build-connected-dataset-v3.pydata room finance files, Week 3 carbon, register, step 2 (stops if register N disagrees with the report date)D1 fund-level/connected-dataset-v1.csv and its lineage file
4build-event-table-v1.pycompany questionnairesfund-level/pooled-event-table-v1.csv
5build-fema-frequency-v1.pyFEMA declarations, facility locationsfema-distinct-disasters-v1.csv, fema-frequency-v1.csv
6build-simulated-baseline-v1.pyfacility locations, Harvest Ridge questionnaire, the two pre-registration decisionsD3 simulated-extended-data-v1.csv (every value labelled simulated)
7build-exposure-model-v1.pysteps 3 to 5, Week 3 exposure, registerD2 exposure-lines-v1.csv, reverse-stress-v1.csv, exposure-model-v1.html
8recompute-sterling-v1.pySterling source files, steps 4 to 7recompute-sterling-v1.csv (an independent re-calculation of Sterling's numbers)
9scan-data-room-measures-v1.pydata roommethodology/d5-scan-files-v1.csv, d5-scan-hits-v1.csv (which companies state a cost-and-saving energy measure)
10build-d5-initiatives-v1a.pyquestionnaires, EIA, register, step 7first-pass initiatives table (kept only because step 12 reads it)
11build-d5-ranking-v1a.pystep 10first-pass ranking (same reason)
12build-g2-view-v1.pysteps 7, 10, 11D4 matrix d4-matrix-v1.csv, d4-aggregates-v1.csv (the other files it writes are first-pass and superseded)
13build-d5-initiatives-v3.pyquestionnaires, EIA fuel prices, register, step 7d5-initiatives-v2.csv and methodology/g2-assumptions-v1.csv
14build-d5-ranking-v3.pystep 13, registerD5 d5-ranking-v2.csv, d5-sensitivities-v2.csv
15build-g2-view-v3.pysteps 7, 12, 13, 14d4-summary-v1.csv, g2-lineage-v3.csv, the one-page view and the four company packs
16extract-six-values-v2.pythe data room (six companies' acquisition, board, budget, status and plan files), the spine company tablemethodology/register-extraction-six-v2.csv (documented exit timing and gross margin for the six companies outside the priority four, each with its cited line)
17build-connected-dataset-v5.pythe spine tables, Week 3 carbon, registers, steps 2 and 16D1, ten companies fund-level/connected-dataset-v3.csv and connected-dataset-lineage-v3.csv (a basis column: spine, register or documented)
18sweep-portfolio-v8.pythe whole portfolio folder of the data room, sweep-commissioning-v7.csv, sweep-judgments-v4.csv, sweep_logic_v6.py, sweep_logic_v8.pythe evidence sweep: sweep-findings-v8.csv (stated values), sweep-qualitative-v8.csv, sweep-review-queue-v8.csv (paragraphs that still need a person's reading: empty in the delivered run), sweep-qualitative-coverage-v8.csv. Stops if the sweep logic no longer matches its commissioning record
19build-exposure-model-v5.pysteps 4, 7 and 17, d2-source-inputs-v1.csv, d2-company-rubric-v1.csv, EIA prices, Week 3 exposure, registerD2, ten companies fund-level/exposure-lines-v5.csv and exposure-summary-v5.csv, late-exit-sensitivity-v2.csv (Patriot at its late exit), methodology/d2-inputs-v4.csv (every cited input, re-opened) and assumptions-register-v9.csv
20build-d4-d5-v4.pysteps 13, 14, 17 and 19, questionnaires, the Ridgeline planD4 and D5, ten companies d4-matrix-v4.csv, d4-summary-v4.csv, d5-initiatives-v5.csv, d5-ranking-v5.csv, d5-value-protected-v5.csv
21build-response-cost-v3.pystep 20, response-cost-inputs-v3.csv, response-cost-components-v1.csv, assumptions-register-additions-v7.csv, step 19 registerfund-level/d5-value-protected-v6.csv (cost to respond beside value protected, payback in days) and methodology/assumptions-register-v10.csv
22build-explorer-v6.pysteps 16 to 21 (the D1, D2, D4 and D5 files), explorer-inputs-v1.csv, explorer-owners-v5.csv, explorer-page-corrections-v1.csv, the page assets, the data room lines those files citeThe explorer fund-level/explorer-v6.html (one self-contained page, data embedded) and explorer-v6-lineage-v1.csv (every figure to its source file and column)
23build-revision-layer-v1.pysteps 16 to 22 (D1, D2, D4, D5 files and the simulated baseline), d2-source-inputs-v1.csv, revision-benchmarks-v1.csv, the DOE industrial assessment database and the EIA reference files, the data-room lines the estimates citeThe revision data layer fund-level/connected-dataset-v4.csv and connected-dataset-lineage-v4.csv, exposure-lines-v6.csv and exposure-summary-v6.csv, d4-matrix-v5.csv, d4-summary-v5.csv, d5-initiatives-v6.csv, d5-ranking-v6.csv, and methodology/revision-validation-v1.csv (the cross-check of each estimate)
24build-w4-final-v1.pystep 23, the explorer builder build-explorer-v6.py (its script is read for the page logic), explorer-owners-v6.csv, explorer-inputs-v1.csv, explorer-page-corrections-v1.csv, process-snapshot-v3.json, this note, revision-benchmarks-v1.csv, the page assetsThe delivery set in fund-level/delivery-v1/: 01-client-deck-v3.html, 02-explorer-v8.html, 03-exposure-model-v3.html, 04-materiality-and-initiatives-v5.html, 05-data-book-v2.html, 06-re-run-guide-v3.html, the portal kith-week4-portal-v2.html, delivery-lineage-v1.csv and a data/ folder of the data files

Steps 16 to 21 are the ten-company rebuild and are the current set for D1, D2, D4 and D5. Steps 1 to 15 still run because the rebuild reads some of their outputs (the initiatives and ranking tables of steps 13 and 14, the pooled event table in step 4, the exposure lines of step 7). Step 22 builds the explorer from the outputs of steps 16 to 21. Steps 23 and 24 are the current set for everything the reader at Northwood receives: they supersede the explorer of step 22 with 02-explorer-v8.html, built on the revision data layer.

Updating next quarter

  1. Data. Replace the data-room files the scripts read (the spine tables portco_kpis.csv and portfolio_companies.csv, the questionnaires, facility locations, supplier file, budgets) and the Week 3 carbon outputs. Each lineage file lists, for every figure, the file, line, period and grade it came from, so you can see what to swap. After swapping a file, update its pin: run shasum -a 256 <file> and put the result in inputs.sources in manifest-v1.json.
  1. Reference data. Download the newer EIA, FRED, FEMA and IAC files into reference-data/, re-pin each (shasum -a 256 <file>, then inputs.sources in manifest-v1.json), edit the typed constants for any FEMA file (see the next section), then update the matching rows in the assumptions register (value, source, source_date, period). External data carries its own vintage; it is not "stale" because it is older than the report date.
  1. Assumptions. For steps 1 to 15, change values in the register, not in the scripts; for steps 16 to 21 see "Updating the ten-company rebuild" below. Items that are judgment, not data, and need an owner's decision each time: the discount rate premium (A-02), hold years N (A-37), exit multiples (A-29), the severity and likelihood bands in the charter, and the Clearview idle-fuel estimate (g2-assumptions-v1.csv, grade C). The ranking is rerun at 8, 10 and 15 percent; a changed premium can change the order.
  1. Periods and dates. Set report_date with the setup script. Remaining hold N depends on it: after step 2 of the run order, compare the N rows in methodology/register-extraction-v5.csv with row A-37 of the register and edit A-37 to match; step 3 stops until they agree (within 0.015). Then work through the typed periods below.
  1. Pre-registration. The extended-data script fixes its random seed and sampling rules in a decision recorded before the data is generated. For a new run, record a new decision and change the two ids and the seed at the top of build-simulated-baseline-v1.py.
  1. Output names. Do not rename outputs. The scripts refuse to overwrite, and later scripts read earlier outputs by fixed name, so every run goes in its own new folder (setup step 2).
  1. Review the explorer's wording. The headline sentences, section titles and notes are typed in build-explorer-v6.py, not generated; the opening sentence and the totals are computed from the book selected. Read the page against the new numbers before it goes anywhere. The builder stops if the company totals it re-sums stop equalling the signed summary, if a cited line no longer carries its words, or if an owner row matches no backlog item.

Updating the ten-company rebuild (steps 16 to 21)

  1. Company inputs are two tables, not code. methodology/d2-source-inputs-v1.csv holds every input behind the exposure model: company, field, value, unit, the data-room file and line, and the text that must be on that line. For a new period, replace the rows; step 19 re-opens each line and stops if the text is not there. methodology/d2-company-rubric-v1.csv holds the judgments for the customer carbon-data line, per company: revenue type (earns today or hope to win), likelihood level (none, some, clear), whether the company can comply, and its share of revenue where it gives one.
  1. Rows to add or replace in that table: insurance_premium_YYYY and insurance_premium_2026_expected, energy_kwh and energy_therms, energy_cost_reported_usd, customer_requirement_revenue_exposed (with the share rows for Trident), own_event_property_damage_usd, the Keystone sessions_* and revenue rows, event_no_loss_statement, customer_statement_*, and late_exit_midpoint_date (the date used for the Patriot late-exit sensitivity).
  1. Still typed in code. Change these, then re-run in order: extract-six-values-v2.py lines 59 to 140 (acquisition dates of the six, and for every exit window and gross margin the file, line and expected text; the acquisition dates are asserted against the company master); build-d4-d5-v4.py from line 103 (the list of company-identified measures for the six: company, file, line, expected text) and the line numbers of Ridgeline's plan table (its figures are read from the file); and the client's severity and likelihood scales (SEV, PRI, BANDS) near the top of build-exposure-model-v5.py and build-d4-d5-v4.py. These are rules and citations, not company results, but they are in code.
  1. EIA price. Step 19 reads the U.S. total commercial price from the "U.S. Total" row of the reference file and states it in the line note and in A-42; it does not compare it with a typed number. It does check that the signed Week 3 energy shock for Sterling, Harvest Ridge and Meridian equals 20 percent of the energy cost in the energy_cost_reported_usd rows; if you replace the Week 3 outputs, update those rows.
  1. Physical-weather rules (A-46). The company's own reported event sets its impact. A loss a property owner bore is not the company's. A combined figure for two events is apportioned by closure days and the combined figure is kept as the upper bound. A company that reports no loss is shown as "reported: no loss", not as a banded zero. An unreported dollar loss is derived only from the company's own cited files (Keystone: cancelled sessions times revenue per session at gross profit), otherwise it is "not quantified". The halve-and-double frequency test runs for every line.
  1. Review queue and judgments. Step 18 leaves out paragraphs already judged in sweep-judgments-v4.csv. A paragraph the logic cannot place appears in sweep-review-queue-v8.csv with its proposed reading. A person reads it and records one row (para_id, answer, direction, value, note, judged_by, judged_date, basis) in a copy of the judgments file saved under a new version name; change the name on the line that reads it in sweep-portfolio-v8.py and re-run. The delivered queue is empty.
  1. If you change the sweep logic (sweep_logic_v6.py or sweep_logic_v8.py) step 18 stops, because the logic's fingerprints no longer match the commissioning record. Run test-sweep-v7.py (it also runs test-sweep-v1.py); first give its output file a new version name, because the scripts never overwrite, then change the commissioning file name in sweep-portfolio-v8.py.
  1. Cost to respond (A-44, A-45, response-cost-inputs-v3.csv, response-cost-components-v1.csv): the day rate and effort ranges are Kith's estimate, graded C. Replace them with your own rates and quotes.

Updating the explorer (step 22)

  1. Owners and first moves are rows, not code. explorer-owners-v5.csv has one row per backlog item: decision-holder, delivery lead, first move and Northwood contact, each with a basis tag (company file, adjacent role, or Kith proposal) and the file, line and words that support it. Replace the roles and moves with your own; the builder re-opens every cited line and stops if the words are not there, and it stops if an item has no row.
  1. Cited inputs. explorer-inputs-v1.csv carries the facts the explorer takes from the data room beyond the data layer (a company's term-loan balance and maturity). Replace the rows and the cited lines for a new period.
  1. Corrections. explorer-page-corrections-v1.csv lists two small corrections to text the signed data layer carries, each with its evidence line. Delete a row once the data layer is fixed; the builder stops if a correction no longer finds its text.
  1. Typed in the builder. The company display names, the risk short names, the default thresholds (large from 15 percent of EBITDA, likely from a 35 percent chance over the hold) and the page wording.

Updating the revision layer (steps 23 and 24)

  1. Benchmarks are rows, not code. revision-benchmarks-v1.csv holds each outside figure: the Marsh US casualty rate change, the RIMS cost of risk, the McKinsey disruption interval and loss share, the EPA fleet mpg, the EV kilowatt-hours per mile and price gap, the FHWA mileage, and the Patriot exit midpoint. Refresh each row from its source before a new run; the script stops if a row lacks a source, vintage or retrieval date.
  1. Computed from the pinned files. The water price is the median of the TN and KY plants in the industrial assessment database; the six measures' cost and payback ranges are quartiles of the matching recommendation codes in the same file; the insurance base is the median premium over revenue of the companies that state a 2025 premium (d2-source-inputs-v1.csv). None is typed.
  1. Judgments typed in the script. The supplier dependence factors (0.25 for people and facility services, 0.5 for equipment and consumables), the 20 percent water shock share, the choice of rubric level for the two customer lines, the first-half-2029 exit for Patriot, and the electric-vehicle count of eight. Change them in build-revision-layer-v1.py, then re-run steps 23 and 24; the cross-check file shows what each one affects.
  1. Review the wording. The delivery pages' headline sentences are typed in w4_content_v1.py, w4_deck_v1.py and w4_process_v1.py; the figures beside them are computed, and build-w4-final-v1.py stops if a claim in the wording stops being true of the data. Read each page against the new numbers before passing it on.
  1. Simulated values. The Harvest Ridge drought curtailment line rests in part on the simulated baseline by a specific exception; replace the simulated values with meter readings and the line becomes measured. The water price line beside it uses the reported volume only.

Checking a run

For the delivered run, handover-list-v17.csv lists a fingerprint (SHA-256) for every received file and every delivered output. A re-run from the same inputs and the same report_date rebuilds all 77 delivered outputs byte for byte; if one differs, an input has changed. To fingerprint a file: shasum -a 256 <file>.

Typed periods and counts to review

Some labels and counts are typed into the scripts. A new quarter needs a person to check each against the new data; the scripts fail loudly where a count or window no longer matches.

ScriptWhat is typedChange when
build-connected-dataset-v3.pycarbon period "CY2025 (as of 2025-12-31)" and finance period "LTM to 2026-01-31" (line 37); "FY2025 actual" labels (lines 145 to 146); "spine as of 2026-01-31" (line 152)the carbon year or the financial year-end moves; the script stops if the files lack the labels
extract-register-values-v5.py"FY2025 actual" labels (lines 161, 165); Sterling and Clearview exit windows in the note text (line 183)a company's exit plan or the year changes
build-fema-frequency-v1.pyFEMA window 2016-02-07 to 2026-02-06, 10 years (line 30); expected 141 rows and 96 distinct disasters (lines 48 to 49)you download a newer FEMA file: set the window, the row count and the disaster count to the new file
recompute-sterling-v1.pythe same FEMA window (line 59)with the FEMA window
build-exposure-model-v1.py"LTM to 2026-01-31" and "CY2025" in the line notes (lines 151, 159, 171, 203, 235, 267)with the financial year-end
build-g2-view-v1.pythe same EBITDA period in the page note (line 176)with the financial year-end
extract-six-values-v2.pythe six companies' acquisition dates (lines 59 to 60); "FY2025 actual (period exception)" labels for NovaCare and Pinnacle gross margin (lines 95, 103)a company is acquired or sold, or the financial year moves
build-connected-dataset-v5.pythe same carbon period "CY2025 (as of 2025-12-31)" and finance period "LTM to 2026-01-31" (line 42); "FY2025 actual" gross-margin labels for Meridian (lines 157 to 158); "spine as of 2026-01-31" (line 171)the carbon year or the financial year-end moves
build-exposure-model-v5.py"LTM to 2026-01-31" in the EBITDA period (line 152) and "CY2025" and "LTM to 2026-01-31" in the line periods (lines 227 to 381); the period text in the register rows A-38 to A-43 and A-46 to A-50 from line 466with the financial year-end and the carbon year
build-explorer-v6.pythe note "twelve months to January 31, 2026" and "Q4 2025" in the page notes; the default thresholds 15 and 35 percent; company display nameswith the financial year-end, the data-room date or Northwood's own scale
build-revision-layer-v1.pythe EBITDA period "LTM to 2026-01-31" in the line notes and the water-price window "FY2016 to 2026" in one line period (lines 155, 191, 213, 251, 280, 340, 349, 574)with the financial year-end and the vintage of the industrial assessment file
build-w4-final-v1.pythe data-room date "January 31, 2026" and "Q4 2025" in the explorer's page notes (lines 92, 268, 274)with the data-room date or the financial year-end
w4_content_v1.pythe data-room date "January 31, 2026" in the delivery pages' notes (lines 136, 240, 255, 319)with the data-room date or the financial year-end
w4_model_v1.pythe data-room date and "Q4 2025" in the forcing-date wording (lines 191, 235, 239)with the data-room date or the financial year-end
test-sweep-v7.pya judged date in a seeded test fixture (line 115)never: it is a test fixture, not data; change it only if you change the fixture

Known issues in the delivered run

  1. Register note A-13 is wrong; the value is right. The note on row A-13 of assumptions-register-v3.csv says the questionnaire-implied Harvest Ridge electricity price is about 15.8 cents per kWh. It should read about 13.5 cents per kWh (the A-19 electricity cost divided by the questionnaire kWh), nearly double the EIA industrial average of 7.13. No model value used 15.8: D5 computes the price from A-19 and the kWh, and the stored EIA ratio of 0.5274 matches 13.5. The register stays as delivered and this entry is the correction.
  1. No IAC cross-check was run at this stage; it was done later in the revision layer, where 2 of 2 company paybacks sit inside the DOE lighting range (solar and HVAC remain untested). The DOE IAC database is in reference-data/ and pinned, and the charter (rule R8) makes it a cross-check only. No cross-check rows were produced: all five quantified measures come from company questionnaires, so none is benchmarked against IAC. Adding that benchmark is a new step, not a fix to an existing one.
  1. The D4 matrix runs through the first-pass chain (steps 10 to 12), as described under the run order.
  1. The cost to respond is an estimate with secondary sourcing. The day rate rests on two benchmark pages whose addresses and check dates are in d5-response-cost-method-v2.md; one could not be fetched directly (its figures were read from a search excerpt), two statements in version 1 that the pages did not carry were removed, and there is no independent project-fee check (A-44, grade C).
  1. The sweep is the evidence record only for the inputs it has rules for. Initiative costs and savings, insurance history, energy units and event figures come from cited questionnaire and plan lines that the exposure model re-opens and asserts (d2-source-inputs-v1.csv) and that an independent check re-opens again; no check ties them to the sweep's findings (a logged decision).
  1. The held-out check supports the review agent and led to a fix in the sweep script. Twelve paragraphs the sweep had resolved were read blind by the review agent and compared with labels (sweep-holdout-score-v1.csv): the agent matched 11 of 12 answers, the script 8 of 12. The script read a leading "No" as "no" where the codebook reads an informal or sized lesser fact as "partial" (items 2, 4 and 7). Sweep logic v8 now queues such answers; twelve paragraphs were queued (sweep-qualexist-queue-v1.csv) and all are judged in sweep-judgments-v4.csv (nine partial, three no). The labels for the twelve held-out items are the builder's proposals approved by the owner and the rule was written after the three misses were seen, so the held-out batch is now in-sample; the nine other paragraphs the rule queued, of which six came back partial, are the out-of-sample evidence. A check re-derives from the qualitative rows that no unjudged "no" has a sized lesser fact after its first sentence.
  1. One workbook in Meridian's folder states a 35.0 percent gross margin. It is headed with a different fund's name and shows exactly 35.0 percent in every period; it is excluded and 30.0 percent (the company's own FY2025 budget file) is used (a logged decision).
  1. Derived energy is a low case. Gas is priced at the industrial price; Sterling's reported gas cost is about five times that figure and unexplained (register A-24), so the derived energy lines for the four companies that report units may be understated; each carries a gas flag and none is large enough to move a severity band.
  1. The explorer's exit year is a model, not a plan. It is the report date plus the remaining hold, which is a judged midpoint or a one-year floor for most companies; the page says so and marks floored holds as earliest modelled. Check it against Northwood's own exit plans.
  1. The owners and first moves are Kith proposals. Each is tagged with its basis. Where a company's file names the role for the topic the tag says so; the first moves are always proposals. Two text defects in the signed data layer (a questionnaire line cite and a truncated clause) are corrected on the page through the corrections file and the data layer still carries them.
  1. d2-inputs-v4.csv is a byte copy of d2-source-inputs-v1.csv, written when step 19 re-opened and asserted every cited line; it adds no evidence beyond the input table.

What a re-run does not do

It does not make the judgments above for you, and it does not read documents it was not pointed at. A company with no stated cost-and-saving measure shows "not quantified", never a filled-in number. Figures marked grade C or PROXY stay marked until a company supplies its own number.

02Back-end note: how it was built

Everything in the build is one of six things, and each number has a trail

This note is for whoever audits or maintains the work. It shows what each part of the process does and how the parts link, without opening a single script. Counts are the record as at 11 October 2026, before the final (G7) freeze.

THE TRAIL FOR EVERY NUMBER: rule, script, check, review, decision, record1. Rules16 rules in the chartersay what must be true2. Scripts24 run-order stepsmake every output file3. Checks65 gate checks, 33 seeded teststest the rules; can stop the build4. Reviewers6 reviewer rolesread-only agents; raise findings5. Decisions180 logged, datedthe owner's words, as given6. Record289 chained entriesappend-only; an edit breaks the chain

Rules

The charter's 16 rules say what must be true: units, scenarios, bands, missing data, ranking, simulated data and the customer data requirement. Nothing on a page departs from them unless a dated decision says so.

Scripts

24 run-order steps (87 script versions are kept) build every file from the data room and the reference data. They stop with a plain message if an input, header or period no longer matches, and never overwrite.

Checks

Gate checks re-derive figures from source and fail on a defect; each is proved with seeded failures before it is trusted (33 test files). A check that passes is stronger evidence than a reviewer's opinion.

Reviewers

Reviewer agents, each read-only and given one question, work from the same packet: Evidence review, Audience read, Presentation review, Method review, Process review. They raise findings; they cannot change a file.

Decisions

180 decisions are logged with the date and channel, and the owner's words wherever the owner decided. A builder's proposal is marked proposed until the owner ratifies it.

The record

289 entries are linked to one another, so an earlier entry cannot be altered without breaking the chain. 259 journal entries say what was done, why, and what was left open.

03Back-end note: gates and review

8 gates, each closed only when its checks pass and the owner signs

Each gate froze the files it covered. Where a check failed or a reviewer raised a must-fix finding, the work went back, was versioned and re-run. Dates are when each gate was frozen; findings are those raised at that gate.

GateClosedChecksReviewersMust-fix raisedShould-fix raisedFiles frozen
G0 Foundations2026-10-0910Process review224378
G1 Data and models2026-10-0925Method review, Evidence review49100
G2 Outputs2026-10-0920Evidence review, Presentation review, Audience read526104
G3 Close2026-10-0914Process review86103
G4 Rebuild2026-10-1022Evidence review, Method review, Process review924250
G5 Explorer2026-10-1110Evidence review, Audience read317176
G6 Handover2026-10-1112Process review36175
G7 Revision and final deliverysigned and frozen 11 October 202613Method review, Evidence review, Audience read, Process review38 findings, all fixed52

Gate G7 covers the revision of this delivery set after G6: the decisions logged after G6, a delivery check with seeded failures, and a reviewer round on method, evidence, audience and process before it was frozen. What was not reviewed: there was no Presentation-lane read of the final pages; the counts in this note are the record at the date shown and are not recomputed; and fixes made after the G7 round were re-run through the delivery check and the re-run test, not re-reviewed.

04Back-end note: what was caught

223 findings were raised on this build; 205 are fixed and every one has a recorded reason

Reviewers worked from the same packet, read-only, with one question each. These are examples of what they caught, in plain terms.

223
findings raised, by reviewers and by the builder
205
fixed, with a recorded reason
8
accepted or deferred, with a reason
10
still open (process items)

G0 · Process review must-fix

Caught: A five-year forward hold was applied to every company, but Sterling is already 5.7 years in and Meridian exits in the first quarter of 2027.

Done: Hold years are now set per company from its documented exit plan.

G1 · Evidence review must-fix

Caught: Clearview's weather events were counted over one year when the company lists three, which overstated how likely they are.

Done: The window is now the three years the company reports.

G4 · Evidence review must-fix

Caught: Ridgeline's growth-plan investment was stated in its own plan, but the model said it was not stated.

Done: Quantified from the plan's own investment table and shown as a company plan, not capitalised.

G4 · Method review must-fix

Caught: Ridgeline's weather impact had been taken as a property owner's cost, not Ridgeline's own.

Done: Carried as Ridgeline's own small event; the owner-borne losses are shown as context.

G4 · Method review must-fix

Caught: Pinnacle's largest event had combined two storms, lifting its severity band.

Done: Apportioned by closure days; the combined figure is shown only as an upper bound.

G5 · Audience read must-fix

Caught: Exit dates were labelled 'planned exit' though they are Kith's model, not a Northwood plan.

Done: Relabelled 'modelled exit', with floors marked as the earliest modelled date.

G5 · Audience read must-fix

Caught: Paybacks in days and exit values read as sized benefits though they rest on a what-if.

Done: Labelled 'if the half-lost case happens' and 'what-if, not likelihood-weighted'.

G6 · Process review must-fix

Caught: The decision list shipped to the client was missing most decisions its own files cite.

Done: Rebuilt from the record so every cited decision is present.

Who raised what
  • Process review: 100
  • Evidence review: 36
  • Audience read: 33
  • Method review: 30
  • Builder: 12
  • Presentation review: 11
  • Builder (held-out check): 1