Financial Model Updater

Step 5: Rebase the estimates

Reported to adjusted. The reported figures and the basis the model runs on are different metrics. Bridge them, one line per adjustment, and say which basis the model uses:

Reported operating income | + stock compensation | + restructuring | + acquisition amortisation | = adjusted operating income

When the company changes its own definition of the adjusted measure, that change is a finding: name it, show both definitions on the affected period, and treat the periods either side as not comparable until the bridge is restated.

Estimate change by driver. Show the walk, not just the new numbers. Each step is a driver, and the steps close on the new estimate:

Prior FY26 EPS | volume | price/mix | gross margin | operating expense | share count | tax | = new FY26 EPS

Then compare the revised estimates against consensus, with the consensus vintage stated.

Done when the bridge's closing figure equals the model's new estimate cell for every period shown, and every adjustment line and driver step names its source or its reason.

Step 6: Restate the valuation as a delta

A new price target on its own tells the reader nothing about what moved it. Report the change and its drivers:

Then decompose the change into estimate revision, discount rate or multiple change, capital structure, and the roll forward of one period, so the components sum to the total change.

The stance the number implies, and the vocabulary available for it, come from .agents/skills/research-conventions/references/judgment.md.

Done when the decomposition sums to the change in fair value, and the spot price and implied return both carry the same as-of.

Step 7: Recalculate, audit, and say which

python .agents/skills/xlsx/scripts/recalc.py <file> 60      # until status is "success"
python .agents/skills/xlsx/scripts/audit.py  <file> --strict # fix every fail

Then reopen with data_only=True and read the Checks roll-up, if the workbook carries one. A cached value shows a passing check on a broken model, which is why the delivery message states plainly whether the file was recalculated or whether displayed values are the ones that were already there.

Done when recalc.py reports success, audit.py --strict reports no fail, the Checks roll-up reads OK, and the delivery message says the workbook was recalculated.

Step 8: Deliver

Everything under {task}/:

• the updated workbook, carrying the Change Log sheet,
• the estimate-change summary: what changed, why, whether it is thesis-changing or noise,
• the valuation delta from Step 6.

The message that delivers them carries three things: the readiness posture, read from the table in .agents/skills/research-conventions/SKILL.md against the state the update leaves the model in, where an assumption row means a load-bearing input rests on a chosen number and a requires rebuild row means the model cannot represent the change at all; the rows that were not written and what they would take; and the monitoring items the update creates, in the table shape .agents/skills/research-conventions/references/judgment.md gives.

Done when the posture is stated with the specific row responsible for it, and every unwritten mapping row appears in the message.

When the model cannot be edited safely

Some workbooks cannot take a write without breaking. The tells: the target cells hold formulas rather than inputs, the file drives external links or macros, it carries pivot caches or objects openpyxl drops on save, or it is the user's system of record and they have not asked for it to be changed.

Deliver a control pack instead: the original untouched, plus a companion workbook whose sheets say what they hold, New Data, Bridge, Implied Impact, Change Log. The pack stands alone, since a cross-workbook link cannot be resolved: values taken from the original are typed in with a provenance comment naming the sheet and cell they came from, and everything computed inside the pack is a live formula under the conventions in .agents/skills/xlsx/SKILL.md.

Done when the original's hash is unchanged, and every sheet in the pack names the cells of the original it corresponds to.

When to rebuild instead

An update assumes the model's shape still fits the company. These changes break that assumption, and updating through them produces a model that ties and misstates:

• the company re-cut its segments, or changed what a segment contains,
• the accounting basis changed: a new standard adopted, a revenue-recognition change, a reporting-currency change,
• a merger, divestiture or spin changed the entity the model describes,
• the driver the update needs does not exist in the model at all.

Say so rather than forcing the write, and route it: .agents/skills/3-statements/SKILL.md for the operating model, .agents/skills/dcf-model/SKILL.md for the valuation. The mapping table is what carries into the rebuild, since it already says what the new shape has to hold.