Budget versus actual

Working state: notes/HANDOFF.md carries the current branch, the open decisions and what is established versus assumed. CLAUDE.md carries the rules.

Analysis, August 2026. Published here as part of the source archive, and deliberately not part of the app's projection — see the interface rule below for why those are different things.

Every section below is written twice: In plain terms for anyone, and The evidence for anyone who wants to check it. The plain version never states anything the evidence does not support — it is the same finding in fewer syllables, not a softer one.

Why this is a separate document from the app

The app answers: how big is the gap, and what would close it? It is built from the budgets the town published, and someone can disagree with every assumption in it and still accept the arithmetic.

This document answers something else: did the money the town budgeted match the money it actually spent?

That second question is a much heavier thing to say out loud. It needs much better evidence, and if any part of it turns out to be wrong, it hands a critic a free way to throw out the first question along with it. So the two stay apart, and the traffic between them runs one way:

Something learned here can change a number the app assumes. The app never repeats an accusation from here.

Each finding is tagged CROSSES (it changes the app) or STAYS (it does not leave this page).

What we are working from, and what we are missing

Two sources, and they do not always agree.

The FY27 workbook (sources/budget-workbooks/fy27-proposals.xlsx) gives 351 line items with both halves for FY25, and eight months of FY26. That is what the first version of this document worked from.

The mirrored budget documents (sources/district-budget/) go back to FY17 and print actual columns beside budget columns, with each document stating its own column kinds. scripts/extract_budget_history.py reads them, taking only the column a document itself labels, and the series are published at /data/total-salaries-history.csv and /data/total-expenses-history.csv.

Where the two overlap they agree closely: FY25 actuals are identical to the dollar, and FY23 and FY24 differ by $10,000 and $5,000 on roughly $15.6M of salaries — 0.06% and 0.03%.

scripts/extract_line_history.py does the same for every line the documents print, not just the totals — 24,337 readings across 32 documents, normalised to 397 distinct lines, published at /data/line-history.csv. That is what makes it possible to ask whether the same lines miss every year, which is the question §2 and §3 turn on.

What that series can and cannot support. Each line's budget and actual are read from the same row of the same document, so a per-line comparison is sound. The lines do NOT sum back to the district totals — between 0.11% and 2.68% out, worst in FY19 and closest in FY23, measured as the sum of the settled lines against the documents' own TOTAL rows. So this file never apportions the total variance between lines. It asks which lines miss and how often, and that is all it asks.

The limit that bounds everything below

The documents disagree with themselves, by more than the effect this document is trying to measure.

Four times in the corpus, a single document prints the same total for the same year twice and gives two different numbers:

In fy24-approved-budget.txt the two FY22 expense totals sit 257 lines apart, one after the salary tables and one after the expense tables, in the same file.

The budget-versus-actual variances measured below run between 0.1% and 0.5%. The documents' disagreement with themselves runs to 1.5%. So a year where a document contradicts itself cannot be used, and the three years that survive are the three where nothing does.

This is not an accusation. Summary blocks printed at two points in a long document, one of them updated and one not, is the most ordinary thing in the world. It is a limit on what can be concluded, and it is the reason this document is careful about small percentages.

FY21 is not a usable year

Every document that reports FY21 gives it an "actual" column, and in 116 of 117 lines that column is identical to the budget, to the dollar — 99%. In ordinary years the figure is 10% to 21%, and those are lines that genuinely do not move: fixed stipends, flat contracts.

A district does not spend its budget exactly on 117 lines. That column is the budget, printed under an "actual" heading — most likely because the books were not closed when the document went out. FY21 was the first full COVID year, which is the obvious candidate for why, though nothing published says so.

So FY21 is excluded from every comparison below. It had been one of three years supporting §1, which is the sort of thing that only shows up when you go line by line.

Two words used below

• Appropriation — the amount Town Meeting votes. Not the same as what gets spent.
• Encumbered — money promised under a signed contract but not yet paid out.

And one thing we DO have, which an earlier version of this document said we did not. The Town Accountant's own year-to-date budget report for FY26 through 31 March, obtained by records request. It is the ledger rather than a budget document: original appropriation, transfers and adjustments, revised budget, expended, encumbrances, for all 67 general fund departments. Extracted to /data/town-ledger-fy26-q3.csv, which reconciles to the report's own GRAND TOTAL before it will write.

It matters because this document repeatedly says money moved between lines mid-year is invisible from here. At department level it is not. In FY26, 28 of the 67 departments had money moved into or out of them: $489,411 in, $148,177 out, of which $76,394 went to the school department — while school non-recurring expenses gave up $4,223. Transfers are real, ordinary and routine, and now they can be pointed at.

What it does not do is settle which school LINES moved: the whole school department is a single row. And it is one snapshot of one year.

What we still do not have: the individual payments, the school department's own line-level ledger, and any year but FY26. Most of what follows still compares columns in budget documents. That is enough to see that a number moved. It is not enough to say why.

Nothing in this document should be described as anything stronger than a gap that has not been explained yet.

1. "The town spends less than it votes, every year" — that is not what the record shows

In plain terms

The earlier version of this document said the town votes more than it spends, every year, and rested it on one year. FY25 was the only year where both halves existed.

The archive changed that. The district's own budget documents go back to FY17 and print actual columns beside budget columns, so the comparison can now be made for several years. It does not hold.

The evidence

Two years where the district's documents give both figures, no document contradicts itself about either, and the year is a real one — see the note on FY21 above:

One lands within a third of a percent and the other is over. Against FY25's −3.01% they are a different order of magnitude, and they do not point the same way. Read the line-by-line totals in §2 alongside them: five years there, never more than 0.9% off.

The FY25 budget is stated two ways, and one of them is explicable

FY25 is the year the original claim rested on: a shortfall of $761,249 against a budget of $25,321,760.

There are two published totals for FY25 salaries in the workbook and a third in the FY26 approved budget document, and a line-by-line comparison explains the salary side entirely. Of 117 FY25 lines both sources state, 115 are identical. Two are not:

Salary Reserve is $347,338 — exactly the workbook's internal salary difference. Its function group is literally TOTAL SALARIES; it is a contingency line that sits in the totals block, and one of the workbook's two totals counts it while the other does not. That is not two sources disagreeing. It is one total struck before a contingency and one struck after, which is ordinary, and the earlier version of this section was wrong to present it as a contradiction.

The expenses are a different matter and are not explained. The workbook says $8,165,299 and the FY26 document $7,695,034 — $470,265 apart. Matching the expense lines between the two sources by name, 142 line up and only three differ, by about $1,000 between them. The lines agree and the totals do not, and nothing published resolves it.

So the range on FY25's budget is narrower than this document previously claimed, and it does not cross zero:

FY25 came in under budget by somewhere between $290,984 and $793,130. The original $761,249 sits inside that range at the top of it. What cannot be said is which end is right, and the $470,265 of it that turns on the expense total is unresolved.

The town has its own number for this, and it is $603,885.97

None of the arithmetic above was necessary to establish that FY25 came in under. The district said so itself, in public, and the figure is in its own minutes.

School Committee, 3 September 2025. The Chair read a statement about a memo from the Superintendent to himself, the Town Manager and the Chair of the Finance Committee "regarding a budget surplus that was recently discovered in the amount of $582,115.44", and described "the unhappy discovery of this surplus".

School Committee, 17 September 2025. "the surplus number has gone up to $603,885.97, the change is due to closing out purchase orders from FY25."

The minutes record what that meant in the room: a member describing "an unknown surplus of money that was not spent and instead was given back to the town", another that "we had drastic reductions in services to students. We really could have used those funds to keep some of the positions."

$603,885.97 sits inside the range this analysis derived — $290,984 to $793,130 — and nearer the top of it. It is also the better figure to quote, for a reason that has nothing to do with arithmetic: it is the town's own, arrived at by closing the books rather than by subtracting two columns, and it is what the people who voted the budget were told.

What this project should say about FY25 is that figure. Not $761,249, which is one subtraction among several, and not a range, which invites an argument the town has already settled.

One thing named in those minutes that we cannot check

The 17 September minutes list, among the issues discovered, "double booking of the para salaries". That phrase appears once in 1,383 meeting documents and is never explained, and it points at the exact line the in-district special education rate is built on.

So it was tested. It cannot move that rate. The para escalator is a compound rate anchored on FY18 and FY27, and FY25 is an interior point: dropping FY25 entirely leaves the rate at 12.78%, and so does assuming FY25 was overstated by 15%. The straight-line fit moves from R² 0.89 to 0.85. The rate is what it was.

What we cannot say is what was double-booked, in which year, or by how much. The phrase is evidence that the district found something and said so. It is not a quantity, and it is not treated as one here.

What this does not affect

Nothing the app computes. Every projection starts from the FY27 adopted column, and the two lines that differ — a salary contingency and a dues line — are not special education and appear in none of the FY25 figures the model reads. Checked line by line rather than assumed.

What we cannot tell

Which figure is right. Both are the district's own. The likeliest ordinary explanation is that they are drawn at different moments — an appropriation as voted against one revised during the year — but nothing published says so, and we are not going to guess at it.

What this means for the app

Nothing, and that is worth stating. Every projection on the site is computed from budget columns only, and this is a question about the distance between two columns rather than about how fast a line grows. It does mean the earlier "three-quarters of a million a year" framing should not be repeated. It was one year, measured against one of two sources that disagree.

2. The total is quiet. Everything inside it is not

In plain terms

This was the strongest finding in the first version and it survives — the difference is that it rested on one year and now rests on five.

The district lands within about one percent of its budget every year. Underneath that, individual lines miss by tens and hundreds of percent, in both directions, and cancel out.

The evidence

Five years, never more than 0.9% off. Now the lines inside those totals:

What is new, and it changes the reading

98 lines have four or more usable years. Seven of them miss the same way every year, and all seven are small — the largest averages $20,754 on a $20 million budget.

So the misses are noise, not padding. Nobody is quietly over-providing a line year after year. Lines overshoot and undershoot and the total comes out flat because there are a hundred of them and they are independent.

That is a better answer than the first version gave, and a less suspicious one.

What we cannot tell

Districts move money between lines during the year with proper approvals. A line that overspent may have been topped up legitimately, and we cannot see those votes.

2b. The whole budget, every group, every year

An earlier version of this section ranked groups by pooled dollar variance, looked at the top six and called it a sweep. That is the method this very document shows does not work: a group whose overspends and underspends cancel reads as quiet when nothing inside it is, and athletics — which §4 is entirely about — never appeared. scripts/analyze_variance.py replaces it and reports every group, every year, four ways.

Coverage first

576 usable line-years, 143 distinct lines, FY18 to FY25 — covering 89% to 96% of the district's budget in FY18–FY23. FY25 is new and it is thinner: 43 lines against 106 in FY23, because only one document so far restates FY25 line by line. Its −2.61% is measured on $7.7M, not on the whole budget, and should be read as a first reading rather than as the year's verdict.

FY23 read as 82% until September 2026, and the thinness was ours: P.S. Teachers/Regular (1-2) and E.S. Teachers/Regular (3-5) carry a grade range in the line's name, the reader split each row at that opening bracket, and 1 and 2 went in at the front of the figures with every real one shifted along behind them. FY23 and FY24 for both lines were published as 1, 2, 3 and 5. About $2.5M of FY23 salary fell out of the measured base as a result. It is in.

3,158 line-years were excluded and each exclusion has a stated reason: 2,046 have only one of the two figures, 750 are under $10,000, 301 are FY21, 51 are cells two documents disagree about, 8 are in FY16 and FY17 — four usable lines each, which is not a measurement of a budget and is excluded by a stated floor of twenty — and 2 are parse artifacts. 212 lines in the FY27 workbook are never measured here at all — mostly lines the older documents do not carry.

The four measures, because one hides what another shows

• net — pooled variance, what a treasurer feels
• gross — over and under added as absolutes, what a budget officer feels
• churn — gross over net. 1.0 means every year misses the same way; high means the group only looks calm because its years offset
• spread — worst year to best

The whole budget, by year

Never more than one percent off in either direction, and one year over.

Salaries come in under. Everything else comes in over.

(The two do not sum to 576: 35 line-years are on lines the FY27 workbook no longer carries, so they have no section to be filed under.)

The shape of a district provisioning payroll for a roster it does not always fill, and provisioning everything else about right. Both are within one percent, and they very nearly cancel — which is why the whole-budget figures above are so quiet.

Almost nothing misses the same way twice

Of every group with four or more measured years, exactly one is over in every year (custodial supplies, +$83,408) and exactly one is under in every year (general education paraprofessionals, −$49,298). Both are small.

At line level it is the same: 7 lines out of 141 miss the same way every year, the largest averaging $20,754.

There is no systematic padding in this budget. That is the single most important finding in this section and it is a negative one.

What there is instead is drift

27 groups moved materially between their first two measured years and their last two:

The clearest case: a line that stopped being maintained, and was then fixed

COMPUTERS — Purchase & Lease:

The budget sat at $39,000 while spending ran $165,107 and $186,722. Then the FY26 budget is $243,450. Equipment leases did the same thing — $30,420 budgeted against $96,907 spent, then $85,000 in FY26.

So this is not a district that cannot budget. It is a line that drifted for four years and was then corrected, which is the same shape as the athletics coaching line in §4. Both are worth saying because a reader looking only at the middle years would conclude something else.

By function group, the largest net movers

And the largest movers, line by line

Health insurance is the one that matters to the gap. Only +1%, +0%, −1%, +3%, +7% — but on the largest non-salary line in the budget that is $293,023, and it is drifting upward. It is the only overspend attached to a line big enough to move the projection, and the model already escalates it at 9%.

And the groups that cancel themselves out

Ranked by churn — these look calm in the net and are not:

Athletics is in that table, and that is why the first attempt missed it. Its net is −$6,787 on athletic expenses and −$69,968 on athletic salaries — unremarkable numbers hiding years that run from −19% to +29%. §4 is about the year those swings were largest.

3. One cost cannot be predicted — and it misses in BOTH directions

In plain terms

This is still the most important finding here, and half of it was wrong.

When a child's plan requires a school the district cannot provide, the town pays another school. It does not set the price, does not choose how many children need it, and cannot say no.

The first version said the district pads what it can predict and under-provides what it cannot. The first half roughly holds. The second does not. Out-of-district tuition is not systematically under-budgeted — it is unpredictable, and it misses badly in whichever direction it feels like.

The evidence

Out-of-district tuition, budgeted against spent:

Three years over, three under, and a range from +27.0% to −51.5%. Twice it cost a quarter more than budgeted; twice it cost half of what was set aside. FY25, newly measurable, lands −37.1% — the third consecutive year it came in far under, which is the first thing in this series that looks like a pattern rather than noise. Three years is still three years, and rule 6 says so.

Special education staffing over the same years:

Five of six under, and never by more than 5.9% — modest, and much steadier than tuition. FY25 rests on a single line, $445,328 of it, because only one document restates FY25 at line level so far. It is consistent with the others and it is not independent evidence of anything.

So the shape is real but the language has to change: this is a line nobody can forecast, not a line anybody is short-changing.

What this already changed in the app

The special education work of 28 August reached the same conclusion from a completely different direction — eleven budgets of the tuition line with a straight-line fit of R² 0.10 and no trend to measure — and the model now holds the line flat and publishes a range of priced scenarios instead of a rate. These two methods disagree about nothing, which is the most reassuring thing in this document.

4. Some lines were set at zero and spent against anyway

In plain terms

FY25 athletics spent $281,665 more than it was budgeted — more than double. That looks bad until you look at where it came from.

$155,614 of it was coaching pay on a line budgeted at zero dollars — while coaches were, obviously, being paid all year. It was not an overspend so much as a cost that was not written down.

The following year it was fixed: coaching was budgeted at $159,444, close to what it had actually been costing. Athletic transportation got the same treatment after overspending — $40,000 the year it cost $87,822, then $127,550 the year after.

So the honest read is: a gap in the budget paperwork, found and corrected. Worth saying because if you only looked at FY25 it would look like something was being hidden, and it was not.

The evidence

Total FY25 athletics overspend $281,665, of which $155,614 was on lines budgeted at $0.

Why we checked, and the good news

The app's whole argument about athletics fees rests on what athletics costs. If those figures were built on broken numbers, the fee analysis would collapse.

They are not. The app uses the FY27 budget — two years after the correction. We checked this specifically because it could have gone the other way.

STAYS.

5. A correction to our own argument — and then a correction to the correction

In plain terms

The first version of this section said special education is budgeted above what gets spent, and used it to argue that the district manages the line carefully. It rested on FY25, where the gap was −12.4%.

Across five years it is not that clean.

The evidence

Special education, staffing and tuition together:

Two years over, four under, and the three largest misses are the three most recent. FY25's −27.5% is measured on a much smaller base than the years above it — $1.6M against $4.0M — so it belongs in the direction of travel and not in the magnitude. This is not a line consistently provisioned above what it costs; it is a line that moves around, mostly because the tuition half of it does.

And a figure that has to be withdrawn

The first version supported this with "13% a year while the rest of the budget grew 3.4%". That comparison ran from an FY23 actual to an FY26 budget — the exact error rule 1 exists to prevent, and it is corrected at the head of analyses/sped-and-funds.md. Budget to budget the figure is nothing like 13%.

The sentence it was supporting can stand on its own without it: special education staffing comes in close to budget, and the app should never describe those lines as overrunning. Only the tuition line does that, and it does the opposite just as often.

What we cannot tell

Whether the FY22 and FY23 underspends are unfilled positions, delayed placements, or money moved elsewhere with approval. The three look identical from here.

6. The one year the para budget did not match what was spent

The FY25 minutes name "double booking of the para salaries" as one of the things the district found. It is never explained anywhere else. But it is testable, because the para lines have a budget and an actual for six years:

For five years the para budget and the para spending track within about three percent, and the actual usually comes in slightly above. FY25 is the only year that breaks it, and it breaks it by $159,649 — a quarter of the $603,886 surplus the district reported for exactly that year.

That is what a double-booked line looks like from outside: money budgeted twice, spent once. It does not prove the district's account, but it is consistent with it in the right year and in no other, which is about as much as an outside check can offer.

What it does not do is move any rate. The escalator on this line is anchored on FY18 and FY27, so FY25 is an interior point — removing it entirely changes nothing. And the same growth appears in actual spending as in budgets: 9.07% a year against 10.89%.

7. The athletics budget is net of fees, and the app does not know it

Found while sweeping, and it is the one finding here that changes something published.

In the general fund, two athletics lines are budgeted at nothing from FY26 on:

In the athletics revolving fund, which is funded entirely by user fees, the same two costs appear by vendor: ArbiterSports $59,400 for officials and scheduling, and Prime Time Sports $25,421 for uniforms — within $113,602 of purchase of service, on top of $30,514 of salaries for four staff. Total fee-funded athletics spending: $146,911.

The costs left one pot and arrived in the other. That is rule 11 in its plainest form: the general fund line is net of what fees pay for, and nothing marks it as net.

What that does to the app. PROGRAM_TOTAL_ADOPTED is $217,908 and the fee model asks what fee would cover $345,458 — while today's fee already covers $146,911 that is not inside either figure. Gross athletics is about $364,819 and fees already pay roughly 40% of it. The published $960 self-funding fee, the $358,380 revenue peak, "fees cover 54%" and conclusion 5 are all built on a base that excludes what the fee already buys.

model/athletics.py states the opposite in a comment — FEE_REVENUE_IN_BUDGET = False, because "fees flow through revolving accounts, so the program costs shown are GROSS". Fees flowing through a revolving account is precisely how the general fund line ends up net.

Not corrected yet, because correcting it means rebuilding the fee curve on a gross base and $146,911 is a floor rather than an established total.

And athletic transportation, which is not the same story

Athletic transportation sits in the general fund under 3510, and its actuals are recorded there like any other line — $39,880, $40,000, $87,822, then $127,550 budgeted for FY26 and cut to nothing in the FY27 adopted budget.

It is not fee-funded: none of the revolving fund's $109,503 of named vendors is a bus company. But two things about it are unresolved. FY24's actual is exactly its $40,000 budget, which charter invoices do not do. And the district's own comment on that row reads "Actuals in munis are tracking well below FY26 budget, can FY27 be reduced? Yes, level funded" — with $47,847 spent and $13,169 encumbered at three quarters through, against $127,550. The app treats $127,550 as the cost of putting the buses back.

What would change any of this

1. Which FY25 budget figure is right. $25,321,760 or $24,883,376. One question, and $438,384 of a $761,249 headline turns on it.
2. The town's accounting records. Everything here compares columns in budget documents. It cannot see money moved between lines during the year — the normal, approved way a district covers an overspend, and the ordinary explanation for several findings above.
3. FY26 year-end. Nothing here reaches it.
4. Why FY21's actual column is its budget. Probably books not closed in a COVID year. Probably. It is currently an inference from a pattern in the numbers.

Still to do

• Findings 2, 3 and 5 have been redone across five years. Finding 4 has not — it is specific to FY25 athletics, and it describes a gap that was found and corrected, so it is the one finding that does not need a longer series to mean what it says.
• Reconcile the line-level series to the district totals. It is 0.1% to 1.6% out, which is fine for the per-line questions asked here and not fine for apportioning a total.
• Extend the series past FY23. Coverage thins because fewer documents report recent actuals; the FY27 workbook covers FY23 to FY25 and has not been merged in line by line.
• Ask how coaching pay was handled in FY25 against a $0 line (finding 4).
• Check the athletics general-fund lines against the athletics revolving fund.