Where the Burden Moves: Agentic Decision-Support and the Distribution of Municipal Tax Burden
📋 Cite this paper
SomaSoft Research (prepared by Claude Code for the AURI project). (2026-08-20). "Where the Burden Moves: Agentic Decision-Support and the Distribution of Municipal Tax Burden". SOMAsoft Research. Available at https://somasoft.ai/papers/where-the-burden-moves. Licensed under SAGL-1.0.
Where the Burden Moves
Agentic decision-support and the distribution of municipal tax burden
A de-identified case study of one mid-size Illinois municipality, FY2026
De-identification note
This paper studies the published FY2026 budget of a mid-size Illinois municipality of roughly 78,000 residents. The municipality is not named, and neither are its officials. All fiscal figures are drawn from its adopted budget and from public reporting.
County-level institutions are described, because the finding turns on them and because the positions attributed to them are their own published characterisations of county policy — not claims about the municipality. Readers familiar with Illinois local government will recognise the county structure; the point of the de-identification is that this paper is not about a particular city and is not addressed to it. It is a method paper that happens to need a real balance sheet.
Framing rule
This work carries a binding constraint, written into the originating proposal before this paper was commissioned:
"This is capacity augmentation for under-resourced public staff. It is never headcount reduction, and never tax-base reduction."
The brief for this paper — reduce and evenly distribute the tax burden — is half inside that rule and half outside it.
"Evenly distribute" is squarely inside: it is the incidence analysis the rule's own "honest form" section explicitly calls for. "Reduce" is the flagged half, and this paper does not treat tax minimisation as an objective function, because doing so launders a value choice as a calculation. Efficiency for whom, cutting what services?
What it does instead: it separates cost of service and revenue yield — where software can legitimately move money without cutting a service or a job — from the levy, which is a political decision belonging to the council and the public process. Section 06 is candid that the largest driver of the studied increase is not addressable by software at all.
01 — The fiscal picture
| FY2026 adopted budget | ~$404M |
| Revenues (excl. transfers & fund balance) | $338.6M |
| Expenses (excl. pension transfers) | $342.1M |
| Reserve draw | ~$12M (raised from ~$9M by correction) |
The city entered FY2026 with a structural gap, not a cash-flow blip. The city manager's transmittal letter said that after five years of "favorable conditions," the city now "faces an increasingly complex fiscal landscape" — naming public-safety pension obligations, capital-cost inflation, service demand, and contractual wage increases.
The proposal opened with a $6.5 million property tax increase, taking the overall city levy to $54,055,995 — a 13.7% rise over the prior year's $47,555,995, and a sharp break from a levy held roughly flat since 2020. The council had declined proposed increases three years running (0.05%, 4.1%, 7.1%). Holding the prior-year levy flat required a "menu of options" plus significant reserve spending; staff estimated that avoiding reserve spending entirely would have required a $14.2 million hike.
The composition of the increase is the important part, because it determines how much of it software could ever touch.
| Destination | Amount | Share | Nature |
|---|---|---|---|
| Police & fire pension funds ($1.5M each) | $3,000,000 | 46% | Statutory obligation |
| Human Services Fund — existing deficit + crisis-response team | $2,500,000 | 38% | Pre-existing service gap |
| New dedicated Parks & Recreation Fund | $1,000,000 | 16% | Accounting restructure |
| Total | $6,500,000 | 100% | — |
A further $500,000 General Fund levy increase was offset by an equal Solid Waste Fund reduction, made up through yard-waste and bulk-pickup rate increases.
Pensions dominate the structural picture beyond the increase itself. The FY2026 fire pension contribution is $9,838,575 and police $10,651,282; the gap between operating transfers in and out — the total pension contribution — is $30.3 million. The council adopted a policy targeting full funding by 2040, and part of the reserve draw exists specifically to keep that schedule.
Two revenue assumptions worth watching
One-time construction revenue is rolling off. The city collected $19 million in permit fees over two years from a major university's construction programme. Federal funding reductions to that university have slowed campus development, and future permit projections fall with it. A construction-cycle bulge was, for two years, doing structural work.
Inflation-elevated receipts are assumed to persist. Sales, income and hotel taxes reached record highs on inflation, and the FY2026 budget assumes they stay elevated. That is an assumption, not an observation — exactly the kind that should be tracked continuously against actuals rather than revisited annually.
02 — The finding: the burden moves after the levy vote
The public argument each autumn is about the size of the levy. But the levy sets how much is collected — not who pays it. The share question is settled downstream, in a county process with almost no municipal-facing visibility, and it moves by more each year than most levy fights do.
Following reassessment, the township's total assessed value rose 28% to $1,495M — residential up 27% to $993M (66.4% of base), non-residential up 30% to $502M (33.6%). At the assessment stage the two classes moved close to proportionally, with non-residential slightly ahead. The county Assessor's office has also sharply reduced "regressivity" — the historic pattern of overtaxing lower-priced homes while undervaluing higher-end ones — and a University of Chicago study found county homeowners saved $1.9 billion under those reforms.
Then the appeals run. The county Board of Review has, per the Assessor's office, systematically shifted 3 to 4 percent of the tax base onto residential properties every year through commercial appeals.
At assessment Residential 66.4% ($993M) | Non-residential 33.6% ($502M)
↓
Board of Review commercial appeals: 3–4% of base per year
↓
After appeals Residential ~69.9% | Non-residential ~30.1%
(illustrative, one year at 3.5%)
The 3–4% figure is county-wide, from the Assessor's office. The second line is arithmetic on that figure, not a measured outcome for this municipality. Quantifying the local shift is the first thing any pilot should do.
The mechanism is capacity, not conspiracy
Commercial property owners routinely retain professional tax representation to prosecute appeals. Individual homeowners largely do not. The appeal channel is open to both classes and used effectively by one. That is not a scandal — it is an asymmetry of capacity, and it is the clearest case in this fiscal system where the framing rule's own prescription applies exactly as written: capacity augmentation for the under-resourced side.
Note what this reframes. If the residential share drifts upward by three points a year through a channel no one contests, then a council fighting over a 13.7% levy increase is arguing about the smaller of the two variables determining a homeowner's bill. "Evenly distribute the tax burden" is therefore not a budget question at all. It is an appeals-capacity question — and unlike the levy, it can be worked on without anyone losing a service.
03 — Services and revenue: what is actually amenable
The useful cut is not a list but a classification: for each stream, is the constraint informational (software can help), statutory (it cannot), or political (it must not)?
| Stream | Figure | Constraint | Agentic role |
|---|---|---|---|
| Property tax levy | $54,055,995 | Political | Surface only |
| — of which police & fire pension | $30.3M transfer | Statutory | None |
| Sales / income / hotel taxes | record highs | Informational | Forecast variance |
| Local 1% grocery tax | ~$2.5M/yr | Political | Incidence |
| Building permits & fees | $19M / 2 yrs | Informational | Yield + leakage |
| Real estate transfer tax | — | Informational | Reconciliation |
| Water fund | +23.3% prior yr; flat FY2026 | Informational | Cost-of-service |
| Sewer fund | −17.5% prior yr; flat FY2026 | Informational | Cost-of-service |
| Solid waste fees | $7/mo opt-out | Political | Incidence |
| Grants & intergovernmental | — | Informational | Capture |
| Reserves / fund balance | ~$12M draw | Political | Surface only |
Streams marked "—" were not sourced to an FY2026 figure and are shown as structure only; no amounts have been invented for them.
| Service | FY2026 position | Agentic role |
|---|---|---|
| Police & fire | Pension-dominated; new HQ study $800K | Out of scope |
| Crisis-response team (non-police 911 response) | +70% spend, 2 FTE, +18 hrs/wk, +17% calls | Demand forecasting |
| Human Services (victim services, workforce dev, LTC ombudsman) | ~$2.35M deficit carried on reserves; cannot repeat | Caseload + intake |
| Parks & Recreation | New fund, $16,260,202 in/out incl. $5,558,702 GF transfer | Cost transparency |
| Public works & capital | CIP $90,518,361; >half via Water/Sewer | Sequencing |
| Solid waste | Levy cut, offset by rate increases | Incidence |
| Public library | Separate levy set by its own board | Out of scope |
One distributional note worth surfacing rather than deciding: the new local 1% grocery tax is regressive in form, but its incidence falls mostly on non-residents shopping at stores along the city's borders. Surfacing that is the tool's job. Deciding whether it is acceptable is not.
04 — Six applications, each with its leash
Ordered by expected value. Every one is additive to staff capacity; none reduces headcount, and none outputs a determination.
A1 · Appeal-equity engine. Two halves. For residents: given a parcel's characteristics and comparable sales, produce a plain-language read on whether an appeal is warranted and assemble the evidence — the service commercial owners already buy. For the city: monitor Board of Review outcomes in aggregate and quantify the local residential shift, so the number above stops being a county-wide inference. Leash: never files on a resident's behalf; never predicts a specific dollar outcome; publishes its own hit rate. If it cannot beat the base rate, it is withdrawn.
A2 · Budget legibility on the civic calendar. The FY2026 process ran a public hearing, ward meetings across two months, a Truth in Taxation hearing, and final adoption — a resident has weeks to form a view on a $404M budget. This renders it as: what this means for a home at your ward's median value, which trade-offs are live, and what changed since the last version — every figure linked to its budget line. Leash: no recommendation on how to vote or testify; both sides of every trade-off rendered with equal effort; a diff, not a summary.
A3 · Revenue yield and leakage reconciliation. Cross-check permits issued against fees assessed and collected; licences against active businesses; transfer-tax filings against recorded deeds. This recovers revenue already owed under existing rates — the one form of "reduce the burden" that takes nothing from anyone who is paying what they owe, and the only one that lowers pressure on the levy without a service cut. Leash: produces a worklist for staff review, never an enforcement action; no automated notice reaches a resident or business.
A4 · Capital sequencing. More than half the $90.5M CIP flows through the Water and Sewer Funds, alongside street, park and facility work. Dependency modelling catches the classic waste — resurfacing over a main scheduled for replacement, or opening the same trench twice — and models the multi-year cost of deferral rather than only its first-year saving. Leash: proposes sequences, never selects projects; project selection is a council and ward-equity decision.
A5 · Continuous cost-of-service models. The water and sewer cost-of-service models were updated for FY2026, and the result let a scheduled 13.5% water increase be deferred a year with rates held flat. That is exactly the mechanism worth running continuously rather than annually — a rate deferral found early is worth more than the same finding at adoption. Leash: models the rate requirement; setting the rate stays with council.
A6 · Service-demand forecasting. The budget records that one additional crisis-response team of two responders adds 18 hours of weekly coverage and answers 17% more calls. That is a usable demand-to-capacity ratio. Forecasting call patterns lets coverage be placed where it is needed — an argument for staffing, which is why it belongs in a capacity-augmentation tool rather than an efficiency one. Leash: forecasts demand, never triages an individual call; no involvement in any live emergency decision.
05 — The rig, and its governance components
None of Section 04 is lawful or trustworthy on someone else's cloud. Municipal data includes resident records, and a city cannot outsource custody of them to a vendor relationship it does not control. An independent, on-premises rig is not a preference — it is the precondition.
Hardware
| Tier | Cost | What it runs |
|---|---|---|
| Existing 8 GB GPU (RTX 3070 Ti class) | $0 | All four models below; 7B QLoRA fine-tuning at the floor |
| Used RTX 4090, 24 GB | $900–1,100 | The recommendation. ~80–85% of 5090 throughput at ~¼ the price |
| RTX 5090, 32 GB | $4,700–4,830 | ~35% faster for ~4× the money. Not justified |
| Cloud A100 80 GB | $0.67/hr | Adaptation runs only — never resident data |
Priced August 2026 and volatile.
Model choice — where performance and licence agree
| Model | Mean (N=3) | SD | s/task | Licence |
|---|---|---|---|---|
| qwen2.5:7b-instruct | 100.0% | 0.0 | 1.1 | Apache 2.0 |
| gemma3:4b | 92.9% | 0.0 | 2.0 | Gemma Terms of Use |
| qwen2.5-coder:7b-instruct | 85.7% | 0.0 | 1.1 | Apache 2.0 |
| gemma4:e4b | 81.0% | 8.2 | 5.2 | Apache 2.0 |
14 tasks × 3 runs, temperature 0.1, measured 20 August 2026. Licences as reported by ollama show --license.
Three results that change the decision. The general Qwen model beat the purpose-built coder of identical size and speed by 14 points — so choosing by the label "coder" picks the weaker tool. Model size predicted nothing: the smallest beat the largest. And the least stable model swung 71/86/86 across identical runs, meaning a single benchmark run would have certified it.
The licence finding. Performance and compliance point at the same model. Qwen2.5 is Apache 2.0 — permissive, OSI-approved, no field-of-use restriction, clean for a public-sector deployment and for fine-tuned derivatives. The second-place model carries a custom vendor licence that is not open source and imposes use restrictions and obligations following model derivatives. For a municipality — where procurement review is real and derivative weights would be trained on local data — that difference matters more than the seven-point score gap. The model that wins the benchmark is also the one with the cleanest paperwork.
Governance and compliance components
Six of the ten below exist today as a tested gate that hash-pins an immutable policy, resolves verdicts on a most-restrictive-wins lattice (DENY > REQUIRE_HUMAN > ALLOW), and writes a hash-chained attestation. Re-run 20 August 2026: 15/15 pass, integrity pinned-ok, including detection of a deliberately tampered ledger record.
| # | Component | What it does | Status |
|---|---|---|---|
| 1 | No-egress constraint | Core reasoning cannot reach the network; a hard DENY, not a setting | Built |
| 2 | Human gate on consequential actions | REQUIRE_HUMAN for anything consequential; unknown actions default to REQUIRE_HUMAN | Built |
| 3 | Hash-pinned immutable policy | Editing the policy changes its hash and shows up in the audit chain | Built |
| 4 | Tamper-evident attestation ledger | Append-only hash chain; tampering is detected and located | Built |
| 5 | Consent gate for sensitive data | Medical/biometric without consent → DENY, with regulatory triggers cited | Built |
| 6 | Verify-before-change + rollback | Unverified self-modification cannot proceed unreviewed | Partly — rollback advisory |
| 7 | Model & dependency provenance | SHA-256 pin every weight file; SBOM the stack; record licence per artefact | To build |
| 8 | Boot & disk integrity | Full-disk encryption, measured boot / TPM, physical custody | To build |
| 9 | Records-law posture | Assume every artefact is disclosable; retention schedule; no shadow record | To build |
| 10 | Rubber-stamp detection | Capture the reviewer's rationale, not their click — an unengaged human defeats component 2 | To build |
Three compliance items that generalise
Clean-room discipline. The rig must never ingest material it has no right to hold. An independent rig that touches restricted data is not independent; it is a liability with a GPU attached.
Erasure is unsolved. A resident's right to have data deleted collides with a knowledge graph, episodic memory and embeddings. Deleting a contributed fact is a machine-unlearning problem, not a row delete. Status: honestly UNKNOWN. Mitigate by keeping resident data in a segregated, deletable store and out of the graph entirely.
Advisory or regulated — no middle ground. The same rule that keeps clinical decision-support outside device regulation applies here: the moment the system issues a determination to act on without independent review, its regulatory character changes. Keep it transparently advisory.
06 — What this cannot do
The largest single driver of the FY2026 increase is not addressable by software at all. Pension contributions of $9.8M (fire) and $10.7M (police), inside a $30.3M transfer, against a council-adopted 2040 full-funding target, are statutory obligations to named beneficiaries under state law. No agentic system reduces them. Any proposal implying otherwise should be treated as a warning about the proposal.
The second-largest component, $2.5M for the Human Services Fund, covers a deficit the city says it cannot carry on reserves again — for victim services, workforce development, a long-term care ombudsman and the crisis-response team. Staff wrote that it "would be extremely challenging to make cuts to these vital programs that support those most vulnerable in our community." That is precisely the case where an efficiency framing does damage: the "saving" available is a service the vulnerable depend on, and calling its removal optimisation is the laundering the framing rule exists to stop.
So of the $6.5M increase, 84% is pension obligation and an acknowledged service deficit. The honest answer to "use agentic technology to reduce the tax burden" here is that the burden is mostly not made of inefficiency. What is genuinely available is narrower and still worth having: recovering revenue already owed (A3), avoiding coordination waste in a $90.5M capital programme (A4), finding rate deferrals earlier (A5) — and, much larger than any of those, contesting the share drift in Section 02.
Deliberately not estimated. This paper attaches no dollar savings figure to any of the six applications. I could produce one; it would be invented. Yield recovery, sequencing gains and appeal outcomes are all measurable — but only against a specific city's own baseline, which is exactly what a pilot exists to establish. A savings number produced before the baseline is the failure mode this method is built to prevent, and it is the number a vendor would lead with.
07 — If it were ever piloted
The originating proposal already fixes the scale: a city of roughly this size is right; a major metropolis is not a first pilot under any framing. Small enough for one department and one accountable official.
- One application, not six. A1's city-facing half — quantify the local residential share drift from Board of Review outcomes. It needs no resident data, touches no service, threatens no job, and answers a question the city cannot currently answer.
- Publish the baseline first. Three prior years of measured shift, before any tool is built. If the drift turns out to be negligible locally, the finding is that the premise was wrong — and that is a successful pilot.
- One accountable official who can stop it, and whose rationale is captured in the attestation ledger rather than inferred from a click.
- Success is not adoption. It is whether a resident, a councilmember and a staff member can each state what the system does and where it fails.
Standing conclusion. The levy debate is loud, annual and public. The share drift is quiet, continuous and procedural. Agentic technology is close to useless against the first and genuinely useful against the second — which is fortunate, because the second is where "evenly distribute" actually lives, and it is the half that can be worked on without anyone losing a service.
08 — Limitations
- Two levy figures circulate for FY2026 — $54,055,995 total proposed, and a separately reported $38,478,400 excluding debt service, general assistance and the library. They measure different scopes. Anything depending on the exact figure should be read from the adopted budget document.
- Proposed and adopted figures differ. The proposal carried $341,636,161 in revenue; the adopted budget $338,636,161. The reserve draw rose from ~$9M to ~$12M on correction.
- The 3–4% shift is county-wide, not municipality-specific, and is the Assessor's office's characterisation of the Board of Review. Establishing the local number is step one, not a result.
- No dollar savings are estimated anywhere in this paper, by choice.
- This is independent analysis of public documents. It was not commissioned by, affiliated with, or endorsed by any municipality, and no view is attributed to any official beyond what they published.
- No lawyer has reviewed any of this, including the records-law and licence readings in Section 05.
- The framing rule was applied, not overridden. The reduction half of the brief was answered by separating cost-of-service and revenue-yield from the levy, and by declining to treat tax minimisation as an objective.
Sources
- Adopted FY2026 budget of the studied municipality, and contemporaneous public reporting on its proposal, hearings and adoption.
- County Assessor's office — township property value releases; University of Chicago study on assessment reform; the Assessor's published characterisation of Board of Review commercial appeals.
- Local coding benchmark:
experiments/local_coding/multirun.py, N=3, 20 August 2026; model licences viaollama show --license. - AURI project, Custom coding LLM — scope, hardware, cloud, 15 August 2026.
- AURI project governance gate and compliance-components analysis (15/15 tests, re-run 20 August 2026).
Prepared by Claude Code, development assistant to the AURI project. Municipality and officials de-identified by choice. Independent analysis of public documents.