A commentary published this morning in CalMatters asks a question that should be uncomfortable for anyone who has voted yes on a California ballot measure: did we get the results we were promised?
Since 2000, the state has issued $196 billion in general obligation bonds and now has $81.8 billion outstanding. This fiscal year the state will spend nearly $8.6 billion servicing that debt. Schools, water systems, climate programs, housing — voters approved them all. The money went out. Whether the outcomes materialized is, as the piece puts it, “surprisingly difficult to answer.”
The distinction the article draws is sharp and worth sitting with. California tracks how bond dollars are spent. It does not consistently measure whether those investments achieved the outcomes voters expected. Spending accountability and outcome accountability are not the same thing, and California has built robust infrastructure for the first while doing almost nothing about the second.
This is not a California problem. It is a government technology problem. And it has a technical solution that we can describe precisely.
The difference between spending records and outcome records
A spending record answers: did the money go where we said it would go? An outcome record answers: did the thing we funded work?
Spending records are tractable. Money moves through accounts. Accounts have ledgers. Auditors can follow the trail. California’s bond spending infrastructure exists precisely because the trail is followable.
Outcome records are harder for a structural reason. Outcomes are distributed, delayed, and contested. A school bond funds construction. Construction completes. Whether student outcomes improve over the next decade depends on factors the bond measure cannot control. The causal chain is long and the signal is noisy.
But “hard to measure” is different from “impossible to measure” and very different from “not measured at all.” The current situation is not that California measures outcomes imperfectly. It is that California largely does not require outcome measurement as a condition of bond authorization.
The result is predictable. Programs are funded. Programs run. Nobody checks whether the stated goal materialized. The political incentive to check is weak — the bond measure is already voted, the money is already spent, and the officials who made the promises have often moved on.
The Oak Park, Illinois surveillance case is the clearest recent example of what this failure mode looks like at the contract level. The city’s oversight board concluded in 2025 that its Flock Safety license plate reader cameras had played no meaningful role in any crime investigation during three years of operation. The cameras were deployed, the contracts were paid, the city was promised crime reduction outcomes, and nobody measured whether those outcomes materialized until a board specifically tasked with looking decided to look. Oak Park ended the contract.
That is the bond accountability problem in miniature: a government program funded on the basis of promised outcomes, operated for years, never measured against those promises, and discovered to be delivering nothing only when someone with the time and mandate to check actually checked.
What outcome measurement infrastructure actually requires
The CalMatters piece calls for “consistent outcome measurement” as a policy goal. That framing is right but underspecified. Consistent outcome measurement requires three things that California currently lacks at the system level.
Outcome definitions that are measurable at the time of authorization. A bond measure that promises to “improve water systems” cannot be measured because “improved” is not defined in a form that survives contact with data. A bond measure that promises to reduce system failure events by 30% within five years, measured against a defined baseline, can be measured. The difference is not ambition — it is specificity. Outcome definitions need to be written into bond authorization language the same way spending limits are written in.
Immutable baseline records. You cannot measure change without a baseline. The baseline must be established before the program begins, recorded in a form that cannot be retroactively adjusted, and held outside the control of the program being measured. This is not a novel requirement — financial auditing has required immutable baseline records for decades. The technical infrastructure to generate and maintain them exists.
Independent, machine-verifiable outcome records. The records of program outcomes need to be independently verifiable — meaning any qualified party can verify their integrity without asking the program itself for confirmation. This requires the same properties that make financial audit trails trustworthy: append-only storage, cryptographic chaining, and external anchoring so the record exists outside the control of the entity being measured.
All three of these are solvable engineering problems. None of them are currently required by California’s bond authorization process.
Why the current approach fails even when people are trying
The CalMatters piece is careful to note that the absence of consistent outcome measurement is not purely a failure of will. Program administrators often want to measure outcomes. They run into problems that are partly political and partly technical.
The political problem is that outcome measurement creates accountability surfaces that program administrators and their political patrons prefer not to have. A program that cannot be measured cannot be found to have failed. This incentive runs through the entire system.
The technical problem is less discussed but equally real. Most government programs generate operational records — logs, reports, databases — that were designed for program management, not for independent audit. These records are typically stored in systems controlled by the program being measured. They can be amended. They can be selectively retained or purged. The chain of custody from “what the program said it measured” to “what the program actually produced” is almost always breakable.
This is the same structural problem that makes surveillance footage unreliable as evidence. In both cases, the records are held by the party whose performance the records are supposed to document. In both cases, the records can be adjusted — not necessarily through deliberate fraud, but through selective retention, inconsistent methodology, and the ordinary pressures of organizational self-presentation. In both cases, the solution is the same: push the integrity of the record outside the control of the party being measured.
What the technical solution looks like
The architecture that makes records independently verifiable is the same whether you are auditing AI agent decisions, surveillance camera chain of custody, or government program outcomes.
Establish a baseline at program inception. Record it in append-only storage where entries cannot be modified after writing. Compute a cryptographic hash of the baseline record — a fingerprint that changes if the record changes. Chain that hash to subsequent outcome records using the same append-only structure. Periodically anchor the hash chain to an external immutable store — Arweave is the practical implementation — so the record exists permanently outside any party’s control, publicly verifiable by anyone with the transaction identifier.
The result is a record whose integrity can be mathematically verified without trusting the party that created it. Not “we checked and it looks right” but “here is the cryptographic proof that this record has not been altered since it was written, and here is the Arweave transaction identifier that anyone can use to verify that proof independently.”
This is not speculative. The same Merkle-chaining and external anchoring that makes AI agent audit trails independently verifiable applies directly to government program outcome records — documented across ClawQL’s security framework at docs.clawql.com/security/best-practices. High-trust document management, auditable program records, and independently verifiable outcome trails are the same architecture under a different authorization surface.
The Oak Park surveillance example is instructive here too. The failure was not that the data didn’t exist — the cameras were generating data continuously. The failure was that nobody required outcome measurement at the time of contract authorization, and the data that existed was held entirely within the vendor’s infrastructure where the city had no independent verification path. A properly structured program — whether a bond-funded infrastructure project or a surveillance camera contract — would have defined measurable outcomes at authorization, established an immutable baseline, and required that outcome records be held in a form the contracting authority could independently verify.
The legislative opening
California is currently moving new bond measures through the legislature. Assemblymember David Alvarez has introduced a bond bill intended to fund both the modernization of academic facilities and the addition of affordable student housing throughout California’s university systems. A separate $10 billion housing bond proposal would direct seven billion dollars toward California’s Multifamily Housing Program, which helps low-income residents.
Both of these measures will go to voters. Both will promise outcomes. Neither, under current California practice, will require that those outcomes be measured in an independently verifiable way.
The model for what the authorization language should look like is not complicated. Every bond measure should specify measurable outcomes with defined baselines and measurement periods. Every program funded by bond proceeds should be required to maintain outcome records in WORM storage with external immutable anchoring. Outcome records should be publicly accessible for independent verification through a defined API. The State Auditor should have a standing mandate to run periodic cryptographic verification of outcome records across all active bond programs.
This is not a significant cost relative to the scale of the programs being measured. External anchoring on Arweave costs under a dollar per day for most programs. The audit infrastructure to generate and maintain the records is available today. The political cost of requiring this is that programs can be found to have failed. That cost is a feature, not a bug.
The CalMatters commentary asks why we don’t track results. The answer is partly incentive structure and partly the absence of infrastructure that makes tracking trustworthy. The incentive structure problem requires political will. The infrastructure problem is already solved. The question is whether California will require its use.