Agent-Driven Policy
June 16, 2026
A legislator is not, fundamentally, a holder of opinions. A legislator is a cognitive engine asked to convert the chaos of a society into a small number of binding, enforceable, legitimate rules—and the engine is catastrophically underpowered. The defining failure of modern government is not corruption or cowardice; it is throughput. A single human mind, backed by a thinning staff, cannot read a four-thousand-page omnibus, scan fifty jurisdictions for what already worked, weigh how severe a problem truly is, estimate whether it can be moved at all, ground a decision in the research, and predict how millions of people will respond. So each of those faculties gets outsourced to whoever arrives with the answer pre-chewed—and the only actors who can afford to pre-chew it are the best-funded interests. The agent does not replace the legislator. It rebuilds the missing faculties one by one.
The first faculty is Comprehension: the ability to see how the system actually works—what the existing law already says, where it contradicts itself, which statutes are dead, who really benefits. Today this faculty barely exists; legislators vote on text they have not read and cannot, structurally, find time to read. An agent reads all of it, continuously, and turns the opaque corpus of accumulated law into a queryable map.
The second faculty is Significance: the discipline of deciding which problems are even worth a law. Legislative attention is the scarcest resource in a republic, and it is allocated by noise—by whichever crisis trends, whichever lobby shouts loudest. An agent can triage a thousand candidate problems by reach, severity, and reversibility, turning a politics of reaction into a politics of deliberate prioritisation.
The third faculty is Tractability: the sober estimate of how hard a problem is to actually move. Most political energy is spent on problems that look urgent but are structurally immovable, while tractable wins go unnoticed. An agent can model expected effect size against implementation difficulty, separating the problems a law can solve from the ones it will only perform solving.
The fourth faculty is Diffusion: the capacity to learn from everyone who already tried. The fifty states and the hundred-ninety countries are a vast, running experiment, and almost none of that evidence reaches the drafter in time. An agent mines the entire global record of policy—what spread, what worked, what backfired—and delivers proven templates instead of blank pages.
The fifth faculty is Evidence: the loyalty to what the research actually shows rather than what the talking point asserts. The evidence base is enormous and growing, and it is almost entirely unscanned by the people writing law. An agent grounds every claim in the studies, the trials, and the data—and, critically, supplies that grounding without a client behind it.
The sixth faculty is Simulation: the power to test a law before it is binding. We ship software behind a staging environment and a rollback button; we ship law to a continent on a floor vote and a hope. An agent war-games legislation against a synthetic population, surfacing the second- and third-order effects—the cobra-breeders, the gaming, the perverse incentives—in silico, before they hit reality.
The seventh faculty is Composition: the act of turning settled intent into precise statutory text. This is the one task already visibly migrating to machines, from a city ordinance drafted by a chatbot to a national drafting assistant trained on a million sections of law. Done well, it collapses the cost of writing good law; done carelessly, it floods the system with bad law faster than ever.
The eighth faculty is Constituent Sensing: the ability to hear what the public actually needs, directly and at scale, rather than through the filter of whoever can manufacture the loudest voice. Today a representative’s sense of the public is a handful of town halls and a flood of form letters; when millions of comments arrive, the genuine signal drowns. An agent listens to all of it, strips out the astroturf, and renders the real distribution of need.
The ninth faculty is Deliberation: the discipline of forcing a proposal to survive its strongest objections before it becomes law. Legislatures vote under time pressure and tribal reflex, rarely steelmanning the other side or naming who pays. An agent cross-examines every bill—generating the best case against, the trade-offs, and the role-reversal test—so the decision rests on public reasons, not on whoever held the floor.
The tenth faculty is Oversight: the loop that learns whether a law actually worked. Most legislation is passed once and never revisited, accumulating as dead statute no one tests. An agent measures every law against its own stated goals, flags failure early, and triggers the revision or repeal that turns lawmaking from a one-way act into a system that learns.
This article is a field guide to the ten capabilities of the Augmented Legislator—the Legislative Intelligence Stack, where Constituent Sensing brackets the front of the cycle and Oversight the back, with Deliberation standing between knowing and writing. Each capability is treated identically: a precise Definition, its Place in lawmaking in five aspects, the twelve principles that make it powerful, the three patterns by which it operates, the key mechanisms with real working examples, the way agents change the game, the four principles of that shift, and the honest advantages and disadvantages. The article closes with a phased Action plan for building the Stack inside a real legislature without surrendering the one thing that must remain human: the vote.
Summary
1) Comprehension
What it is — The faculty of seeing the existing system as it really is: the full corpus of law, its contradictions, its dead letters, its true beneficiaries.
How it works — Continuous reading and structural mapping of statutes, precedents, and proposed text into a queryable model.
Why it matters — You cannot reform a system you cannot see; comprehension is the precondition for every other faculty.
Failure mode — Voting blind: passing text no human has read or understood, captured by whoever summarises it.
2) Significance
What it is — The triage faculty: deciding which problems are meaningful enough to deserve scarce legislative attention.
How it works — Scoring candidate problems by reach, severity, urgency, and reversibility into an explicit priority order.
Why it matters — Attention is the binding constraint of a republic; misallocating it wastes the whole machine.
Failure mode — Government by trending crisis: loud problems crowd out large ones.
3) Tractability
What it is — The realism faculty: estimating how hard a problem is to actually move with a law.
How it works — Modelling expected effect size against implementation difficulty, cost, and resistance.
Why it matters — Effort spent on immovable problems is the largest hidden waste in politics.
Failure mode — Performative legislation: passing laws that look like solutions but cannot bite.
4) Diffusion
What it is — The learning faculty: mining other jurisdictions for policies that already worked.
How it works — Scanning the global record of adoption, outcomes, and failures to surface proven templates.
Why it matters — Most problems have been solved somewhere; reinvention is pure waste.
Failure mode — Parochial blindness: drafting from scratch while the answer sits in another statehouse.
5) Evidence
What it is — The grounding faculty: tying decisions to what research and data actually show.
How it works — Retrieving, weighing, and citing studies, trials, and evaluations against each claim.
Why it matters — Without evidence, law is narrative; with it, law can be corrected.
Failure mode — Lobbyist epistemics: the best-funded interest supplies the “facts.”
6) Simulation
What it is — The foresight faculty: testing a law against a model of the world before it is binding.
How it works — War-gaming policy on synthetic populations and economic models to expose second-order effects.
Why it matters — Unintended consequences are where good intentions go to die.
Failure mode — Shipping to 330 million people with zero unit tests.
7) Composition
What it is — The drafting faculty: converting settled intent into precise, conflict-free statutory text.
How it works — Generating and red-lining legal language grounded in the existing corpus.
Why it matters — The gap between intent and text is where loopholes and litigation live.
Failure mode — Legislative spam: cheap drafting that floods the system with volume, not law.
8) Constituent Sensing
What it is — The input faculty: hearing what citizens actually need, at scale, beneath the manufactured noise.
How it works — Collecting, deduplicating, and classifying public input while filtering astroturf and fraud.
Why it matters — A representative who cannot hear the represented governs blind to them.
Failure mode — Mistaking the loudest manufactured campaign for the public will.
9) Deliberation
What it is — The reasoning faculty: stress-testing a decision against its strongest objections.
How it works — Generating the opposing case, the trade-offs, and the role-reversal test.
Why it matters — A law unexamined by its best critics is a law waiting to fail.
Failure mode — Tribal reflex: passing on “our side” rather than on public reasons.
10) Oversight
What it is — The feedback faculty: learning whether a law actually worked after passage.
How it works — Measuring real outcomes against stated goals and triggering revision or repeal.
Why it matters — Without a feedback loop, laws accumulate as dead, unexamined sediment.
Failure mode — Ghost laws: passed once and never revisited.
The Capabilities
1) Comprehension
Definition
Comprehension is the faculty of accurately seeing the system a legislator proposes to change—the full body of existing law, its internal contradictions, its obsolete provisions, and its real-world beneficiaries—before touching it.
It functions as the legislature’s situational awareness layer: the precondition that makes every downstream faculty possible, because no problem can be triaged, no law simulated, and no text drafted against a system that is invisible to the person governing it.
Place in lawmaking: 5 aspects
The precondition for legitimacy
A vote on unread text is a vote without consent of the mind that casts it.
Comprehension is what converts a signature into an actual decision.
The map of the existing corpus
Statute is accreted over centuries; no single mind holds it.
Comprehension turns that sediment into a navigable structure.
The contradiction detector
New law collides with old law in ways drafters rarely foresee.
Comprehension surfaces conflicts before they become litigation.
The dead-letter finder
Much law is obsolete, redundant, or never enforced.
Comprehension distinguishes living rules from fossils.
The beneficiary lens
Every rule moves value to someone; the question is whom.
Comprehension makes the distributional reality legible.
Why it works: 12 principles
Externalised memory — it stores the corpus outside any single overloaded staff.
Structural reading — it maps relationships (this section amends that one), not just words.
Completeness — it reads all of the text, not the fraction a human samples.
Cross-reference — it links proposed text to every statute it touches.
Provenance — it traces where language came from and who supplied it.
Comparability — it sets current law beside the proposed change, clause by clause.
Continuity — it persists across electoral cycles, immune to staff turnover.
Speed — it reads in minutes what once took staff weeks.
Searchability — any clause becomes retrievable on demand.
Version awareness — it tracks how text mutated across drafts and amendments.
Scale-invariance — a thousand-page bill is no harder to read than a one-pager.
Neutrality — it gives every provision equal attention, not selective focus.
Three major patterns of how it works
Ingest → structure → query
Ingest the raw corpus and the proposed text
Structure it into linked clauses, definitions, and cross-references
Expose it to natural-language interrogation
Compare → flag → explain
Compare new language against existing law
Flag conflicts, redundancies, and dead letters
Explain each flag in plain language
Trace → attribute → expose
Trace clauses to their textual origin
Attribute them to a source (agency, interest, model bill)
Expose the provenance to the legislator and the public
Key mechanisms and how they work
A. Statutory research and pruning systems
Models that read the entire code and locate the relevant, redundant, or obsolete law.
Example: Stanford’s RegLab built a statutory-research system that identified relevant law with 94–99% reliability; deployed with the San Francisco City Attorney, it produced an ordinance cutting more than a third of the city’s mandated reports.
B. Code-scale deregulation analysis
Running an entire administrative code through analysis to flag what is unnecessary.
Example: Ohio ran its roughly fifteen-million-word administrative code through an AI analysis that flagged two million words and some 900 rules for removal, putting the state on track to cut nearly a third of the code.
C. Provenance and model-legislation detection
Computational comparison that reveals who actually wrote a bill.
Example: The “Copy, Paste, Legislate” investigation analysed nearly a million state bills and found more than 10,000 copied almost verbatim from interest-group “model legislation,” over 2,000 of which became law.
How AI changes the game: definition
AI turns comprehension from a sampling problem into a total-coverage problem—reading the whole corpus, mapping its structure, and answering questions about it in real time—while shifting the risk from “we missed something” to “we over-trusted the summary.”
In short: the legislator can finally read everything.
Four principles of how AI changes the game
From sampling to totality — from reading a fraction of the text to processing all of it.
From text to structure — from prose pages to a linked, queryable graph of law.
From periodic to continuous — from a one-time read to an always-current model of the corpus.
From opaque to attributed — from anonymous clauses to traceable provenance.
Advantages and disadvantages
Advantages
Ends the absurdity of voting on unread text.
Surfaces conflicts and dead letters before they cause harm.
Exposes hidden authorship and beneficiaries.
Gives a small office the reading capacity of a large institution.
Disadvantages
A confident, wrong summary is more dangerous than an honest gap—automation bias is real.
Whoever tunes the comprehension model shapes what the legislator “sees.”
Structural maps can flatten the deliberate ambiguity that law sometimes needs.
Total legibility of the corpus is also a tool for those who would exploit it.
2) Significance
Definition
Significance is the faculty of deciding which problems are meaningful enough to warrant scarce legislative attention—weighing how many are affected, how severe the harm, how urgent the timing, and how reversible the damage.
It functions as the legislature’s triage layer: the discipline that allocates the single most constrained resource in a republic—the finite attention of its lawmakers—toward the problems that actually matter rather than the ones that merely shout.
Place in lawmaking: 5 aspects
The attention allocator
There are always more problems than legislative slots.
Significance decides what gets a hearing and what does not.
The severity weigher
Not all harms are equal; some are catastrophic, some cosmetic.
Significance ranks by magnitude, not volume of complaint.
The reach estimator
A problem affecting millions differs from one affecting hundreds.
Significance scales attention to population touched.
The reversibility filter
Irreversible harms deserve priority over recoverable ones.
Significance privileges the problems that cannot wait.
The agenda guard
Agendas are captured by whoever manufactures urgency.
Significance defends the agenda against manufactured noise.
Why it works: 12 principles
Comparability — it puts dissimilar harms on a common scale.
Proportionality — it matches attention to magnitude.
Explicitness — it makes the priority order visible and defensible.
Resistance to noise — it discounts volume in favour of severity.
Forward weighting — it privileges the irreversible and the compounding.
Coverage — it scans the whole problem space, not the trending slice.
Auditability — it leaves a record of why a problem was prioritised.
Multi-dimensionality — it weighs reach, severity, urgency, and reversibility together.
Counterfactual framing — it asks what happens if nothing is done at all.
Stakeholder breadth — it counts the silent affected, not only the vocal.
Recurrence sensitivity — it flags chronic problems that never spike but never resolve.
Revisability — priorities update as conditions change.
Three major patterns of how it works
Scan → score → rank
Scan the full landscape of candidate problems
Score each by reach, severity, urgency, reversibility
Rank into an explicit priority order
Aggregate → weight → triage
Aggregate signals of harm across data sources
Weight by magnitude and population
Triage into act / monitor / ignore
Compare → justify → publish
Compare a problem against the current agenda
Justify its place with explicit criteria
Publish the reasoning for scrutiny
Key mechanisms and how they work
A. Severity thresholds and common currencies
Institutions already triage life-and-death allocation with explicit severity metrics.
Example: The UK’s NICE allocates health spending against an explicit cost-per-quality-adjusted-life-year threshold, with a formal “severity modifier” that raises the bar a society will pay for the most severe conditions—a working machine for ranking meaningfulness.
B. Evaluation coverage as a significance signal
Knowing which programs are unexamined reveals where attention is missing.
Example: Reformers behind the U.S. evidence-based-policy movement estimate that only a small fraction of public spending is rigorously evaluated, and propose setting aside as little as 1% of program funds for evaluation—evidence that significance is currently unmeasured.
C. The ghost-law problem
Laws passed and never revisited are significance failures by default.
Example: Scoping reviews of ex-post legislative evaluation find that the societal impact of most laws is rarely measured after passage, leaving “ghost laws” on the books with no one asking whether they still matter.
How AI changes the game: definition
AI turns significance from an implicit, noise-driven reflex into an explicit, continuous triage—scoring a thousand candidate problems by reach and severity in the time a staffer reads one lobbyist memo—while raising the danger that whatever the model fails to count becomes invisible.
In short: prioritisation becomes deliberate, not reactive.
Four principles of how AI changes the game
From loudest to largest — from the problem that trends to the problem that matters.
From episodic to continuous — from crisis-driven attention to standing triage.
From implicit to explicit — from gut ranking to a defensible, published score.
From narrow to comprehensive — from the visible slice to the whole problem space.
Advantages and disadvantages
Advantages
Protects the agenda from manufactured urgency.
Surfaces large, quiet problems that never trend.
Makes prioritisation transparent and contestable.
Aligns scarce attention with actual magnitude of harm.
Disadvantages
What the model cannot quantify, it may silently de-prioritise.
Severity scoring embeds contestable value judgments as if neutral.
A triage metric, once public, becomes a target to be gamed.
Quantified significance can crowd out legitimate moral salience that resists numbers.
3) Tractability
Definition
Tractability is the faculty of estimating how hard a problem is to actually move—how large an effect a law can realistically produce, against how much cost, complexity, and resistance it must overcome.
It functions as the legislature’s realism layer: the discipline that separates problems a law can genuinely solve from problems a law can only perform solving, redirecting effort from the immovable to the achievable.
Place in lawmaking: 5 aspects
The effect-size estimator
Some interventions move the needle; many do not.
Tractability forecasts the realistic magnitude of impact.
The difficulty appraiser
Implementation, enforcement, and compliance all cost.
Tractability prices the friction of making a law bite.
The resistance map
Every law meets opposition proportional to whose value it moves.
Tractability anticipates where the law will be fought.
The leverage finder
Small, well-placed changes can outperform sweeping ones.
Tractability locates the high-leverage intervention point.
The futility filter
Some problems are structurally beyond a single statute.
Tractability flags where law is the wrong instrument.
Why it works: 12 principles
Expected value — it weighs impact by probability of success, not hope.
Cost realism — it counts implementation and enforcement, not just intent.
Resistance modelling — it forecasts opposition and capture.
Leverage focus — it seeks the minimal change with maximal effect.
Mechanism clarity — it demands a causal story for why a law would work.
Boundary honesty — it admits where law cannot reach.
Comparability — it ranks interventions by achievability, not ambition.
Path dependence — it accounts for what current structures actually permit.
Time horizon — it distinguishes quick wins from slow burns.
Reversibility of the fix — it favours interventions that can be undone if wrong.
Enforcement realism — it weighs whether a rule can actually be policed.
Coalition feasibility — it estimates whether the votes and allies exist to pass it.
Three major patterns of how it works
Model → estimate → discount
Model the causal mechanism
Estimate the raw effect size
Discount by implementation difficulty and resistance
Decompose → locate → target
Decompose a problem into movable and immovable parts
Locate the high-leverage component
Target the intervention there
Forecast → stress → revise
Forecast the expected outcome
Stress it against opposition and evasion
Revise the ambition to match what can bite
Key mechanisms and how they work
A. Effect-size evidence from trials
A growing body of randomised trials gives realistic priors on how much an intervention moves.
Example: The development-economics network J-PAL has run nearly a thousand randomised controlled trials across more than eighty countries, producing concrete effect sizes that tell a drafter whether a given lever historically moved the outcome at all.
B. Calibrated, low-cost interventions
Cheap, well-targeted changes can have outsized, measurable effects.
Example: The UK’s behavioural-insights work found that a single rewritten tax-reminder letter—telling recipients most neighbours had already paid—was estimated to raise tens of millions a year, a high-tractability win invisible to grand legislation.
C. The futility signal from backfires
History records interventions whose tractability was misjudged and which moved the problem the wrong way.
Example: Research on “three-strikes” sentencing found it flattened the penalty gradient so severely that eligible offenders became measurably more likely to commit violent crimes—an immovable problem made worse by a law that looked decisive.
How AI changes the game: definition
AI turns tractability from a gut feel into a modelled estimate—pulling real effect sizes from the global trial record and weighing them against implementation friction—while risking false precision that dresses guesswork as forecast.
In short: ambition gets calibrated to what can actually move.
Four principles of how AI changes the game
From hope to expected value — from “this should work” to “this historically moved X.”
From intent to friction — from the goal to the real cost of enforcing it.
From sweeping to leveraged — from grand gestures to minimal high-impact changes.
From certainty to calibrated doubt — from false confidence to honest probability.
Advantages and disadvantages
Advantages
Redirects effort from immovable problems to achievable ones.
Grounds ambition in real historical effect sizes.
Exposes the implementation friction politicians routinely ignore.
Surfaces cheap, high-leverage interventions that never make headlines.
Disadvantages
Effect sizes from one context transfer imperfectly to another.
Quantified tractability can bias toward the easily measured and against the structurally important.
A low-tractability score can become an excuse for inaction on hard, vital problems.
Modelled forecasts carry false precision that invites over-trust.
4) Diffusion
Definition
Diffusion is the faculty of learning from every jurisdiction that already faced a problem—mining the fifty states and the hundred-ninety countries for the policies that spread, the ones that worked, and the ones that backfired.
It functions as the legislature’s import layer: the mechanism that converts the world’s running policy experiment into proven templates, so a drafter starts from what already succeeded elsewhere rather than from a blank page.
Place in lawmaking: 5 aspects
The laboratory harvester
Sub-national and foreign governments are live experiments.
Diffusion harvests their results for reuse.
The template supplier
Most problems have a workable solution somewhere.
Diffusion supplies it instead of a blank draft.
The failure archive
Other jurisdictions have already made the mistakes.
Diffusion imports the warnings, not just the wins.
The implementation-detail carrier
The difference between success and failure is often a detail.
Diffusion transfers the how, not only the what.
The patchwork tracker
Reforms move unevenly across dozens of legislatures at once.
Diffusion keeps the moving map current.
Why it works: 12 principles
Reuse — it avoids reinventing solved problems.
Evidence of feasibility — a policy that ran elsewhere is proof of possibility.
Outcome transfer — it carries results, not just designs.
Failure avoidance — it imports others’ mistakes as warnings.
Detail fidelity — it transfers the implementation specifics that decide success.
Timeliness — it surfaces proven options before the drafting deadline.
Breadth — it scans more jurisdictions than any human could track.
Context matching — it weights examples by similarity to local conditions.
Adaptation over copying — it adjusts templates rather than transplanting them blind.
Recency — it privileges current results over stale precedent.
Counter-diffusion awareness — it tracks where reforms were repealed or banned, not only adopted.
Source diversity — it draws from many jurisdictions, avoiding single-model dependence.
Three major patterns of how it works
Scan → match → adapt
Scan the global record for analogous problems
Match the closest proven policy
Adapt it to local constraints
Trace → evaluate → import
Trace where a policy spread
Evaluate its measured outcomes
Import the version that worked
Detect → warn → adjust
Detect where a policy backfired
Warn the drafter of the failure mode
Adjust the design to avoid it
Key mechanisms and how they work
A. Cross-jurisdiction policy mining
The “laboratories of democracy” only help if someone reads the results.
Example: When New York implemented cordon congestion pricing it explicitly followed London’s earlier rollout, down to the implementation detail of pairing the charge with expanded bus service to absorb displaced drivers.
B. Tracking a reform across many legislatures at once
Proven templates spread unevenly and fast across dozens of statehouses.
Example: Right-to-repair legislation has now been introduced in all fifty U.S. states—a patchwork no single staffer can track, but exactly the moving map a diffusion agent maintains.
C. Global outcome evidence for a single policy
The same policy run in many countries yields a distribution of outcomes to learn from.
Example: More than forty countries have adopted sugar-sweetened-beverage taxes, leaving a documented range of consumption effects—from modest to large—for the next adopter to study before drafting.
How AI changes the game: definition
AI turns diffusion from occasional, anecdotal borrowing into systematic, continuous mining of the entire global policy record—surfacing proven templates and documented failures on demand—while risking the uncritical transplant of policies whose context does not travel.
In short: the drafter starts from what already worked.
Four principles of how AI changes the game
From anecdote to corpus — from a remembered example to the whole record.
From design to outcome — from copying a law’s text to copying its measured results.
From wins-only to failures-included — from cherry-picked success to honest distribution.
From snapshot to live map — from a one-time scan to a continuously updated tracker.
Advantages and disadvantages
Advantages
Eliminates the waste of reinventing solved problems.
Carries implementation details that decide success or failure.
Imports others’ mistakes as cheap warnings.
Keeps a live map of reforms moving across many jurisdictions.
Disadvantages
A policy that worked in one context can fail in another; transplant is risky.
Diffusion can entrench convergence and suppress local experimentation.
The same machinery lets interest groups spread model legislation faster, too.
Outcome data from abroad is uneven, lagged, and sometimes politicised.
5) Evidence
Definition
Evidence is the faculty of grounding legislative decisions in what research and data actually show—retrieving, weighing, and citing the studies, trials, and evaluations that bear on a claim, rather than the assertions supplied by whoever is in the room.
It functions as the legislature’s grounding layer: the discipline that ties law to reality and, decisively, supplies that grounding without a client—breaking the monopoly under which the best-funded interest is also the source of the “facts.”
Place in lawmaking: 5 aspects
The reality anchor
Law detached from evidence is narrative with force.
Evidence keeps the claim tethered to the world.
The subsidy replacement
Today, research and drafting labour is donated by lobbyists.
Evidence supplies the same subsidy with no donor attached.
The claim auditor
Every justification rests on an empirical premise.
Evidence checks whether the premise is true.
The uncertainty reporter
Honest evidence carries its own error bars.
Evidence states what is known and what is not.
The correction enabler
Only an evidenced law can be falsified and fixed.
Evidence makes legislation a testable hypothesis.
Why it works: 12 principles
Loyalty to data — claims stand or fall on quality, not source.
Causal identification — it distinguishes correlation from cause.
Effect sizes — it asks not just whether, but how much.
Provenance — it attributes every fact to a traceable source.
Uncertainty honesty — it reports confidence, not just conclusions.
Independence — it owes nothing to the interest that benefits.
Falsifiability — it names what would prove the claim wrong.
Replication weighting — it trusts findings that reproduce over one-off results.
Conflict reconciliation — it resolves contradictory studies rather than cherry-picking one.
Robustness balance — it weighs novel findings against established ones.
Relevance filtering — it prefers evidence from comparable populations and contexts.
Method transparency — it exposes how a conclusion was reached, not just the conclusion.
Three major patterns of how it works
Retrieve → weigh → cite
Retrieve the relevant research
Weigh it by quality and relevance
Cite it against the specific claim
Synthesise → reconcile → report
Synthesise findings across studies
Reconcile conflicting results
Report a confidence-weighted conclusion
Verify → flag → correct
Verify each cited source exists and says what is claimed
Flag fabrication and overreach
Correct before the claim is acted on
Key mechanisms and how they work
A. The standing evidence base
A vast, mostly unread body of rigorous research already exists.
Example: The UK’s What Works network spans policy areas accounting for hundreds of billions in public spending, synthesising evidence for decision-makers—proof that the supply of evidence already outstrips the bandwidth to use it.
B. Institutionalised evidence mandates
Governments have legislated the demand for evidence even where the labour is scarce.
Example: The bipartisan U.S. Foundations for Evidence-Based Policymaking Act required agencies to build evidence-building plans and appoint Chief Evaluation Officers—an explicit statutory demand for grounding that agents can help supply.
C. Verification against hallucination
The grounding faculty fails catastrophically if the “evidence” is invented.
Example: In Mata v. Avianca, lawyers filed a brief citing six entirely fabricated precedents produced by a chatbot; in a separate case, an expert defending a deepfake statute filed sworn testimony with AI-hallucinated citations and was excluded—proof that an evidence agent without a verification layer is worse than none.
How AI changes the game: definition
AI turns evidence from a scarce, lobbyist-supplied subsidy into an abundant, on-demand, client-free resource—retrieving and weighing the research behind any claim in seconds—while introducing a new failure mode: confident fabrication that must be caught before it is cited.
In short: the subsidy finally has no master.
Four principles of how AI changes the game
From supplied to retrieved — from facts handed over by an interest to facts pulled from the record.
From assertion to citation — from “studies show” to a traceable source.
From scarce to continuous — from a one-off literature review to standing grounding.
From trust to verification — from believing the output to checking its provenance.
Advantages and disadvantages
Advantages
Breaks the lobbyist monopoly on policy information.
Grounds every claim in a traceable, weighable source.
Reports uncertainty instead of false certainty.
Makes law a falsifiable hypothesis that can be corrected.
Disadvantages
Hallucinated citations can launder fabrication as scholarship.
Automation bias leads officials to over-trust the cited output.
Evidence informs but cannot settle value disagreements—it can smuggle values as facts.
The training data and the model’s tuner both shape what counts as “evidence.”
6) Simulation
Definition
Simulation is the faculty of testing a law against a model of the world before it becomes binding—war-gaming its effects on a synthetic population and economy to expose the second- and third-order consequences a drafter never imagined.
It functions as the legislature’s staging-environment layer: the missing test harness that lets a society run a law in a sandbox—surfacing the gaming, the perverse incentives, and the distributional losers in silico—before it ships to millions.
Place in lawmaking: 5 aspects
The consequence engine
Laws fail at the second order, not the first.
Simulation reveals the downstream effects.
The gaming detector
Every rule is an optimisation target for those it binds.
Simulation surfaces the evasion in advance.
The distributional X-ray
Aggregate effects hide who wins and who loses.
Simulation shows the losers before the vote.
The rollback substitute
Law has no easy undo; mistakes are costly.
Simulation is the cheap rehearsal that prevents them.
The behavioural realism layer
People respond, adapt, and evade.
Simulation models behaviour, not just arithmetic.
Why it works: 12 principles
Foresight — it moves error discovery before enactment.
Behavioural modelling — it captures how people actually respond.
Distributional resolution — it disaggregates winners and losers.
Adversarial testing — it lets the rule be gamed in safety.
Cheapness of failure — a failed simulation costs nothing.
Scenario range — it explores many futures, not one forecast.
Iteration — it lets the law be revised before it bites.
Assumption transparency — it states what the model takes for granted.
Sensitivity analysis — it shows which assumptions the result depends on.
Emergence capture — it surfaces effects no one deliberately designed in.
Calibration — it is checked against real-world outcomes where they exist.
Comparability — it scores the proposal against the status quo and alternatives on a common scale.
Three major patterns of how it works
Model → run → observe
Model the population and economy
Run the proposed law against it
Observe the emergent effects
Perturb → adapt → expose
Perturb the system with the new rule
Let simulated agents adapt and evade
Expose the gaming behaviour
Score → compare → revise
Score outcomes across scenarios
Compare against the status quo
Revise the law before enactment
Key mechanisms and how they work
A. Microsimulation and rules-as-code
Tax and benefit law is already tested against synthetic populations before scoring.
Example: France’s OpenFisca encodes tax-and-benefit law as executable code so a reform can be simulated before it is passed; open successors such as PolicyEngine put the same capability in a citizen’s browser.
B. Agent-based models of the economy
Whole economies can be simulated as interacting agents for “what-if” policy design.
Example: Central banks, including the Bank of England, have moved agent-based macroeconomic models from the seminar room into the operating toolkit, and the EU funded agent-based engines explicitly for policy design.
C. Behavioural and synthetic-population simulation
Generative agents now reproduce real human responses closely enough to poll.
Example: A Stanford study built generative agents of over a thousand real people from interviews; the agents reproduced their human counterparts’ survey answers about 85% as accurately as the humans reproduced their own answers two weeks later. In a separate model, simulated workers spontaneously learned to avoid a tax code—surfacing the gaming before it could hit the real economy.
How AI changes the game: definition
AI turns simulation from siloed, expert-only microsimulation into a general staging environment for any law—modelling behaviour, gaming, and distribution across a synthetic society—while risking over-trust in models that are fragile, gameable, and only as honest as their assumptions.
In short: law finally gets a test harness.
Four principles of how AI changes the game
From arithmetic to behaviour — from static scoring to modelled human response.
From aggregate to distributional — from a single number to who-wins-who-loses.
From forecast to war-game — from one projection to adversarial scenarios.
From narrow domains to all law — from tax-only microsimulation to general policy testing.
Advantages and disadvantages
Advantages
Moves the discovery of unintended consequences before enactment.
Surfaces gaming and evasion in safety.
Reveals distributional losers the aggregate hides.
Makes failure cheap and revision routine.
Disadvantages
Models are fragile; a buggy simulation can mislead with authority.
Any simulated metric becomes a target interests will reverse-engineer and game.
Synthetic populations inherit the biases of their training data.
False confidence in a model can be more dangerous than honest uncertainty.
7) Composition
Definition
Composition is the faculty of converting settled intent into precise, conflict-free statutory text—translating a policy decision into legal language that says exactly what it means and collides with nothing it should not.
It functions as the legislature’s drafting layer: the final translation from what we decided to what the statute says, where loopholes, ambiguities, and litigation are either prevented or created.
Place in lawmaking: 5 aspects
The intent translator
A decision is not yet a law until it is text.
Composition renders intent into enforceable language.
The loophole closer
Imprecise drafting is where evasion lives.
Composition tightens the text against exploitation.
The consistency keeper
New text must cohere with the existing corpus.
Composition harmonises language across statutes.
The accessibility shaper
Law that no citizen can read loses legitimacy.
Composition can render text in plain language too.
The throughput multiplier
Drafting capacity caps how much law a body can produce.
Composition raises that ceiling—for better or worse.
Why it works: 12 principles
Precision — it says exactly what is meant.
Consistency — it aligns with definitions already in force.
Completeness — it anticipates the cases the rule must cover.
Conflict-freedom — it avoids collision with existing law.
Traceability — it links each clause to its intent.
Revisability — it red-lines and iterates quickly.
Legibility — it can produce a human-readable companion.
Speed — it produces a working draft in minutes, not weeks.
Edge-case coverage — it anticipates the situations a rule must handle.
Definitional discipline — it reuses terms already defined in the corpus.
Enforceability — it writes text that can actually be applied and adjudicated.
Plain-language parity — it keeps the readable version faithful to the legal one.
Three major patterns of how it works
Intent → draft → red-line
Capture the settled intent
Draft the statutory language
Red-line against corpus and edge cases
Generate → check → harmonise
Generate candidate text
Check for conflicts and loopholes
Harmonise with existing definitions
Translate → simplify → publish
Translate legal text into plain language
Simplify for public comprehension
Publish both versions together
Key mechanisms and how they work
A. Direct AI drafting by elected officials
Legislators are already drafting real bills with language models.
Example: A Porto Alegre councillor had a chatbot draft a municipal ordinance from a forty-nine-word prompt and the council passed it unanimously; a Massachusetts state senator used the same tools to draft an AI-regulation bill that he said got him “about seventy percent of the way there.”
B. Corpus-grounded drafting assistants
National drafting offices are building assistants trained on the full body of law.
Example: The UK’s Office of the Parliamentary Counsel built a drafting assistant grounded in roughly 1.5 million sections of legislation and tens of thousands of court cases, generating explanatory material and supporting precise legal language.
C. The volume warning
Cheap drafting raises throughput, which is not the same as raising quality.
Example: Lowering the cost of writing bills has already produced a flood—on the order of a thousand AI-related bills introduced in a few months of a single U.S. session—demonstrating that composition without judgment yields volume, not law.
How AI changes the game: definition
AI turns composition from a scarce, specialist bottleneck into an abundant, on-demand capability—drafting and red-lining precise statutory text grounded in the corpus—while collapsing the cost of producing bad law just as fast as good law.
In short: drafting stops being the bottleneck—and judgment becomes it.
Four principles of how AI changes the game
From scarce to abundant — from a specialist queue to on-demand drafting.
From blank page to grounded draft — from starting cold to starting from the corpus.
From opaque to legible — from impenetrable text to a plain-language companion.
From production-limited to judgment-limited — the constraint moves from writing to deciding.
Advantages and disadvantages
Advantages
Collapses the cost and delay of precise drafting.
Closes loopholes by red-lining against the whole corpus.
Produces plain-language versions that raise legitimacy.
Gives a back-bencher the drafting capacity of a leadership office.
Disadvantages
Cheap drafting floods the system with volume over quality.
Whoever owns the drafting model can steer statutory language at scale.
Undisclosed AI authorship raises real accountability and legitimacy questions.
Fluent text can mask substantive errors a human would have caught.
8) Constituent Sensing
Definition
Constituent Sensing is the faculty of hearing what the public actually needs—mapping the real preferences, burdens, and priorities of citizens at scale, and separating genuine signal from manufactured noise.
It functions as the legislature’s input layer: the mechanism that lets a representative perceive the people they serve directly, rather than through the filter of whoever can afford to manufacture the loudest voice.
Place in lawmaking: 5 aspects
The representation anchor
A representative who cannot hear the represented is one in name only.
Sensing restores the direct line between citizen and lawmaker.
The signal–noise filter
Organised campaigns drown out individual citizens.
Sensing separates substance from orchestrated volume.
The burden detector
Citizens pay a silent “time tax” they rarely write letters about.
Sensing surfaces friction, not only stated opinion.
The preference map
Opinion is distributed unevenly across issues and groups.
Sensing renders what the public actually wants, by segment.
The astroturf shield
Manufactured, bot-driven, and duplicated input corrupts the record.
Sensing detects and discounts it.
Why it works: 12 principles
Directness — it perceives citizens without an intermediary.
Scale — it processes millions of inputs, not a sampled few.
Signal extraction — it separates substance from form-letter volume.
Authenticity detection — it flags fabricated or duplicated comments.
Disaggregation — it sees subgroups, not just the average.
Inclusivity — it hears those who lack organised representation.
Multilingual reach — it understands input in any language.
Continuity — it listens between elections, not only at them.
Burden sensitivity — it detects friction and “time tax,” not only opinion.
Proportionality — it weights by genuine prevalence, not manufactured volume.
Privacy preservation — it can aggregate without exposing individuals.
Responsiveness — it routes real concerns to the relevant decision.
Three major patterns of how it works
Collect → dedupe → classify
Collect inputs across every channel
Remove duplicates and astroturf
Classify by topic, sentiment, and segment
Aggregate → weight → surface
Aggregate by genuine prevalence
Weight by authenticity
Surface the real distribution of need
Detect → verify → route
Detect a genuine concern
Verify it is organic
Route it to the relevant faculty
Key mechanisms and how they work
A. Public-comment analysis at scale
Agencies receive millions of comments; AI categorises, deduplicates, and flags bot-generated versus substantive input.
Example: Federal comment-analysis pipelines now compress what once took weeks of manual review into hours, triaging millions of public comments on proposed rules into topics and genuine-versus-duplicate buckets.
B. Astroturf and fraud detection
Natural-language analysis reveals manufactured campaigns hiding inside the record.
Example: Of the roughly 22 million comments on the FCC’s net-neutrality repeal, analysis found about 18 million were fake, with fewer than 800,000 genuinely organic—exactly the noise a sensing agent must strip out.
C. Direct citizen-engagement channels
Standing platforms let citizens shape decisions between elections, not only at them.
Example: Singapore’s REACH e-engagement platform gathers citizen feedback on policies directly—a model for structured, ongoing constituent input rather than episodic polling.
How AI changes the game: definition
AI turns constituent sensing from a sampled, gameable trickle into scaled, authenticated, real-time perception of the public—hearing millions directly and filtering the manufactured—while raising the danger of ever more convincing synthetic “citizens.”
In short: the representative can finally hear the represented.
Four principles of how AI changes the game
From sample to population — from a few loud voices to the whole distribution.
From form-letter to substance — from counting volume to extracting signal.
From episodic to continuous — from election-day to always-on listening.
From gameable to authenticated — from astroturf-vulnerable to fraud-aware.
Advantages and disadvantages
Advantages
Restores the representative’s direct line to the represented.
Surfaces silent burdens the vocal never raise.
Filters manufactured campaigns from genuine concern.
Hears the unorganised, multilingual, and marginalised.
Disadvantages
Generative AI also makes synthetic “constituents” cheaper and more convincing.
Aggregating citizen input at scale raises surveillance and privacy risks.
Sensing can be mistaken for a mandate, bypassing deliberation.
What the model labels “noise” may include real but unconventional voices.
9) Deliberation
Definition
Deliberation is the faculty of stress-testing a decision through argument—steelmanning the opposition, surfacing the trade-offs, applying the role-reversal test, and forcing a proposal to survive its strongest objections before it becomes law.
It functions as the legislature’s reasoning layer: the adversarial discipline that converts evidence and simulation into a justified choice, ensuring a law is defended on public reasons rather than passed on tribal reflex.
Place in lawmaking: 5 aspects
The objection generator
Drafters see the case for, rarely the strongest case against.
Deliberation manufactures the best opposing argument.
The trade-off namer
Every law has costs its sponsors prefer to leave implicit.
Deliberation makes who-benefits-and-who-pays explicit.
The role-reversal test
A rule acceptable from power may be intolerable from opposition.
Deliberation tests it from every position.
The public-reason filter
Justifications by tribe or creed do not bind a plural society.
Deliberation demands reasons any citizen could accept.
The blind-spot finder
Authors cannot see what they did not think of.
Deliberation surfaces the unconsidered.
Why it works: 12 principles
Adversarial rigour — it attacks the proposal to find its weaknesses.
Steelmanning — it builds the strongest version of the opposing case.
Impartiality — it judges arguments by merit, not source.
Trade-off candour — it names costs, not only benefits.
Reversibility test — it checks the rule from every stakeholder’s position.
Public reason — it requires justifications independent of tribe or creed.
Assumption surfacing — it exposes hidden premises.
Perspective breadth — it represents absent and minority viewpoints.
Consistency — it treats like cases alike across time and party.
Falsification framing — it asks what would change the conclusion.
Proportionality — it favours the least-restrictive effective means.
Humility — it admits the limits of what is known.
Three major patterns of how it works
Propose → attack → defend
State the proposal
Generate the strongest objections
Force a defence or a revision
Reframe → reverse → test
Reframe from each stakeholder’s view
Apply the role-reversal
Test for acceptability across positions
Expose → weigh → justify
Expose the trade-offs
Weigh them openly
Justify the choice on public reasons
Key mechanisms and how they work
A. Multi-perspective argument generation
Agents can argue every side of a question, attacking and defending in turn.
Example: Work on adversarial and multi-agent reasoning shows that pitting models against each other to attack and defend a claim surfaces weaknesses a single pass misses—a standing red-team for legislation.
B. Structured value frameworks
Explicit principles convert open argument into disciplined judgment.
Example: The Apolitical Politics framework already codifies the role-reversal test, public reasons, and trade-off candour—the exact criteria a deliberation agent can apply, clause by clause, to every bill.
C. Distributional and objection mapping
Surfacing who loses, and why, turns abstract objection into specific accountability.
Example: The same microsimulation and synthetic-population tools that reveal distributional losers feed deliberation by naming, concretely, whose interests a law moves—so the objection is grounded, not rhetorical.
How AI changes the game: definition
AI turns deliberation from a scarce, often-skipped luxury into a standing adversarial process—generating the strongest objections, the role-reversal, and the trade-off map for every proposal—while risking persuasive argument detached from truth.
In short: every bill can be cross-examined before it is passed.
Four principles of how AI changes the game
From advocacy to adversary — from arguing one side to attacking every side.
From implicit to explicit trade-offs — from hidden costs to a named distribution.
From tribal to public reasons — from “our side” to justifications all can assess.
From skipped to standing — from rushed votes to routine cross-examination.
Advantages and disadvantages
Advantages
Forces a proposal to survive its strongest objections.
Makes trade-offs and losers explicit before the vote.
Represents absent and minority perspectives.
Anchors the decision in public reasons, not reflex.
Disadvantages
Fluent argument can persuade without being true—rhetoric outruns evidence.
Endless deliberation can become a tactic for delay.
The model’s framing of “the other side” embeds its own biases.
Manufactured objections can obstruct as easily as improve.
10) Oversight
Definition
Oversight is the faculty of learning whether a law actually worked—monitoring its real-world effects after passage, evaluating it against its stated goals, and triggering revision or repeal when it fails.
It functions as the legislature’s feedback layer: the loop that converts a law from a one-time act into a testable hypothesis, closing the cycle so that legislation learns instead of accumulating as dead, unexamined sediment.
Place in lawmaking: 5 aspects
The hypothesis closer
A law is a prediction that an intervention will help.
Oversight tests whether the prediction held.
The ghost-law detector
Statutes pass and are never revisited.
Oversight finds the dead letters still on the books.
The enforcement monitor
Text on the page is not the same as a rule applied.
Oversight checks whether the law actually bites.
The sunset trigger
Some laws should expire or be revised on schedule.
Oversight flags when their time has come.
The learning capture
Each law is a lesson for the next.
Oversight turns outcomes into institutional memory.
Why it works: 12 principles
Falsifiability — it treats every law as a hypothesis with a stated test.
Goal anchoring — it measures against the law’s own declared aims.
Continuity — it watches outcomes long after the vote.
Honesty about misses — it surfaces failure rather than hiding it.
Counterfactual measurement — it asks what would have happened otherwise.
Enforcement realism — it checks application, not just text.
Timeliness — it flags failure early, not after decades.
Reversibility — it makes repeal and revision routine.
Comparability — it scores outcomes against the original forecast.
Independence — it evaluates free of the author’s stake.
Transparency — it publishes results, including the failures.
Cumulativeness — it feeds lessons forward into future legislation.
Three major patterns of how it works
Measure → compare → judge
Measure real outcomes
Compare to stated goals
Judge success or failure
Monitor → flag → trigger
Monitor enforcement and effect
Flag drift or failure
Trigger review
Evaluate → publish → feed-forward
Evaluate against the forecast
Publish the result
Feed lessons into the next law
Key mechanisms and how they work
A. Ex-post evaluation at scale
Most laws are never rigorously revisited after passage.
Example: Scoping reviews of ex-post legislative evaluation find the societal impact of laws is “rarely measured” after enactment, leaving “ghost laws” on the books—precisely the gap a standing oversight agent closes.
B. Institutional evaluation mandates
Governments have legislated the demand for post-hoc evidence even where the labour is scarce.
Example: The U.S. Foundations for Evidence-Based Policymaking Act requires agencies to appoint Chief Evaluation Officers and assess the coverage and quality of their evaluations—an oversight mandate agents can help fulfil continuously.
C. Corpus pruning as oversight
AI can identify obsolete or redundant law for repeal.
Example: Stanford’s RegLab statutory-research system and Ohio’s code review both used AI to find outdated statutes—oversight applied to the existing corpus, surfacing what should be revised or removed.
How AI changes the game: definition
AI turns oversight from a rare, after-the-fact audit into continuous, automated evaluation—measuring every law against its goals and flagging failure early—while risking metric-driven judgment that mistakes the measurable for the meaningful.
In short: law finally learns from its own results.
Four principles of how AI changes the game
From one-time act to standing hypothesis — from “passed” to “still working?”
From decades to real-time — from belated review to early warning.
From hidden to published — from buried failure to transparent result.
From isolated to cumulative — from each law alone to lessons that compound.
Advantages and disadvantages
Advantages
Closes the loop so legislation learns from outcomes.
Surfaces ghost laws and obsolete statutes for repeal.
Makes failure visible early, when it is still cheap to fix.
Feeds concrete lessons into the next round of lawmaking.
Disadvantages
Metric-driven oversight can optimise the measurable and miss the meaningful.
Continuous monitoring carries surveillance and privacy risks.
Evaluation framed by the model can embed its own definition of “success.”
Automated repeal flags could strip protective laws under the banner of efficiency.
Action plan
The Augmented Legislator is not built by buying a chatbot. It is built by installing the Legislative Intelligence Stack as public infrastructure, owned by the institution and the citizen rather than by a vendor or an interest. The sequence matters: comprehension and evidence first, simulation and composition last, with legitimacy designed in at every layer.
Phase 1: Give the legislature its own eyes and ears (Comprehension + Evidence + Constituent Sensing)
Deploy a public statutory-comprehension model. Stand up a corpus-grounded system over the full body of law, with conflict detection, dead-letter flagging, and provenance on every clause—so no representative ever again votes on text no one has read.
Install a client-free evidence layer. Pair every bill with an evidence agent that retrieves, weighs, and cites the research behind each claim—with a mandatory verification step that catches fabricated citations before they reach the floor.
Stand up a constituent-sensing channel. Collect, deduplicate, and authenticate public input at scale, filtering astroturf, so the representative hears the real distribution of need—not the loudest manufactured campaign.
Mandate provenance and uncertainty. Require that every machine-supplied fact carry its source and confidence, and every machine-flagged conflict carry a plain-language explanation.
Phase 2: Make prioritisation deliberate (Significance + Tractability)
Build a standing problem-triage register. Score candidate problems by reach, severity, urgency, and reversibility, publish the ranking, and force any deviation from it to be justified.
Attach a tractability estimate to every proposal. Require a modelled effect size and an implementation-friction appraisal—grounded in the global trial record—before a bill consumes a hearing.
Log the dropped problems. Publish what the triage de-prioritised and why, so significance failures are visible, not silent.
Phase 3: Test before you ship (Diffusion + Simulation)
Stand up a diffusion service. Maintain a live map of analogous policies across jurisdictions, with measured outcomes and documented failures, so every draft starts from what already worked.
Require a staging run for material laws. Before enactment, war-game the bill against a microsimulation, an economic model, and a synthetic population—surfacing gaming, perverse incentives, and distributional losers—and publish the result.
Adopt rules-as-code. Encode the operative provisions as executable logic so the law can be simulated, queried, and re-tested as conditions change.
Phase 4: Deliberate, then draft with judgment (Deliberation + Composition + Legitimacy)
Cross-examine every bill. Before drafting, require a deliberation pass that generates the strongest objections, names the trade-offs and the losers, and applies the role-reversal test—so the decision rests on public reasons, not on whoever held the floor.
Give every representative their own drafting agent. Open-weighted where possible, auditable, logged, and grounded in the corpus—so the drafting subsidy that lobbyists once monopolised belongs to the back-bencher and the public, not the best-funded interest.
Bind the Stack to the Apolitical Politics oath. The agent supplies the mechanism, the evidence, the simulation, and the text; the human owns the role-reversal test, the public reasons, the candid naming of who benefits and who pays, the preference for reversible choices—and the vote. The machine does the is; the accountable human, within inviolable rights, chooses the ought.
Distribute ownership as the antidote to capture. Refuse a single official model. The defence against an agent that writes laws for its owner is many competing, auditable agents owned by representatives, parties, and citizens—legitimacy through plurality, not monopoly.
Phase 5: Oversee, measure, and learn (Oversight)
Treat every law as a standing hypothesis. Attach stated goals and a falsification test at passage; monitor real-world outcomes and enforcement; flag ghost laws and trigger the revision or repeal of what fails.
Publish the guarantees. Track time-to-evidence before a vote, the share of bills with a published simulation and an ex-post evaluation plan, the proportion of proven cross-jurisdictional policy actually surfaced, and the falsification record of laws once passed.
Keep a human override at every layer. When a model fails, revert to human procedure; treat every automated step as advisory to an accountable person.
Named deliverable: the Legislative Intelligence Stack Charter—a per-office specification of the ten capabilities, their ownership and audit rules, their binding to the legislator’s oath, and the published outcome metrics by which the citizen judges whether the augmented legislature is, in fact, governing better. The law was never too complex for democracy. Democracy was simply never given enough mind. The Stack gives it one—without ever taking away the vote.
Recent Posts
The Czech Agentic State :: An Architecture
June 8, 2026
The Agentic State: The Sixteen Principles
June 5, 2026