Hello (Hammer) Worlds

You finish a report. Your name is on it. Months later, somebody with authority to say no challenges one figure.

Which file did it come from? Which month? What denominator sat behind it? Was it measured, cited, or assumed? Does the source support the sentence you wrote, or only a nearby sentence that sounded close enough at the time?

That is when analytical work becomes expensive. Producing the number was quick. Reconstructing why anybody trusted it is not.

Today we are releasing Hammer Worlds publicly.

Hammer Worlds is for people who do not merely produce numbers. They have to defend them months later.

Hello, world

Every report assumes a world.

It assumes a set of sources. It assumes a date. It assumes which populations can be compared, which prices mean the same thing, which missing rows mean nothing, and which conclusions may follow from available evidence.

Most reports leave that world implicit. Hammer makes it explicit.

Before a Hammer world helps write anything, it introduces itself:

• What evidence it holds.
• When that evidence was read.
• Which rules govern its use.
• Which comparisons are valid.
• What it has not seen.
• Which questions its sources cannot answer.
• What evidence would settle each refused question.

This is not a caveat page at the back. It is the beginning of the work.

A world model needs more than records. It needs a corpus, the evidence available inside the world, and physics, the rules governing what can follow from that evidence.

Together, corpus and physics give an agent Judgement.

Frontier world models

People rarely begin with a clean analytical question. They begin with a goal: show that a service can be paid for, defend an existing code, challenge a proposed payment change, or build a case a plan’s actuary will accept.

A goal is not yet a claim. A correct refusal can still leave its writer at a dead end.

Hammer maps the frontier between the report somebody wants to write and the evidence available to support it.

Each part of the argument lands somewhere:

• Settled here. This world holds the evidence and names the route to it.
• Settled elsewhere. A relevant source exists, and the world tells you whether it has been acquired, converted, or remains unread.
• Yours to declare. Your payer mix, contract terms, population, costs, and operating assumptions belong to you. No public corpus can supply them.
• Refused permanently. Available records cannot settle the claim. Hammer keeps that wall in place and names the nearest supportable question beside it.

The frontier does not write the report, settle claims, rank alternatives, or invent missing evidence. It shows where each part of the work belongs before you spend time answering the wrong question.

Who this release is for

This release is built for somebody whose deliverable faces adversarial review after a lag:

• A market access or HEOR lead building the payment argument a payer analyst will read.
• A vendor-side evidence owner deciding which reimbursement claims can survive actuarial review.
• A coverage-argument or comment-letter writer defending fee-schedule and payment claims before a Medicare contractor or another reviewer.

Their shared problem is not access to another number. It is knowing which number will bear weight, preserving where it came from, and refusing the sentence that will collapse when somebody checks it.

This release is not built for high-volume market maps, denial-rate estimates, quick coverage conclusions, or an unweighted market average for a deck. Its sources cannot support those claims. Hammer says so before they travel.

What ships in the first public world

The released world holds:

• Published commercial negotiated rates for 115 billing codes. Source: rate extract read 27 August 2026 and shipped inside world vintage 31 August 2026.
• Medicare Physician Fee Schedule and Clinical Laboratory Fee Schedule records, including status and professional or technical components.
• Medicaid National Correct Coding Initiative procedure-to-procedure edits, with their own quarter named separately from payer-rate vintage.
• Medicare Local Coverage Articles read for furnishing and interpretation requirements.
• Hospital price-transparency filings under 45 CFR 180, kept separate from payer negotiated rates because a filed charge is not a paid rate.
• Published effect sizes carrying study population, interval, and PMID, with observational and randomised evidence separated.
• A source register distinguishing evidence that is shipped, available but unread, unavailable, or gated by licence.

You can look up what a named commercial payer published it pays for a named billing code. Each result keeps its source vintage and contract-row count attached. It is not an industry average and does not pretend to describe a payer mix you did not declare.

You can position that rate against a Medicare fee schedule where a national dollar anchor exists. Where the schedule provides relative value units instead, Hammer returns those units without inventing a dollar conversion from factors the world does not hold.

You can check whether a code pair carries an edit in the Medicaid edition of the National Correct Coding Initiative file. A missing rule returns unknown, never permission.

You can read what a Medicare contractor’s Local Coverage Article says about who may furnish or interpret a code. That result stays bound to the contractor, enrolment type, and source that published it.

You can model screening yield over sensitivity, specificity, prevalence, and population you declare. Every input is restated on the answer. Careful arithmetic over assumptions remains an assumption.

Every figure keeps its evidence attached

A rate cell carries its own contract-row count, spread relative to its median, and row-count thickness relative to the strongest cell for that code in the same extract. The comparison cell and its count remain visible so another reader can redo the arithmetic.

These measures are evidence, not verdicts. Hammer does not decide whether a wide spread is acceptable for your use. It tells you how thin the evidence is and leaves that consequential decision with you.

A reporting-floor rule is enforced. A rate cell backed by fewer than 10 contract rows is refused rather than served. Source: reporting-floor rule in the released world. Every affected answer distinguishes a withheld cell from a genuinely absent cell. Once those states are flattened, no later analysis can recover the difference.

Every answer also carries a receipt. A later reader can hand it back and either reproduce the answer under the same world and build or receive a refusal stating what moved.

The failure being prevented is quiet drift: last quarter’s analysis returns a different figure and nobody can tell whether the model changed or the world did.

A refusal is an answer

The released world publishes 25 permanent refusals. Source: released world manifest, vintage 31 August 2026. Each refusal names what would be required to settle that class of question.

This world will not tell you:

• A denial rate or probability that a claim gets paid. Negotiated-rate files contain prices, not adjudications, so the necessary denominator does not exist.
• Whether a payer covers a code. A published rate says what a payer pays when it pays, not whether it will.
• That a missing rate means non-coverage. The supportable statement is that no rate was recorded in the disclosure read for that vintage.
• A cross-payer market average without weights you declare. An unweighted average describes this corpus’s payer mix, not your market.
• A clinical conclusion about a named device or test from procedure-pricing records.
• What was in force after the world’s vintage. Rates and policy documents move on different clocks, so a claim carried into a later vintage must be rechecked.
• Utilisation, and therefore revenue, from a price. Volume lives in adjudicated claims, not negotiated-rate disclosures.
• Facts about your own book. Your contracts, payer mix, population, costs, and run rate are yours to declare.

The released world documents that upstream payer files are typically removed after about 90 days. Source: its shipped world description. Preserving vintage and receipt is therefore part of the answer, not administrative metadata.

If the first useful thing Hammer does is refuse a sentence, the product is working. Read what would settle it.

Measured, cited, or assumed

Every figure carries an origin:

• Measured: read from a register held by this world.
• Cited: published elsewhere, with its source attached.
• Assumption: declared by the user.

A result computed from several inputs inherits the weakest origin among them. Arithmetic over assumptions remains an assumption.

Before a claim enters a report, Hammer checks whether the figures served by the world support it. Arithmetic claims are decided by arithmetic. Claims about named records or text are checked against those records.

Neither route falls back to the other. Rewording a failed quantitative claim does not buy approval from a weaker check. Anything asserted without a check remains visible in the final report.

Two examples

Professional and technical payment are not interchangeable

For CPT 91200, the 2026 Medicare Physician Fee Schedule held in this release serves a global row of 1.01 relative value units, a professional component of 0.33, and a technical component of 0.68. Source: 2026 CMS Physician Fee Schedule corpus shipped with this world.

The rows stay separate. Adding either component to the global row counts the same work twice. Comparing a professional-only payer rate with a global rate compares different services and produces a spread that does not mean what the table says.

Hammer also declines to turn those units into dollars because the world does not hold the annual conversion factor or locality adjustment required for that step.

Predictive value depends on population

A screening model receives sensitivity, specificity, prevalence, and population as declared assumptions. Prevalence keeps its measurement class, such as claims, direct measurement, self-report, or modelling.

What comes back is the corresponding table, predictive values, and every input restated. Hammer does not convert one prevalence class into another because the world holds no factor licensing that conversion.

The report therefore carries the population where the result applies rather than letting a specialist-clinic estimate travel silently into population screening.

Install Hammer Worlds

Hammer Worlds is available through the public Hammer plugin repository. It connects your agent to a hosted world. No corpus is downloaded to your machine, and no API key is pasted into the plugin.

Claude Cowork

Open Cowork > Customize > Plugins. Under Personal plugins, choose +, then Add marketplace. Add hmmrlabs/hammer-plugin, install Hammer Worlds, and complete connector sign-in.

Claude Code

/plugin marketplace add hmmrlabs/hammer-plugin
/plugin install hammer-worlds@hammer

Choose user scope when prompted, then open the connection and complete browser sign-in.

Codex

codex plugin marketplace add hmmrlabs/hammer-plugin
codex plugin add hammer-worlds@hammer

Complete browser sign-in during installation or from the installed connection.

Other Agent Skills clients

npx skills add hmmrlabs/hammer-plugin --skill hammer-worlds -a <profile> --yes

Replace <profile> with a profile supported by the Skills CLI. Where the client installs only the skill, add https://worlds.hammer.ai/mcp through its MCP settings and complete sign-in there.

Plugin is licensed Apache-2.0.

Read release, install Hammer Worlds, and bring the claim you need to defend.