hammer-provider-quality Scope: world hammer-coverage-intelligence vintage 2026-10-01 Revision SHA-256: 42e04526fa4b1834af49bf0f805fb21196ca195ee3006d3ca62a382104a95156 --- name: hammer-provider-quality description: Use when a provider-side quality or clinical director has a portfolio slot to fill and needs to know which improvement programme is worth it, whether a site really is worse than its peers, whether a before-and-after was the intervention or the mean coming back, or what a quality withhold and a savings model are actually worth to them. Carries the screen that runs before the pitch rather than after, as an ordered routing over estimators with nothing computed in this file; the margin rule that turns a boundary into a distance and says what the absence of a margin means; the redesign move each kind of death names, with the step it repairs, how many changes it takes, and the re-screen that has to clear before it is offered at all; a worked example quoted in bounds because bounds are what the receipts carry; and the refusal this seat has to publish rather than work around, because what an avoided admission is worth to you turns on a contract this world has not read. Read hammer-coverage-navigate for tool-level routing, hammer-building-the-evidence for the method underneath, and hammer-writing-the-deliverable when the answer has to become a document. --- # hammer-provider-quality **A director does not arrive with a candidate. They arrive with a portfolio slot and a room to defend.** That is the whole difference between this seat and the contracting one beside it. A contracting lead has a named provider and a named code and wants a figure. A quality director has a budget line, a committee meeting, and several ideas that all sound reasonable, and the expensive mistake is not picking the wrong one. It is picking one, building the programme, running it for a year, and finding out at review that the improvement was the mean coming back. So this file is a screen you run **before** the pitch rather than a defence you mount after it. **`tools/tools.json` is the authority on what exists and what each call takes, and `world.toml` is the authority on what this world declines to answer.** This file is the order to call things in and the reason the order is what it is. Where this file and the tool surface disagree, the tool surface is right and this file is the defect. ## Nothing on this page is computed on this page Every statistical and actuarial technique in this world is an estimator or it is refused. The screen below is a **routing order over estimators**, not a calculator: no reliability is worked out in a chat turn, no variance is combined by hand, no bound is interpolated to a midpoint. If a step needs arithmetic no estimator here performs, the step's job is to **name that gap** and hand it back, because a figure produced in prose carries no receipt and nothing downstream can check it. ## The four reading rules Quoted from the director who specified them, because she is the one who needed them and reached them the expensive way. > 1. **That the money runs backwards.** An avoided admission is revenue lost under > fee-for-service and revenue kept under risk, and this world cannot tell her which she is in. > 2. **Reliability before ranking, always.** `measure_reliability` should be the first call > whenever a comparison is about to become a consequence. The routing table buries it. > 3. **`regression_to_mean_expectation` paired with `competing_event_coupling`** as the standard > two punches on any before-and-after. > 4. **That `withhold_earnback`'s refusal to net is a feature**, because forfeited revenue and > incurred cost point at different remedies. **Rule 1 is the first section of this file's refusal, further down**, because it is not a gap in the data that a fetch would close. It is a fact about your contract that nobody here has read. **Rule 2 is step one of the screen**, which is what this file does about the burial. The tool-level table says to reach for `measure_reliability` first whenever a comparison is about to become a penalty; here it is the step that has to clear before any other step is worth running. **Rule 3 is why step three of the screen carries two estimators rather than one.** A before-and-after has two ways of flattering you at once and they are independent: you selected the cohort for being extreme, and the denominator is a survivor count that mortality moves. **Rule 4 is the one that reads as a limitation until you need it.** Against a declared Medicaid managed care subcontract, `withhold_earnback` served a withheld pool of $960,000 as **released $480,000 and forfeited $480,000, two totals, with no field anywhere on the payload a third could be written into** (r57, her section 7.1). That absence is deliberate. Forfeited revenue is money that never arrived and incurred cost is money that went out, and the remedies are different people's work. `deferred_versus_avoided` holds the same line on the savings side: **prevented $739,300 against deferred $525,000, where the naive slide reads $1,264,300 and roughly forty percent of that total is a timing shift rather than a saving** (r56, her section 7.3). A slide carrying either sum is a slide the finance seat takes apart in the room. ## The screen, as an ordered routing over estimators Four steps, in this order, and the order is not a preference. Each one decides whether the next is worth running at all. | step | the question it settles | the estimator it routes to | |---|---|---| | 1 | At your volume, is the thing you want to compare measurable at all | `measure_reliability` | | 2 | Does a published effect for this mechanism exist, and does its interval cross no effect | `treatment_effect` for a randomised effect, `prognostic_association` for an observational one, never one call naming both | | 3 | Does the movement survive the mean coming back and the competing event | `regression_to_mean_expectation` and `competing_event_coupling`, both, per reading rule 3 | | 4 | Is there a money path, and what kind of money is it | `savings_methodology`, and the section below names the calls it does not make for you | **A candidate that clears all four is worth a slot. One that fails at step one was never an idea, it was a ranking artifact**, and running steps two through four on it produces a persuasive document about a difference nobody can measure. **`candidate_screen` runs that whole routing in one call, under one receipt.** Hand it your declared candidates and it walks the steps in this order, carrying each estimator's own refusal verbatim where a step stopped, and grouping the candidates by where each one died. It scores nothing and orders nothing: what comes back is a partition of the set you declared, and an argument asking for a ranking is refused by name. Call the estimators one at a time instead when you want to see a single step in isolation, or when you are checking one of this file's claims against the build in front of you; the table above is what each step routes into either way. ### Step one needs a between-provider variance, and where it comes from decides what you have `measure_reliability` requires the variance of the true rate across the providers you mean to distinguish between, and **it ships no default on purpose**: the defaulted input is the input that decides the answer, so a default would be this world choosing your conclusion. Declaring it yourself is legitimate and the answer then carries `assumption` as its basis. **`hospital_variance` reads one off the published measure corpus `cms-hospital-compare`**, for one named measure, and hands back the distribution it was computed from. Call it, take the figure, and the answer's basis is `measured` rather than `assumption`. Declaring your own is still legitimate, and the answer then says `assumption` beside it, which is the difference this seat is usually arguing about. Three things about that corpus decide what it can and cannot do for this seat, and each was measured against the file rather than assumed from its name. - **It holds 162 New York facilities against 20 measures, and not one of the 20 is a readmission measure at any horizon.** Both the measure identifier and the measure name were searched for every spelling a readmission goes under and returned nothing, on the same run that returned all 20 names, so the zero is the corpus rather than a search that never ran. `Hybrid_HWM` is the hybrid hospital-wide all-cause risk standardised **mortality** rate, not its readmission twin. So a director whose slot is a readmission programme **cannot read a measured variance for the measure they are actually managed on**. They can read one for a mortality, complication or patient-safety measure, and that is a different measure about a different event. - **19 of the 20 carry a numeric score beside a numeric denominator.** The composite patient-safety indicator `PSI_90` carries `Not Applicable` in its denominator on all 162 facility rows, so it is not a measure any denominator argument can be run against. - **A measured variance does not take your side, and it does not take the other side either.** On `MORT_30_AMI` it came out at 0.000365 and on `MORT_30_HF` at 0.000849, against the 0.0005 the director in the worked example had declared: **below on one measure and above on the other, from the same corpus by the same arithmetic**. Run it. Do not assume which way it moves your case, and do not quote one measure's variance while managing another. **The scale trap, because it does not look like an error.** A variance is in the square of its measure's units, and those units differ by measure: some are percentages, most patient-safety measures are rates per thousand. `measure_reliability` computes its sampling term on a **proportion**. Handing it a percentage-scale variance unconverted is out by ten thousand and returns a reliability just under 1 with a curve saying any denominator will do. A variance above one quarter is refused for exactly this reason, and that refusal is the cheap version of the error. **The expensive version is the one that answers, and it is now checked rather than warned about.** Sending the per-thousand conversion where the measure's score is a percentage lands inside the domain and produces a reliability two orders of magnitude too small, which nothing in the shape of the call reveals. So when you pass the variance as an object naming `hospital_variance` and a `measure_id`, `measure_reliability` re-runs that tool for that measure and compares your value against every field it serves. Which field you sent is named back to you. The corpus-scale field is refused by name. The other conversion does not answer at the value you sent: what comes back declines it, carries no reliability where an answer would carry one, and serves what the field that measure calls for gives, so you can see the figure you were reaching for. A value from another measure is refused and nothing is substituted for it. **Send the object form rather than copying the number out**, because the bare number is a declaration and asks this world to check nothing. ### The gap this world does not close at step four **Step four routes to `savings_methodology` and to nothing else.** That row used to name the whole money surface, and the screen has only ever called a single entry from it: `savings_methodology` takes the mechanism you declared, says which evidence class the case is then held to, and carries no number at all. The rate tools are not reachable from a candidate key, which is why they are not in the screen rather than an oversight in it, and a candidate that came back cleared at step four has had its evidence class settled and its money priced by nobody. So the money is a set of calls you make yourself, with their own arguments: `rate_position` and `rate_comparison` for what payers publish, `filed_charge` for what a hospital filed, `cost_of_care_exposure` to size it over a declared population, `declared_inputs` for your own book, `deferred_versus_avoided` before any total is quoted, and `withhold_earnback` or `tournament_position` where the money path is a contract term rather than a price. Every one of them lands under its own receipt, so the claim report names them separately from the screen's. The money path for an admission runs through the hospital filing side, and the loop that finds which code a filer actually carries before you spend a call guessing is written down in `hammer-provider-contracting` rather than repeated here. **A per-facility cut is served by `filed_charge` under `by_facility`, and it is off unless you send it.** Send it and every facility of that filer arrives at once, each with its own `facility_rows`. Do not read `facility_rows` as `n`: the filer grain's `n` comes off a deduplication whose key carries no facility at all, `facility_rows` is what the same key leaves with the facility tuple added, so the per-facility counts sum to at least the filer's `n` and never to less. A figure at one grain is not the same quantity as a figure at the other under a finer name. What no route reaches at all is a paid amount. Everything here is a published rate or a filed charge, and a business case scored on paid amounts has to say so and price the gap. ## The margin, which is the difference between a boundary and a plan A step that comes back "not cleared" tells you where the line is. **A step that tells you how far off you are tells you whether you are one design change away or three**, and only the second is something a director can act on this quarter. So: **declare your threshold, read the margin against it, and read the absence of a margin as this world declining to pick your threshold rather than as a missing number.** The worked case is the director's own. Her reliability came in at roughly **0.20**, which is the complement of the bound below rather than a figure served on its own, against the **0.7** the provider profiling literature conventionally uses (r52). Two things about that sentence, and they belong in the same paragraph because separating them is how the second gets lost: - The **0.7 is her convention to adopt**, taken from that literature, and it is **not this world's rule**. `measure_reliability` refuses a threshold argument by name rather than ignoring it, because a caller who sends one believes this world has an opinion about whether their measure is fit to report, and it does not. Whether a reliability is high enough to act on is a decision about consequences this world cannot see. - The margin she actually needed was not in reliability units at all. It was **874 to 876 cases to reach 0.7, against a caseload nine times smaller** (r52). That is the number that tells her she is not one design change from clearing at site grain. She needs a different unit of analysis, which is a different move from a bigger sample. Declare no threshold and you get the figure and no margin, and nothing is supplied on your behalf. That is the same refusal one layer up, and it is working. ## When a candidate dies, the death names its own repair The screen is only worth running if a refusal comes back with somewhere to go. Each row below is a move the director in the worked example either made herself or should have been handed. | what killed it | what it means | the move it names | step it repairs | changes it takes | offered only when the re-screen shows | |---|---|---|---|---|---| | Reliability kills a per-site comparison | The sites are not distinguishable at this denominator | Stop comparing sites. Pool the system and use a concurrent comparison group | 1 | Two: the unit of analysis, and the comparison group | The pooled denominator clears step one at **your** declared threshold, with the figure it would then reach | | Regression to the mean eats a before-and-after | You selected on the extreme | A concurrent control, or a pre-period long enough to select on something stable | 3 | One: the comparison, or the selection window | `regression_to_mean_expectation` on the redesigned selection predicts a return small enough to leave a signal | | A competing event eats the benefit | The endpoint is wrong for the horizon | A composite endpoint, or a shorter horizon where the competing risk is small | 3 | One: the endpoint, or the horizon | `competing_event_coupling` on the new endpoint bounds the mortality-consistent share below what you can live with | | The denominator is too thin at any design | The unit of analysis is wrong | Go up a level: service line rather than hospital, system rather than site, year rather than quarter | 1 | One per level, and each level is its own change | The level you moved to clears step one, rather than the next level up doing so | | The mechanism classifies soft | Not fatal. It changes the evidence class and the sponsor | Say so at the start rather than in review | 4 | None to the design, one to the document | `savings_methodology` returns the class you are about to present under, before the room hears a number | **A move is offered only once it has been re-run through the screen.** Telling a director to pool the system is worth nothing if the pooled denominator still fails step one, and worse than nothing if she spends a quarter reorganising the measurement to find out. So a move you hand over has already been checked against the step that killed the candidate, it names the figure that step would then read, and **a move that does not clear the step it claims to repair is not offered at all**. Where no move in this vocabulary reaches the threshold, say that, rather than listing moves that fail. **The cost wall, stated plainly.** This world holds no cost for any of these moves. It does not know what it costs to pool six hospitals into one measured unit, to stand up a concurrent control, or to extend a pre-period. So the sequence above is ordered by **how many changes a move takes and never by what it costs**, unless the costs are yours and declared, in which case they are `declared_inputs` and the ordering is yours too. Inventing a cost here would be the defaulted variance failure wearing different clothes. ## The worked example, in bounds, because bounds are what the receipts carry A director intended to target her weakest cardiac cohort, measured at **25.0 percent against a system mean of 17.4 percent**, a selection gap of 7.6 points. The screen kills it at step one and hands back the pooled concurrent-comparison design. - **At least 79 and at most 80 percent of the variance in that measure is sampling error at her denominator** (r52). That is what the receipt entails and both endpoints are the answer. The reliability of roughly 0.20 quoted above is the complement of this bound, not a separate measurement. - **Regression alone entails that the hospital measures between 0.189 and 0.190 next period with nothing done at all** (r55). Against her 25.0 percent, that interval is what entails a return of **between 6.0 and 6.1 of the 7.6 points**, unaided. Any single figure inside that range, including the tidy midpoint, asserts a precision the receipt does not settle. - **What is left is between 1.5 and 1.6 points**, and that residue is the most any intervention could have been credited with. **A programme delivering 5.5 points would have been underperforming doing nothing**, and it would have been reported as a win. **Two different 1.5-point figures live in this example and they are not the same quantity.** Getting this wrong teaches the exact conflation the screen exists to stop. - **The residue** above: what is left of a 7.6-point **selection gap** once regression's predicted return is taken off (r52, r55). It is about a gap between a site and its system, in one period. - **The movement**: a separate crude **before-and-after** improvement of 1.5 points she was holding alongside, where **at least 18 and at most 21 percent of that movement is consistent with her mortality having risen** (r53, her section 4), because a patient who dies cannot be readmitted and the survivor denominator moves in the direction that flatters. It is about one site across two periods. Different receipts, different periods, different denominators. A sentence saying "about 1.5 points" without saying which one is a sentence a reader will carry into the wrong argument. **And step two is where the screen earns its keep in the other direction.** Her cardiac rehabilitation candidate clears it: all-cause hospitalisation at RR 0.58, 95 percent CI 0.43 to 0.77, with `crosses_no_effect` false, and beside it mortality at RR 0.87, CI 0.73 to 1.04, with `crosses_no_effect` **true** (r50, r51, her section 3.1, Dibben 2021, PMID 34741536). **The second row is the one a vendor would not show, and the screen shows both.** A candidate whose every row reads clean is a candidate somebody filtered. ## The refusal that is the deliverable *"What is an avoided readmission worth to us."* This is the seat's first question and this world cannot answer it, and the reason is not that nobody has packed the data. **An avoided admission is revenue lost under fee-for-service and revenue kept under risk, and which of those you are in is a term of your contract that this world has not read.** The direction of the answer flips on that term, not its size, so there is no conservative version to quote and no assumption to declare that would make it safe. **The tool for it is deferred permanently rather than pending**, which is a different fact from a gap in the corpus and should be reported as one. `world.toml`'s own `cannot_answer` list is the authority on what this world declines and it is where this refusal is recorded; read it there rather than from a copy in prose. So the refusal is the deliverable, and it is a stronger one than a number would have been. What it hands the reader is the question to take to the people who can settle it, in the order it has to be settled: 1. Which line of business this cohort sits in, and whether the contract for it is fee-for-service, shared savings, or full risk. 2. Under that contract, whether an avoided admission reduces revenue, is shared, or is kept. 3. What was actually paid for those admissions, which is an adjudicated claim and is held by the payer rather than by this world or by your own filing. Bring those back and the money path at step four becomes computable over **your** declared inputs, with `declared_inputs` carrying them and `savings_methodology` naming the class of saving you are then presenting. Bring back none of them and the correct sentence is that the clinical case stands on its own receipts and the financial case is unwritten. **Saying that first is what gets the second meeting.** ## The other walls this seat runs into, and where they are written Read them from `world.toml`'s `cannot_answer` list rather than from any restatement, because a copy of that list here would age on its own and the copy a reader meets first is this file. The shapes to expect, so you recognise one when it arrives: - No measure specification of any kind, so what a measure counts is never something this world can tell you, and that is the half a payment argument usually turns on. - No count of readmissions on any programme's definition, and the nearest published record here is a different numerator over a different period in a different population. - No causal effect outside the randomised records, because adjustment is not randomisation. - No evidence about any named device, test or vendor, because this world prices procedures. - No ranking of providers, and a per-facility ordering is the same refusal reached by a finer route. Each of those refusals names what would settle it. **Treat that clause as a work order**, the way `hammer-building-the-evidence` describes: it is usually the cheapest thing in the answer. ## Every answer built on this world closes with a claim report Its shape is defined in `hammer-worlds` and it is not optional. Build the claim list before writing prose, check it, then write only what came back entailed. Two things bite this seat in particular: a bound has to be claimed as its two endpoints rather than as a point, and a threshold you declared is your input rather than this world's finding, so it belongs under `asserted without a check` alongside the decision you made with it. HTTP bindings for this session: Explicit source-reference bindings: SKILL.md: https://space.hammer.ai/skills/global/building-sourced-excel-models.skill.txt?ledger=eccbd49b235df851c1880b208d9d33fb&snapshot=7a53af6a9beb58928d4a64dc92db09c9db65433f9909f1a10d16e302adaf86ba analysis.md: https://space.hammer.ai/skills/global/analysis.skill.txt?ledger=eccbd49b235df851c1880b208d9d33fb&snapshot=7a53af6a9beb58928d4a64dc92db09c9db65433f9909f1a10d16e302adaf86ba coverage-website.md: https://space.hammer.ai/skills/global/coverage-website.skill.txt?ledger=eccbd49b235df851c1880b208d9d33fb&snapshot=7a53af6a9beb58928d4a64dc92db09c9db65433f9909f1a10d16e302adaf86ba rate_position: https://space.hammer.ai/analysis/rate_position/?ledger=eccbd49b235df851c1880b208d9d33fb&snapshot=7a53af6a9beb58928d4a64dc92db09c9db65433f9909f1a10d16e302adaf86ba rate_comparison: https://space.hammer.ai/analysis/rate_comparison/?ledger=eccbd49b235df851c1880b208d9d33fb&snapshot=7a53af6a9beb58928d4a64dc92db09c9db65433f9909f1a10d16e302adaf86ba stack_check: https://space.hammer.ai/analysis/stack_check/?ledger=eccbd49b235df851c1880b208d9d33fb&snapshot=7a53af6a9beb58928d4a64dc92db09c9db65433f9909f1a10d16e302adaf86ba screening_yield: https://space.hammer.ai/analysis/screening_yield/?ledger=eccbd49b235df851c1880b208d9d33fb&snapshot=7a53af6a9beb58928d4a64dc92db09c9db65433f9909f1a10d16e302adaf86ba deferred_versus_avoided: https://space.hammer.ai/analysis/deferred_versus_avoided/?ledger=eccbd49b235df851c1880b208d9d33fb&snapshot=7a53af6a9beb58928d4a64dc92db09c9db65433f9909f1a10d16e302adaf86ba interpretation_requirement: https://space.hammer.ai/analysis/interpretation_requirement/?ledger=eccbd49b235df851c1880b208d9d33fb&snapshot=7a53af6a9beb58928d4a64dc92db09c9db65433f9909f1a10d16e302adaf86ba cohort_figures: https://space.hammer.ai/analysis/cohort_figures/?ledger=eccbd49b235df851c1880b208d9d33fb&snapshot=7a53af6a9beb58928d4a64dc92db09c9db65433f9909f1a10d16e302adaf86ba comorbidity_edges: https://space.hammer.ai/analysis/comorbidity_edges/?ledger=eccbd49b235df851c1880b208d9d33fb&snapshot=7a53af6a9beb58928d4a64dc92db09c9db65433f9909f1a10d16e302adaf86ba fee_components: https://space.hammer.ai/analysis/fee_components/?ledger=eccbd49b235df851c1880b208d9d33fb&snapshot=7a53af6a9beb58928d4a64dc92db09c9db65433f9909f1a10d16e302adaf86ba filed_charge: https://space.hammer.ai/analysis/filed_charge/?ledger=eccbd49b235df851c1880b208d9d33fb&snapshot=7a53af6a9beb58928d4a64dc92db09c9db65433f9909f1a10d16e302adaf86ba hypertension_evidence: https://space.hammer.ai/analysis/hypertension_evidence/?ledger=eccbd49b235df851c1880b208d9d33fb&snapshot=7a53af6a9beb58928d4a64dc92db09c9db65433f9909f1a10d16e302adaf86ba prognostic_association: https://space.hammer.ai/analysis/prognostic_association/?ledger=eccbd49b235df851c1880b208d9d33fb&snapshot=7a53af6a9beb58928d4a64dc92db09c9db65433f9909f1a10d16e302adaf86ba treatment_effect: https://space.hammer.ai/analysis/treatment_effect/?ledger=eccbd49b235df851c1880b208d9d33fb&snapshot=7a53af6a9beb58928d4a64dc92db09c9db65433f9909f1a10d16e302adaf86ba regression_to_mean_expectation: https://space.hammer.ai/analysis/regression_to_mean_expectation/?ledger=eccbd49b235df851c1880b208d9d33fb&snapshot=7a53af6a9beb58928d4a64dc92db09c9db65433f9909f1a10d16e302adaf86ba competing_event_coupling: https://space.hammer.ai/analysis/competing_event_coupling/?ledger=eccbd49b235df851c1880b208d9d33fb&snapshot=7a53af6a9beb58928d4a64dc92db09c9db65433f9909f1a10d16e302adaf86ba withhold_earnback: https://space.hammer.ai/analysis/withhold_earnback/?ledger=eccbd49b235df851c1880b208d9d33fb&snapshot=7a53af6a9beb58928d4a64dc92db09c9db65433f9909f1a10d16e302adaf86ba tournament_position: https://space.hammer.ai/analysis/tournament_position/?ledger=eccbd49b235df851c1880b208d9d33fb&snapshot=7a53af6a9beb58928d4a64dc92db09c9db65433f9909f1a10d16e302adaf86ba measure_reliability: https://space.hammer.ai/analysis/measure_reliability/?ledger=eccbd49b235df851c1880b208d9d33fb&snapshot=7a53af6a9beb58928d4a64dc92db09c9db65433f9909f1a10d16e302adaf86ba declared_inputs: https://space.hammer.ai/analysis/declared_inputs/?ledger=eccbd49b235df851c1880b208d9d33fb&snapshot=7a53af6a9beb58928d4a64dc92db09c9db65433f9909f1a10d16e302adaf86ba cost_of_care_exposure: https://space.hammer.ai/analysis/cost_of_care_exposure/?ledger=eccbd49b235df851c1880b208d9d33fb&snapshot=7a53af6a9beb58928d4a64dc92db09c9db65433f9909f1a10d16e302adaf86ba savings_methodology: https://space.hammer.ai/analysis/savings_methodology/?ledger=eccbd49b235df851c1880b208d9d33fb&snapshot=7a53af6a9beb58928d4a64dc92db09c9db65433f9909f1a10d16e302adaf86ba hospital_variance: https://space.hammer.ai/analysis/hospital_variance/?ledger=eccbd49b235df851c1880b208d9d33fb&snapshot=7a53af6a9beb58928d4a64dc92db09c9db65433f9909f1a10d16e302adaf86ba filer_inventory: https://space.hammer.ai/analysis/filer_inventory/?ledger=eccbd49b235df851c1880b208d9d33fb&snapshot=7a53af6a9beb58928d4a64dc92db09c9db65433f9909f1a10d16e302adaf86ba candidate_screen: https://space.hammer.ai/analysis/candidate_screen/?ledger=eccbd49b235df851c1880b208d9d33fb&snapshot=7a53af6a9beb58928d4a64dc92db09c9db65433f9909f1a10d16e302adaf86ba frontier: https://space.hammer.ai/analysis/frontier/?ledger=eccbd49b235df851c1880b208d9d33fb&snapshot=7a53af6a9beb58928d4a64dc92db09c9db65433f9909f1a10d16e302adaf86ba judgement_check: https://space.hammer.ai/analysis/judgement_check/?ledger=eccbd49b235df851c1880b208d9d33fb&snapshot=7a53af6a9beb58928d4a64dc92db09c9db65433f9909f1a10d16e302adaf86ba claim_check: https://space.hammer.ai/analysis/claim_check/?ledger=eccbd49b235df851c1880b208d9d33fb&snapshot=7a53af6a9beb58928d4a64dc92db09c9db65433f9909f1a10d16e302adaf86ba deliver: https://space.hammer.ai/analysis/deliver/?ledger=eccbd49b235df851c1880b208d9d33fb&snapshot=7a53af6a9beb58928d4a64dc92db09c9db65433f9909f1a10d16e302adaf86ba session_trace: https://space.hammer.ai/analysis/session_trace/?ledger=eccbd49b235df851c1880b208d9d33fb&snapshot=7a53af6a9beb58928d4a64dc92db09c9db65433f9909f1a10d16e302adaf86ba progress_check: https://space.hammer.ai/analysis/progress_check/?ledger=eccbd49b235df851c1880b208d9d33fb&snapshot=7a53af6a9beb58928d4a64dc92db09c9db65433f9909f1a10d16e302adaf86ba model_advisory: https://space.hammer.ai/analysis/model_advisory/?ledger=eccbd49b235df851c1880b208d9d33fb&snapshot=7a53af6a9beb58928d4a64dc92db09c9db65433f9909f1a10d16e302adaf86ba knowledge_navigate: https://space.hammer.ai/analysis/knowledge_navigate/?ledger=eccbd49b235df851c1880b208d9d33fb&snapshot=7a53af6a9beb58928d4a64dc92db09c9db65433f9909f1a10d16e302adaf86ba simulate: https://space.hammer.ai/analysis/simulate/?ledger=eccbd49b235df851c1880b208d9d33fb&snapshot=7a53af6a9beb58928d4a64dc92db09c9db65433f9909f1a10d16e302adaf86ba sensitivity: https://space.hammer.ai/analysis/sensitivity/?ledger=eccbd49b235df851c1880b208d9d33fb&snapshot=7a53af6a9beb58928d4a64dc92db09c9db65433f9909f1a10d16e302adaf86ba interval_check: https://space.hammer.ai/analysis/interval_check/?ledger=eccbd49b235df851c1880b208d9d33fb&snapshot=7a53af6a9beb58928d4a64dc92db09c9db65433f9909f1a10d16e302adaf86ba decision_navigate: https://space.hammer.ai/analysis/decision_navigate/?ledger=eccbd49b235df851c1880b208d9d33fb&snapshot=7a53af6a9beb58928d4a64dc92db09c9db65433f9909f1a10d16e302adaf86ba network_infer: https://space.hammer.ai/analysis/network_infer/?ledger=eccbd49b235df851c1880b208d9d33fb&snapshot=7a53af6a9beb58928d4a64dc92db09c9db65433f9909f1a10d16e302adaf86ba hello_world: https://space.hammer.ai/analysis/hello_world/?ledger=eccbd49b235df851c1880b208d9d33fb&snapshot=7a53af6a9beb58928d4a64dc92db09c9db65433f9909f1a10d16e302adaf86ba say_world: https://space.hammer.ai/analysis/say_world/?ledger=eccbd49b235df851c1880b208d9d33fb&snapshot=7a53af6a9beb58928d4a64dc92db09c9db65433f9909f1a10d16e302adaf86ba Read the companion references here (these replace local file references in the source guide): Build a sourced analysis: https://space.hammer.ai/analysis.skill.txt?ledger=eccbd49b235df851c1880b208d9d33fb&snapshot=7a53af6a9beb58928d4a64dc92db09c9db65433f9909f1a10d16e302adaf86ba OData exploration and analytic cuts: https://space.hammer.ai/odata.skill.txt?ledger=eccbd49b235df851c1880b208d9d33fb&snapshot=7a53af6a9beb58928d4a64dc92db09c9db65433f9909f1a10d16e302adaf86ba Worked analysis reference: https://space.hammer.ai/skills/global/analysis.skill.txt?ledger=eccbd49b235df851c1880b208d9d33fb&snapshot=7a53af6a9beb58928d4a64dc92db09c9db65433f9909f1a10d16e302adaf86ba Building sourced Excel models: https://space.hammer.ai/skills/global/building-sourced-excel-models.skill.txt?ledger=eccbd49b235df851c1880b208d9d33fb&snapshot=7a53af6a9beb58928d4a64dc92db09c9db65433f9909f1a10d16e302adaf86ba Coverage website reference: https://space.hammer.ai/skills/global/coverage-website.skill.txt?ledger=eccbd49b235df851c1880b208d9d33fb&snapshot=7a53af6a9beb58928d4a64dc92db09c9db65433f9909f1a10d16e302adaf86ba Spine cohort and post-acute analysis: https://space.hammer.ai/skills/spine-analysis.skill.txt?ledger=eccbd49b235df851c1880b208d9d33fb&snapshot=7a53af6a9beb58928d4a64dc92db09c9db65433f9909f1a10d16e302adaf86ba hammer-building-the-evidence: https://space.hammer.ai/skills/world/hammer-building-the-evidence.skill.txt?ledger=eccbd49b235df851c1880b208d9d33fb&snapshot=7a53af6a9beb58928d4a64dc92db09c9db65433f9909f1a10d16e302adaf86ba hammer-cost-of-care-case: https://space.hammer.ai/skills/world/hammer-cost-of-care-case.skill.txt?ledger=eccbd49b235df851c1880b208d9d33fb&snapshot=7a53af6a9beb58928d4a64dc92db09c9db65433f9909f1a10d16e302adaf86ba hammer-coverage-navigate: https://space.hammer.ai/skills/world/hammer-coverage-navigate.skill.txt?ledger=eccbd49b235df851c1880b208d9d33fb&snapshot=7a53af6a9beb58928d4a64dc92db09c9db65433f9909f1a10d16e302adaf86ba hammer-provider-contracting: https://space.hammer.ai/skills/world/hammer-provider-contracting.skill.txt?ledger=eccbd49b235df851c1880b208d9d33fb&snapshot=7a53af6a9beb58928d4a64dc92db09c9db65433f9909f1a10d16e302adaf86ba hammer-provider-quality: https://space.hammer.ai/skills/world/hammer-provider-quality.skill.txt?ledger=eccbd49b235df851c1880b208d9d33fb&snapshot=7a53af6a9beb58928d4a64dc92db09c9db65433f9909f1a10d16e302adaf86ba hammer-savings-classification: https://space.hammer.ai/skills/world/hammer-savings-classification.skill.txt?ledger=eccbd49b235df851c1880b208d9d33fb&snapshot=7a53af6a9beb58928d4a64dc92db09c9db65433f9909f1a10d16e302adaf86ba hammer-vendor-evidence: https://space.hammer.ai/skills/world/hammer-vendor-evidence.skill.txt?ledger=eccbd49b235df851c1880b208d9d33fb&snapshot=7a53af6a9beb58928d4a64dc92db09c9db65433f9909f1a10d16e302adaf86ba hammer-writing-the-deliverable: https://space.hammer.ai/skills/world/hammer-writing-the-deliverable.skill.txt?ledger=eccbd49b235df851c1880b208d9d33fb&snapshot=7a53af6a9beb58928d4a64dc92db09c9db65433f9909f1a10d16e302adaf86ba Data: https://space.hammer.ai/data/?ledger=eccbd49b235df851c1880b208d9d33fb&snapshot=7a53af6a9beb58928d4a64dc92db09c9db65433f9909f1a10d16e302adaf86ba Transcript: https://space.hammer.ai/session/?ledger=eccbd49b235df851c1880b208d9d33fb&snapshot=7a53af6a9beb58928d4a64dc92db09c9db65433f9909f1a10d16e302adaf86ba