Trust the process, not the designer

Coordinated entry has been burned before by tools that claimed objectivity, produced opaque numbers, and were later shown to predict poorly. You should not have to trust whoever built HEAT. You should be able to inspect how decisions are made, how mistakes are corrected, and how the instrument changes over time. That is the standard used in medicine, aviation, and safety-critical engineering, and it is the standard this page holds HEAT to.

Public beta in development. Not for production decision making.

Where the line is

HEAT establishes facts. Your written standards decide the order. No vulnerability score, no rank, and no queue position exists anywhere in the tool. When one opening has six eligible households, the CoC's adopted written standards apply these facts (24 CFR 578.7(a)(9); HUD Notice CPD-17-01 II.B.3 requires prioritization to follow documented, publicly available policy).
Nothing here can screen anyone out. Too little or no income, substance use, domestic violence history, service resistance, evictions or credit, and criminal record can never lower anyone's standing, and an automated test proves it on every change (CPD-17-01 II.B.4).
Declining is never read as "no". Every question takes "I do not know" and "I prefer not to answer". A missing answer is reported as missing, named as declined, unknown, or not asked, and no determination is lowered by nondisclosure (CPD-17-01 II.B.11).
Who answered is recorded on every result. HEAT runs in three modes: a person answering for themselves, an assessor working with the person present, or someone answering on their behalf. Question wording follows the mode, and the mode is printed on the result and carried in the record. What runs on this site is a public preview anyone can open. It is not a coordinated entry assessment of record for any community, and no community has adopted it.

Two statements we will defend

On how HEAT is built: HEAT was developed using modern software engineering tools, but every determination is produced by deterministic rules that print themselves, rule by rule, on the result they produce, and can be independently reproduced. No AI, model, or hidden weight participates in any determination. Identical answers always produce identical results.

On bias: We do not claim HEAT is unbiased; no instrument can honestly claim that. We designed it to minimize avoidable bias, we publish every rule, we test for unintended disparities during pilots, and we change the instrument when evidence shows a better approach. A facially neutral fact like emergency-room visits can become a proxy where care access differs by community; that risk is named in our Evidence Register and tested in our validation plan, not hidden.

Color, and five words that stay separate

Two commitments below are not preferences and are not open to being traded for a better looking screen. They are stated here so that a later version of HEAT can be held to them, including by us.

On color: Color describes the state of the record and the workflow, never the value of the person.

On the five words: Need, eligibility, priority, availability, and allocation are separate determinations and must remain visually and logically distinguishable.

The color vocabulary

Five families, one meaning each, applied the same way on every screen. A color is a sentence about the record, so a color with two meanings is a sentence nobody can read.

Red: act now. Something is unsafe right now. On a result, red is reserved for safety and is used for nothing else.
Amber: look here. The record is incomplete, conflicting, or waiting on something. It is an instruction about paperwork and circumstance, never about the person, and it may never land on a determination.
Green: ready. A requirement of the workflow is satisfied, or a step is settled and complete.
Cyan: information. A stated fact or a determination. Every value a determination can take is drawn identically, so no value is highlighted over another.
Neutral: no active status. Context that changes nothing on its own, and is not asking anyone to do anything.

Every color carries a glyph and a word beside it. Nothing on the screen reports a state in color alone, so a reader who cannot see the difference reads exactly the same record.

The five words

Need. What intervention appears appropriate.
Eligibility. Which programs can serve this household.
Priority. How the community's adopted policy orders access when resources are scarce.
Availability. What can actually be accessed now.
Allocation. What was ultimately assigned.

HEAT today establishes need, and the facts that eligibility is decided on. Priority, availability, and allocation belong to the community: its adopted written standards set the order, its inventory sets what is open tonight, and its staff decide what is assigned. Merging any two of these into one indicator is how a tool begins making policy without ever saying that it did.

Written down before we need it

If HEAT ever displays a community's priority groups, all groups will receive identical visual treatment. Priority will never be traffic-light colored, because a color scale on priority is a score, whatever the labels say. We are stating this while HEAT has no priority display at all, which is the only point at which the promise costs us nothing and is therefore worth making. The same rule is already mechanical for the determinations HEAT does make: a test in the automated suite enumerates every value each determination can take and fails the build if any one value is drawn differently from the others.

The standing controls

Everything versioned, nothing edited in place. Questions, rules, and thresholds change only by new version; every result records the versions that produced it, so any determination can be reconstructed later.
Every determination explains itself. The full reasoning trace, rule by rule, is on every result. There is nothing to reverse-engineer because nothing is hidden.
Automated release gates. An automated suite runs on every change: golden scenarios, a practitioner-reviewable chronicity corpus, simulated clients, impossible users, thousands of randomized rule-interaction checks, and a suite proving the factors HUD forbids using to screen people out can never change any determination.
Reading level, gated in three tiers. Everything a person being assessed reads is held below a grade 8 reading level, everything a caseworker reads while using the tool is held below grade 12 with no sentence over 28 words, and reference pages like this one are left uncapped so they can say exactly what they mean; all three are checked by the same automated test on every change.
A public Changelog that includes our mistakes, with root causes, and an Evidence Register stating why each question exists, what was rejected, and what could go wrong. The repository carries a documented harm analysis for each output: which way a wrong answer fails, what that costs a person, and which error the design is built to minimize.
Structured disagreement. Professionals can file "I would have decided differently" from any result, and participants can change any answer at any time and see the result recompute. Every result page carries a route for saying the determination is wrong, with no dead ends: the CoC's own review contact and appeal process where a community has configured one, and the feedback channel built into the tool where it has not. Both routes land in a review queue.
A standing public challenge. Find a scenario where HEAT reaches the wrong determination; accepted cases join the permanent regression suite before any fix is written, credited by the name you choose to leave.
Planned for pilots: monthly non-punitive expert re-derivation of sampled assessments (the medical M&M model), outcome validation against a published baseline, and disparity monitoring with the pilot community's own data.

How feedback becomes change

Every flag gets a written disposition. Nothing is adopted silently and nothing is dismissed silently. A suggestion passes five tests before it becomes a question or a rule.

Construct fit. Does it belong to what HEAT measures, or is it policy that belongs to the CoC?
Necessity. Does it change a determination or drive a concrete action, or is it information for its own sake?
Question cost. The instrument has a hard cap of 36 questions, so an addition has to earn its place against everything already asked.
The screening-out rule. Anything added must be provably unable to lower anyone's standing (HUD CPD-17-01 II.B.4), enforced by an automated test.
Evidence and traceability. Adopted changes ship as new item versions with an Evidence Register entry naming the source; declined suggestions get their reasoning recorded too.

The outcomes are adopt, adapt, answer without changing, or defer to pilot data, and the Changelog shows which one every flag received.