# Triaging harms: which the kernel can touch, and which it must refuse **What this page is for:** it describes how we choose which real-world wrong to test against the protocol next — and, one day, how a machine could scan a whole archive of harms and hand back a ranked shortlist of the ones a checkable record would most cheaply shift. It borrows a habit called *triage* — invented in battlefield medicine and emergency nursing, borrowed from there by software engineering, and here returned to something near its original use: sorting social and legal harms the way a clinician sorts patients. It is written for litigators, campaigners, lawmakers, and anyone deciding where scarce effort should go.
Plainly Triage began in battlefield medicine: when you cannot treat everyone, sort by how urgent and how treatable, and name honestly whom you cannot help. Software borrowed it for bugs; we return it to public harms. The refusals matter as much as the picks — most wrongs are not ours to touch, and saying so is half the value. The scenario corpus answers a question one case at a time — *does a thin, checkable protocol help with this wrong, and where does it stop?* Each answer is expensive: a few hundred lines of researched, machine-validated argument. So the prior question — *which wrong next?* — deserves a cheap, honest, repeatable answer of its own. That is triage.
## The idea, borrowed twice The word comes from the French *trier*, to sort, and its method was worked out on battlefields and in emergency wards. When you cannot treat everyone at once you sort by two things — how urgent the case is, and how much your treatment would actually change its outcome — and then you do the hardest thing in the practice: you name, honestly, whom you cannot help. Software teams borrowed the habit for bugs. Turning it on legal and social harms brings it most of the way home. An engineering team faced with a thousand open bug reports does not work through them in the order they arrived. It **triages**: for each report it asks whether the bug is real and reproducible, whether it is in scope for this system at all, how much damage it does, and how cheap it is to fix — and only then decides what to pick up. The reports it *closes* — "not a bug", "won't fix", "out of scope" — matter as much as the ones it accepts, because a clear reason for closing saves everyone from re-litigating it. A democratic harm can be triaged the same way, and the discipline transfers almost unchanged. The one adaptation: our "won't fix" is a matter of principle, not laziness. A great many wrongs are real, serious, and simply *not ours to touch* — and the protocol is built to refuse them on purpose. ## The pipeline Four stages. A case that fails an early one never reaches the later ones. 1. **Intake, with a citation.** Restate the harm in one line, and name a real source for it. No case enters triage without a documented precedent — the same rule the scenario schema enforces ("no attack without a cited precedent"), applied at the door. 2. **The scope sort — the load-bearing filter.** Ask one question: is the wrong a matter of *substantive judgment* (who is right, what is fair, what is true) or of *process integrity* (was the record made, was the rule cited, is the pattern visible)? - **Substantive → refused, out of scope**, with the reason stated. A child's best interests under Article 8, whether a claim is misinformation, whether a policy was wise — the kernel refuses to decide these (refusal 11), because a machine that adjudicated them would be a machine worth capturing. In the clinical metaphor these are the cases triage cannot save — and naming them, rather than pretending otherwise, is the oldest discipline in the practice. This is not a gap to apologise for; it is the design. - **Process integrity → accepted.** On to prioritise. 3. **The signature check.** Every case that has ever *worked* shared a shape. Does this one carry it? (The catalogue is below.) If it does, the same cheap lever — a contemporaneous, checkable record — is a candidate. 4. **Priority — reach × leverage: the clinician's calculus.** Severity and treatability — how bad is it, and how much would a checkable record actually change the outcome. - **Reach**: how many people, how much public money, how central to the democratic function (1 low – 3 high). - **Leverage**: how much a cheap, checkable record would shift the *cost* of the wrong (1–3). Note the word: *cost*, not *existence*. A record rarely cures a wrong; it makes it visible in time, destroys deniability, and feeds the courts and elections that can. Where even that changes little, leverage is low, and the score says so. ## The recurring signature The wrongs the protocol can touch are not a random set. Across the corpus they keep the same few fingerprints — the "bug categories" of public harm: - **`pattern_of_lawful_acts`** — the wrong lives in an aggregate of individually-lawful acts (a procurement lane, a stop-and-search pattern, a regulator's forbearance). Unreachable one act at a time; visible only in the aggregate. - **`record_absent`** — the act was never recorded contemporaneously, so accountability means reconstructing it, late and by force. - **`record_destroyed`** — the record was made, then destroyed at source. - **`access_gated`** — the record exists, but access is by-permission and can be slow-walked until the answer no longer matters. - **`inaction_no_artifact`** — the wrong is a *non-decision*, which produces nothing to appeal or review. - **`documentary_gate`** — a document requirement excludes people from a right. ## What the scorer will and will not do This is the part that keeps triage honest, and it is worth stating plainly to a non-engineering reader. - **It is a scout, not a judge.** It ranks; a person decides. This is not timidity — it is the kernel's own first principle. A protocol that refuses to let a machine adjudicate substance cannot then build a machine that adjudicates which injustices deserve attention. The scout flags "this wrong has the shape of one a checkable record could reach — a human should look." Nothing more. - **Every score shows its inputs.** There is no hidden weighting: a score is `reach × leverage`, both set by a person's reading and both on the page. The same discipline the kernel demands of a form field — every demand cites its rule — applied to the triage itself. - **Legibility is a lever, not a cure.** Cases marked ⚠ are ones where making the wrong visible changed little: the court voided the procurement lane and returned nothing; the watchdog found the regulator failed and the river stayed dirty. They still rank — deniability destroyed is worth something — but the caveat is shown, never buried. - **Precedents are verified at encode, not at triage.** Triage runs on institution-level leads and is deliberately cheap and provisional — like confirming a bug reproduces only when you pick it up, not when you file it. The full cited-precedent rigour happens later, when a candidate is promoted to a full scenario. ## The current run Auto-filled by `python3 tools/triage.py --write` — regenerate it rather than editing by hand; the block below is generated from `triage/cases.json`. The ranking doubles as a check on the rubric, not just an output: the encoded cases land where their verdicts predict — the ones that *break* sink to the bottom on leverage, the rubric admitting it — and the strongest candidates surface at the top by the process, not by an author's hunch. _Generated by `tools/triage.py` over `triage/cases.json` (29 cases: 22 accepted, 7 refused)._ ### Accepted — ranked by leverage | # | case | score | reach × leverage | signature | status | |--:|------|:-----:|:----------------:|-----------|--------| | 1 | foi-slow-walking | 9 | 3 × 3 | access_gated, record_absent | **encoded** (`foi-slow-walking`) | | 2 | horizon-unauditable-evidence | 9 | 3 × 3 | record_absent, pattern_of_lawful_acts | **encoded** (`horizon-unauditable-evidence`) | | 3 | party-finance-dark-money | 9 | 3 × 3 | pattern_of_lawful_acts, record_absent | **encoded** (`party-finance-dark-money`) | | 4 | revolving-door-lobbying ⚠ legible-not-remedied | 9 | 3 × 3 | pattern_of_lawful_acts, record_absent | **encoded** (`revolving-door-lobbying`) | | 5 | automated-welfare-reasons | 6 | 3 × 2 | record_absent | **encoded** (`automated-welfare-reasons`) | | 6 | exam-algorithm-downgrade ⚠ legible-not-remedied | 6 | 3 × 2 | record_absent, pattern_of_lawful_acts | **encoded** (`exam-algorithm-downgrade`) | | 7 | grenfell-warnings-ignored ⚠ legible-not-remedied | 6 | 3 × 2 | inaction_no_artifact, record_absent | **encoded** (`grenfell-warnings-ignored`) | | 8 | nhs-data-access-opaque | 6 | 3 × 2 | access_gated, record_absent | **encoded** (`nhs-data-access-opaque`) | | 9 | procurement-vip-lane ⚠ legible-not-remedied | 6 | 3 × 2 | pattern_of_lawful_acts, record_absent | **encoded** (`procurement-vip-lane`) | | 10 | regulator-inaction ⚠ legible-not-remedied | 6 | 3 × 2 | pattern_of_lawful_acts, inaction_no_artifact | **encoded** (`regulator-inaction`) | | 11 | citizenship-deprivation-basis ⚠ legible-not-remedied | 4 | 2 × 2 | record_absent | **encoded** (`citizenship-deprivation-basis`) | | 12 | criminal-disclosure-failure ⚠ legible-not-remedied | 4 | 2 × 2 | access_gated, record_absent | **encoded** (`criminal-disclosure-failure`) | | 13 | facial-recognition-no-basis | 4 | 2 × 2 | record_absent | **encoded** (`facial-recognition-no-basis`) | | 14 | ministerial-standards-unenforced ⚠ legible-not-remedied | 4 | 2 × 2 | inaction_no_artifact | **encoded** (`ministerial-standards-unenforced`) | | 15 | ofsted-single-word-judgment ⚠ legible-not-remedied | 4 | 2 × 2 | record_absent | **encoded** (`ofsted-single-word-judgment`) | | 16 | protest-conditions-no-basis | 4 | 2 × 2 | record_absent | **encoded** (`protest-conditions-no-basis`) | | 17 | removal-without-legible-warrant | 4 | 2 × 2 | record_absent | **encoded** (`removal-without-legible-warrant`) | | 18 | send-ehcp-refusals ⚠ legible-not-remedied | 4 | 2 × 2 | pattern_of_lawful_acts, inaction_no_artifact | **encoded** (`send-ehcp-refusals`) | | 19 | voter-id | 4 | 2 × 2 | documentary_gate | **encoded** (`voter-id`) | | 20 | drc-medical-detention ⚠ legible-not-remedied | 3 | 3 × 1 | record_absent | **candidate** | | 21 | government-by-whatsapp ⚠ legible-not-remedied | 3 | 3 × 1 | record_destroyed | **encoded** (`government-by-whatsapp`) | | 22 | infected-blood-coverup ⚠ legible-not-remedied | 3 | 3 × 1 | record_destroyed | **candidate** | ### Recommended to encode next The highest-leverage cases not yet built as scenarios: - **drc-medical-detention** (score 3) — Hospitals detain medically-cleared patients — and hold the bodies of the deceased — until unpaid bills are settled, with no lawful warrant and no route to contest the confinement. _Lead: Transnational Justice Initiative / 9BR Chambers; UN Working Group Against Arbitrary Detention (DRC debt-detention, 2025); via CrowdJustice — non-UK_ - **infected-blood-coverup** (score 3) — Patients are given contaminated blood, not told for years, and the documents that would prove it are lost or destroyed. _Lead: Infected Blood Inquiry (Langstaff, final report 2024)_ ### Refused — out of scope Not defects a checkable record can touch; each is closed with its reason. | case | why refused | |------|-------------| | child-surname-article8 | A substantive judgment about a child's best interests under Article 8 — who is right. The kernel refuses to decide substance (refusal 11); this belongs to a court, and the protocol has no business weighing it. | | doctor-misinformation-gmc | Deciding whether a claim is misinformation is a judgment of truth — a ministry of truth is exactly what the kernel refuses to build (T11, refusal 11). Whether the GMC then acts is its substantive professional judgment. | | policy-wisdom | Whether a policy is wise is a political judgment, not a process defect. There is no record to make, rule to cite, or pattern to surface — only a disagreement, which elections and argument settle, not infrastructure. | | planning-merits | The merits of a planning decision are substantive and political — whether this was the right call — for the planning system and the courts, not a process defect a record can settle. The kernel can make the decision and its reasons legible; it cannot decide whether the decision was right. | | assisted-dying-question | A substantive moral and political question the kernel refuses on principle (refusal 11). It can run a fair, verifiable, receipt-free vote on the matter and hold the rights floor around it; it must never decide the answer — the line that keeps it from becoming a machine worth capturing. | | clinical-best-interests | The best interests of a critically ill patient are a substantive medical and ethical judgment for clinicians and a court, weighing evidence the protocol cannot and must not adjudicate. The kernel can make the decision and its reasons legible; it has no business deciding whether treatment should continue. | | statue-removal | Whether a monument should stand is a substantive question of values and history a community must argue and decide, not a process defect. The kernel can run a fair, verifiable vote on it and record the outcome; it must never decide the answer. | ## Running it, and adding a case ``` python3 tools/triage.py # print this report python3 tools/triage.py --check # validate the corpus only ``` To add a case, append it to `triage/cases.json` with its harm, a lead, the scope (`onstage`/`offstage`), any signatures, and — for onstage cases — a `reach` and `leverage` from 1 to 3. Offstage cases must state why they are refused. That is the whole method: cheap enough to run over a large archive, honest enough that the refusals are on the page next to the picks.