RM170 — Kleene shape: None stays a neighbour, the warning is the product¶

Subject: the authoring shape for ROADMAP.md § RM170. Measurements: CONTRADICTION_CORPORA. Every number this note leans on was measured there; this file does not re-derive them. Date: 2026-09-02.

This is a design record, not a second measurement and not a contract. The live item stays in ROADMAP. If a sentence here and the entry disagree, the entry wins until someone edits it.

The decision this note records. Extend the camp test so a withhold (None) is visible. Do not extend VALID_DIRECTIONS. Do not sum camps into a point. Two enricher owners — civic_draft and a new verification check — emit a warning. The author seeing it and proceeding is the expected close, not a workaround.

1. What this is answering¶

RM170's defect is not "we lack a combined direction." It is: an author can write risk over a variant a source also rebuts, and every gate stays green.

contested_variants cannot see that pair. It counts a variant whose camps include both risk and protective. A Does Not Support row is mapped to None and then skipped, so it never enters a camp. That is why the counter is 0 on both bases, correctly, and why the three VHL variants are invisible to it.

A vector sum (reversal as true→false, refute as true→none, final point = the sum) classifies the two fights honestly and then answers the wrong question. A sum asks what should the cell become? RM170 asked for a finding. Writing the point would reopen two refusals: enricher checks report and never repair, and RM152 refused filling direction from source disagreement.

House three-valued algebra is Kleene over answers (exists, covered, licence known), not a metric on signs. risk ∧ withhold does not become unknown on the authored cell. It becomes a finding beside an unchanged cell.

2. The None algebra¶

A snapshot variant is a set, not a score. Three slots, independently filled:

slot	who fills it	values
camps	every `Supports` row	`∅`, `{risk}`, `{protective}`, `{risk, protective}`
withhold	any `Does Not Support` row	absent, or present with `evidence_status` and the evidence ids
authored	the module row, if any	`risk` / `protective` / `unknown` / `contested` / absent

Rules:

A withhold is not a camp. It does not enter {risk, protective}. That is why today's contested_variants is correctly 0.
A withhold is not protective. None is withheld, never negated (@refutation-withholds).
Combining never picks a point. You publish the set. gene_validity already does this one grain up: an unorderable [definitive, refuted] group is a set, not a verdict.
Kleene stays on answers, not on signs. The authored direction is not rewritten.
Status is part of the withhold, not a weight. accepted and submitted are not unit vectors. A hint that omits status_basis is false: every assert-and-refute pair in CIViC rests on a submitted rebuttal.

Snapshot readings, before any authored row:

camps	withhold	name	what it is
one pole	no	settled	source has a sign
both poles	no	reversal	`contested` — genuine opposition. Still 0 in CIViC
one pole	yes	refute-beside-claim	RM170. A sign sitting next to a withhold
both poles	yes	reversal + withhold	both findings. Unobserved today
empty	yes	refute-only	CHEK2 788, TP53 4968. A withhold with no claim beside it
empty	no	nobody-asked	not on this axis

The authored row is a second set. The finding is the join of the two, not a reduction of either.

3. Per-case advocacy¶

Each case is a different sentence. One code per sentence, or the author cannot tell what to do.

A. Settled source, author agrees¶

Snapshot {risk}, no withhold. Author writes risk.

Solves: silence is correct. No warning. That is a pass, not a skip.

How: camps match; withhold absent.

B. Settled source, author disagrees¶

Snapshot {risk}, no withhold. Author writes protective (or the reverse).

Solves: ordinary source-versus-author, the clin_sig shape. Not RM170. Do not steal this into the muddy-water code.

How: a later, separate check if CIViC concordance is ever wanted. RM152 refused filling direction from this. Leave it parked. The author may outrank a settled source; that is what record_override is for.

C. Reversal (`contested`)¶

Snapshot {risk, protective}.

Solves: the fight contested_variants already counts. The drafter already names those subjects.

How: keep that line. Do not merge it with D. contested stays on VALID_DIRECTIONS as the authored member for "I am saying the sources disagree about the sign." The snapshot statistic is not that cell.

D. Refute-beside-claim — the RM170 case¶

Snapshot {risk} + withhold. Author writes risk (the usual draft).

Solves: "you are shipping a sign the source also rebutted, and nothing said so."

How:

Drafter writes the Supports row. Does not write a protective row. Emits a named warning: variant, evidence ids on both sides, each status, status_basis.
Checker fires on that authored row later, same facts, attested.
Cell stays risk. Finding stays.

This is VHL G104V: accepted claim 7134 + submitted refute 10949.

E. Refute-only, author never writes¶

Snapshot ∅ + withhold. No authored row.

Solves: do not invent a subject. CHEK2 788 is not muddy water about a claim; there is no claim.

How: drafter writes nothing (already). Checker: not a subject. subjects does not count it. A release-record count of refute-only variants may live on the snapshot; it is not a module finding.

F. Refute-only, author then writes a sign¶

Snapshot ∅ + withhold. Author writes risk.

Solves: a different sentence — "you asserted a sign the source has only ever rebutted." Author-versus-refute, not source-versus-itself.

How: own warning code. Same severity (warn, never gate). Same owners. Do not call it D or the subject count silently mixes the three VHL variants with the two claimless ones. Name it: refute-only, no supporting item on this basis.

G. Combination profile (EID 8721)¶

One evidence item, two variants (VHL S183L and VHL D126N).

Solves: the parquet's lie that this is two single-variant rebuttals. _submitted_evidence_row stamps molecular_profile_id from the variant's single-variant profile, so the item is written as two rows claiming MP 2037 and 2406.

How: the finding keys on evidence_id, then fans out to authored rows. One finding, two subjects. Text says "combination genotype, profile 5278." Do not wait for RM174 to ship the check — but do not count "3 variants, 3 independent refutes." Honest count: one single-variant pair (2428) + one combo item touching two loci.

H. Accepted-only snapshot¶

Subject count 0. The class does not exist on this basis.

Solves: silence that is a measured empty set, not a skip.

How: record ran(subjects=N, findings=0) and detail names status_basis=accepted. --offline / no snapshot is skipped, and that is a different row.

I. STRchive `Refuted` + a pathogenic band¶

Not this algebra. One field grades the locus–disease association; the same record still publishes a pathogenic band the drafter writes. Grade versus bands, no direction slot.

Solves nothing in this check. Park it. If evidence is adopted later: draft-repeats and check-repeat-bands grow a sibling warning. Different code, different subject, different skill line. Deciding "one check" for RM170 is deciding whether that field is adopted at all; the CIViC sign does not need that answer.

J. Author sees the warning and keeps the row¶

Solves: the expected close. Muddy water is a fact about the source, not an authoring error.

How: cell unchanged. Warning remains — an outranked source never goes green, the same rule as clin_sig. Optional record_override writes why they kept the sign. Closure binds the bytes. verification.json says the question was put.

4. Two owners, both enricher¶

owner	stage	first sighting for	writes	must not
`civic_draft`	2	anyone drafting from CIViC	the Supports row; a named warning on D / G / (F if they later add)	write the opposite sign; swallow the variant into an aggregate `N rows`
new verification check	5, and inside `enrich()` when a civic snapshot is present (same as `clin_sig`)	hand-authors and ClinVar-drafted modules that never touched `civic_draft`	`verification.json` record + findings with `detail`	repair `direction`; escalate under `--strict`

just-dna-compiler validate / compile is not an owner. Format and compiler never fetch and never hold a CIViC snapshot (Principle 2). Nothing this snapshot produces may enter resolution.csv. Once the check has attested, compile restates the findings as warnings (verification_findings_recorded, S70). That is the echo at stage 6, not the first sign on the road.

close is not an owner. It already refuses only an invalid spec. Warnings do not block a close. That is the line that makes "see it and proceed" a designed outcome.

Severity is settled in the entry: warn in both modes, never gate. Neighbours: @clinsig-never-escalates, @a-source-recuring-is-not-a-strict-matter. A source disagreeing with itself is not an authoring error.

The vocabulary for the finding is VALID_VERIFICATION_CHECKS / VALID_WARNING_CODES, not VALID_DIRECTIONS. RM150 added contested beside unknown so an absence and a finding stopped sharing a member. A refute-beside-claim is another shade; putting it on direction is the old state overload (Principle 5).

5. When the author sees the sign¶

A road sign, not a barrier. Same sign, three posts. Missing a post is a skip, not a pass.

 2 draft ──► 3 curate ──► 4 enrich ──► 5 check ──► 6 validate/compile ──► 6b close
    │                        │            │                │                 │
    │ named warning          │            │ attested       │ echo            │ does not refuse
    │ if civic_draft         │            │ finding        │ of stage 5      │
    │                        │            │                │                 │
    └──── first sighting ────┴────────────┴── guaranteed ──┴── reminder ─────┴── proceed
          for CIViC drafts                    if snapshot
                                              was available

Stage 2 — first post, CIViC drafts only. Today: "N CIViC row(s) refute a predisposition claim… none was drafted." No variant named. Needed: the contested_variant treatment — name the variant, both evidence ids, statuses, basis. The Supports row is still written. The sign says: you are about to curate a risk row that the same snapshot also rebuts.

They continue to stage 3. That is expected. They fill genotype, conclusion, maybe keep direction=risk. The warning does not delete the row.

Stage 3 — no new post. Curation does not fetch CIViC. Do not hide the finding in hints. If they never drafted from CIViC, they have not seen the sign yet. Stage 5 exists for them.

Stage 4/5 — the guaranteed post. Fold the check into enrich() when a civic snapshot is on disk or configured, exactly like clin_sig. Anyone who enriches a germline-direction module against that snapshot sees a paragraph of this shape:

Variant 2428 (VHL G104V): authored direction=risk. CIViC has supporting EID 7134 (accepted) and Does-Not-Support EID 10949 (submitted). Snapshot basis accepted+submitted. A refutation does not establish the opposite sign; this row was not changed.

--strict still compiles. The record is ran, findings ≥ 1, never skipped.

If there is no snapshot, or --offline: skipped, reason named. Not a pass. The skill line is: you have not been told the water is clear; you have been told nobody looked.

Stage 6 — the reminder. validate / compile reprint the attested findings. Green --strict is expected. The warning sits in the list an author is supposed to read on a green run, the same place clin_sig mismatches already live.

Stage 6b — proceed. close does not care. Closure means a person declared these bytes final knowing what the checks said, not that the water was clear. A closed module with this finding is a checked module that chose to keep a sign. A closed module with skipped is an unchecked module. Those must stay distinguishable in verification.json.

After close. Edit the row and the closure drops; a check record bound to the old bytes drops with it. Re-check, re-close. Pass 2 re-enters at 3 or 5, not at 0.

They do not see a flipped direction, a dropped row, a <<REPLACE>>, or a close refusal. Those would turn an expected proceed into a fight with the tool.

6. Path × edge case¶

How the row got there	Snapshot	What they see, when	If they proceed
`civic_draft` VHL G104V	`accepted+submitted`	2: named D warning, risk row written. 5: same finding attested. 6: echo	close ok; record stays
`civic_draft` VHL S183L / D126N	wider	2 and 5: one G finding, two rows, text says combination / EID 8721	close ok
`civic_draft` on `accepted` only	accepted	2: no D (class is empty). 5: `ran, findings=0`, basis named	close ok; silence is measured
Hand-author `risk` on G104V	wider	2: nothing. 5: first sighting. 6: echo	same as draft path from 5 on
ClinVar-drafted VHL panel	wider	same as hand-author — `draft-panel` is not `civic_draft`	same
Hand-author `risk` on CHEK2 788	either	5: F, not D	close ok
Author writes nothing on 788	either	nothing. E. Not a subject	—
No civic snapshot / offline	—	5: `skipped`. Stage 6 does not invent a finding	close ok; unchecked on this axis
They `record_override` "I've read both papers"	wider	warning remains. Override is the reason, not a mute	close ok; next reader sees both
Re-draft after CIViC accepts 10949	later release	2: `differs` / new named warning; 5: still D, status may now be accepted	same; currency is a different check
`draft-repeats --gene DMD`	STRchive	not this sign. Today: nothing. Later: sibling warning on bands	out of scope here
Close without ever enriching	—	no sign at all	legal today; the missing record is the honest state. Do not make close require this check (RM73's gate half, major, blocked)

Proceeding after the sign is the clin_sig contract, copied on purpose: two opinions in the world, the format does not arbitrate, --strict is reproducibility not truth, the finding carries enough to weigh (here: evidence ids + statuses + basis; there: review stars).

7. What the author is looking at¶

Not a new screen. Three existing surfaces, one sentence each.

Draft / enrich / check CLI warning — the named paragraph in §5. This is the muddy-water sign. It is a warning. The command exits 0 in both modes.
verification.json — a new VALID_VERIFICATION_CHECKS member, subjects, findings, detail listing variant keys + EIDs + statuses + basis. skipped if it could not run.
Compile warning list — restatement, so a green strict compile still shows it.

review_queue stays for overrides. Do not dump check findings there unless picking up the S70 sidecar (RM130). Not required to close RM170.

8. What next¶

Do the CIViC half. Do not wait on STRchive or RM174.

Draft line first. Cheapest, covers the path that created the item. Change refutation_states_no_direction to name variants the way contested_variant already does. Include evidence ids, statuses, basis. Pin the phrase. No schema change.
Check second. New VALID_VERIFICATION_CHECKS member. Subject = authored row whose snapshot variant is D or F. Warn both modes. Attest. Fold into enrich() when a civic snapshot is available so a hand-author cannot miss it by only running enrich and not a new command. detail is mandatory: a count with no names is not actionable (S70).
Key findings on evidence_id, then fan out to rows. So G is one finding, two subjects, even while RM174 still mis-stamps the profile column.
Compile echo. Restated warnings from the record. Already the pattern; wire this member into the same list.
Tests from the probe's three variants plus 788. D fires on 2428; G is one finding not two; E does not fire; F fires only with an authored sign; accepted-only is findings=0 with basis in detail; no snapshot is skipped.
Park STRchive as a sibling, or a line under RM165, not a second axis of this check.
RM174 in parallel. Repair 2 (keep the fan-out, publish the real profile). Repair 1 would delete two of the three subjects and make the check look barren.

Not next: new VALID_DIRECTIONS members, a draft filter, a close gate, compiler-side validation, vector sums, adopting STRchive evidence.

The algebra is no longer the block. The remaining block on a general check is STRchive — and that answer is not needed to ship the CIViC sign. Ship the sign. Let "see it and proceed" be the documented happy path.