/human-merge

Many reports in, one decidable surface out. Each input may be excellent. The pile is still undecidable, because the reader cannot tell which claims are corroborated, which rest on one source wearing four hats, and which two reports quietly disagree.

The complete style layer lives in human-output when that skill is installed. Otherwise apply the portable style floor, stopping-point rule, and checker contract in references/human-contract.md. This file carries only what exists because there is more than one report.

The four operations

These are the whole skill. Each has no single-input analogue, which is why nobody who wrote one of the inputs could have done it.

1. Collapse correlated reports to one source. Reports are not sources.
2. Name what is superseded — and whether a live decision already rests on the dead figure.
3. Classify each conflict before touching it. Most are not contradictions.
4. Pool coverage. A hole no single report had appears only at merge time.

Anything else you are tempted to add is human-output restated. Leave it out.

The order the work happens

The sections below are reference, not a running order. Executed top to bottom they would grade confidence before counting sources, which is the error this skill exists to prevent. Run these ten steps in order. Three of the dependencies are load-bearing and named where they bite.

1. Inventory. List every input: author, date, and the basis it states for each claim. Nothing else can start until this exists — every later step reads from it, and a report whose basis you never asked for gets graded generously by default.
2. Reconcile definitions. One word meaning two things manufactures both false agreement and false conflict. Before any comparison, or steps 3 and 6 both run on sand.
3. Same or different. Dedupe claims, so you are comparing findings rather than sentences.
4. Count sources, not reports. Run the five-key test per claim. Independence before confidence — the label in step 7 is capped by the number this step returns, so doing it later means relabelling everything.
5. Order the supersessions. Within one subject, attribute and scope only. Then ask what already rests on the superseded figure.
6. Classify each conflict. After steps 2 and 3, because most apparent conflicts dissolve into a definition or a scope.
7. Grade and label. Now, and only now, assign confirmed, judgement or estimate.
8. Combine, or park. The joint arithmetic, longhand; and the items that do not combine, each with what would make it combinable.
9. Pool the coverage. Walk the decision. Last, because the gap list has to catch anything steps 1 to 8 named without a figure.
10. Write it. Verdict, ledger, numeral trace.

The five relations

Every merge names each of these or says it does not apply. A merge that skips one is a summary, and a summary of six reports is a seventh report.

1. Superseded. Name the earlier report, its figure, the new figure, and the delta. Then the part everyone omits: say whether a decision already rests on the old figure. If one does, that is the verdict of the whole merge.

2. Same or different. Two reports naming one thing are not two findings. Say which it is. If they are one, drop the double count. If they are two, say whether fixing one fixes the other — two problems counted as one is how a fix gets under-sized.

3. Load-bearing. One report matters only because of another. A lease clause is dull until a payback period runs past it. Say which is the carrier and which is the load, and file the pair rather than the clause alone.

4. Together. The joint figure none of them states. Show the arithmetic from named inputs, longhand.

5. Does not combine. Different clock, different population, different unit, different confidence class. Park it, labelled, and say what would make it combinable. Forcing a fit is this skill's failure mode.

Independence: count sources, not reports

Two reports corroborate only if either could have been wrong while the other stayed right. Run the test on the claim, not on the report — one report can be independent on its first claim and derivative on its second.

Ask each report what its basis actually was, then match on five keys:

1. They quote the same document.
2. They interviewed the same person.
3. They inherited the same brief, or the same corpus.
4. One of them read the other.
5. They are sibling agents run over one shared context.

The fifth key is the trap in agent fan-out. Siblings over one context are correlated by construction: their agreement carries no confirming weight and looks identical to genuine corroboration.

Then apply the counterfactual, in words: could report A have been wrong while report B stayed right? If no, they are one source.

If you dispatched the workers, their briefs are your best independence evidence. Read supplied briefs before the reports; if unavailable, use the unknown-independence fallback below. Two workers sent the same context and the same question will agree, and the reports will not say so.

Self-merge. If you wrote some of the inputs, you are not independent of them. Say so in the ledger, and do not count your own report as corroborating your own conclusion.

The fallback, when independence cannot be resolved. Do not guess and do not count optimistically. Declare independence unknown, count those inputs as one source, cap the merged claim at estimate, and print the line:

independence not established

Unknown independence is not independence. That sentence is the rule.

Access caps the label — it is not a second ladder

There is one ladder: the contract's confirmed, judgement, estimate. How a reporter knew a thing does not get its own scale beside that one; it sets a ceiling on where the claim can land. Two ladders is decode work, and the reader should never hold both.

Print the basis in the same breath as the label — "(Confirmed; R5 inspected it)" — rather than in a legend the reader must scroll back to.

No numeric score. A single number invites averaging, and averaging is the failure this skill exists to prevent. A derived or modelled claim inherits its weakest parent's ceiling and adds nothing, so there is nothing left for a second scale to say.

As-of dates and scope boundaries travel with every value here as they do everywhere — that rule is human-output's, not restated. What is this file's: most false conflicts are two true claims about different scopes, which is why scope is a row in the conflict table below.

Merged confidence is a minimum, never an average

• Capped by the weakest input the conclusion actually rests on. Trace which inputs it rests on. Strong inputs on other questions do not lift it.
• Agreement between dependent inputs raises nothing. Only independent methods arriving at the same place raise the label a step.
• One unresolved reversing input caps the merge at estimate, however strong everything else is.
• Coverage is part of confidence. Unexamined is not the same as fine.

Use the contract's three labels: confirmed, judgement, estimate.

Three forbidden moves

• Never average two numbers that disagree. The mean of a right answer and a wrong one is a wrong answer with no author.
• Never drop the outlier. A lone dissent from the only Direct source is often the most valuable report in the pile.
• Never attribute a claim to "the reports" collectively. That launders one source's guess into the voice of all of them.

Classify the conflict before you touch it

Order of arrival is evidence only within the same subject, attribute and scope. Later beats earlier when a newer reading replaces an older one about the same thing. Across different scopes, arrival order is not vintage at all — it is a scope mismatch wearing a date, and treating it as supersession overwrites a true claim with an unrelated one.

A live conflict belongs in the verdict sentence. If the reader would act differently depending on which report is right, that is the finding.

Pool the coverage

Walk the decision, not the reports, and ask which parts no input addressed. Each author saw their own gap and none saw the union. Print the union. Silence is not assent: a report that was never asked about a topic is a gap, not a vote, and the two must stay apart in every count you print. For the same reason, never write that nobody has ever done a thing when what you have is nobody in this set reporting it — absence of a report is not evidence of absence, and the phrasing has to say which one you hold.

The verdict test, and it is a hard one. Every fact named in the verdict either carries a figure or appears in the gap list. No exceptions. A fact strong enough to reach the verdict sentence and weak enough to have no number is precisely the one the reader will assume somebody priced. The merge's own central fact is the likeliest thing to be missing from a gap list assembled by walking the inputs, because it belongs to no single input — walk the verdict too, clause by clause, after the verdict is written.

The queue: reports that arrive over days

Inputs rarely land at once, and the reader acts mid-stream. So every merge after the first has a second job: say what the new report does to a decision already taken.

Three outcomes, and you must choose one out loud:

• Confirms it. Say so in one line and stop.
• Narrows it. The decision stands, its scope or its margin changes.
• Invalidates it. Lead with that. "The figure your Wednesday decision rested on is now wrong" is the verdict, not a footnote to it.

Carry forward what the reader was told last time, so a correction reads as a correction rather than as a fresh opinion.

The ledger, and when the set is too big for one

The ledger is the merge's audit record, and it is a named artefact, not a mood. It is one row per merged claim carrying: the claim, which reports it came from, how many independent sources those reports collapse to, its label, and the report-and-section pointer that lets a reader return to the argument under it. Its closing summary — source collapses, contradictions, supersessions, parked items — is illustrated at the end of the museum example. Whenever this file says "say so in the ledger", that row is where it goes. Sourcing a claim is human-output's rule; what is added here is the source count beside it, because that is the number a single-report writer never had to produce.

Beyond roughly a dozen inputs the ledger is itself a wall, and you must choose tiering or refusal rather than letting it grow. The criterion is whether the set partitions cleanly by decision:

• Tier it when the inputs fall into groups that each bear on one decision, with few claims crossing a group boundary. Merge each group, then merge the merges, keeping pointers and source counts through both passes.
• Refuse and say the set needs splitting when the groups do not separate — when the same claims are load-bearing in several groups at once, so any partition you draw cuts through a relation. Tiering there hides the crossing relation inside a pass nobody reads, which is worse than the wall.

A twenty-line ledger reproduces the problem in a smaller font.

When they do not combine

This is a legitimate deliverable, not a softer merge. A merge with zero combined claims is a valid output. Never manufacture a joint sentence to justify the skill having run.

The shape is exact:

1. The refusal, first line, stated plainly.
2. One named reason per parked item — not one reason for the set.
3. What would make it combinable, per item.
4. Then each finding in full, in decision order, with its own verdict, its own grade, and its own back-pointer.

Worked example — a museum archive

The museum example and numeral trace are bundled separately. Read them for an explicit onboarding or demonstration request, or when a procedural ambiguity remains after reading the core rules. Ordinary merges skip the example. It illustrates this procedure and supplies no evidence about the current inputs.

Keep the example's frame discipline on every route: a chosen analysis frame, such as a time horizon, is a choice rather than a source finding. Label it where first used and expose any decision-reversing alternative alongside it.

Resolve the reference relative to this installed skill directory. If it is missing or unreadable, report that gap and use the core procedure; do not reconstruct the example or resolve unknown input facts from it. Report any remaining procedural limitation. Return to Checking after the example; the ordered procedure, ledger and complete numeral trace remain mandatory.

Checking

Resolve this installed skill's directory and run node <skill-dir>/references/tools/human-output-check.mjs <draft-path> on the merged draft. The bundled checker grades the mechanical half: printable characters and width inside figures, bar arithmetic, percentage sums, sentence length, acronym expansion, one marked recommendation.

It cannot grade whether independence was established, whether a conflict was classified honestly, or whether a dropped claim was superseded rather than merely inconvenient. Nor can it tell an invented figure from a traced one — that limit and its remedy are the checker contract, references/human-contract.md section 3. So do the one check that catches the worst defect: freeze the inputs, then compare every numeral in the merge against them one at a time, and print the trace where the reader can see it. A merge with no trace should be read as a merge whose losses were not counted.

Then three checks the numeral pass does not cover, because each has caught a merge that had already passed it:

1. Every non-numeric claim gets the same pass. Numerals are the easy half. Run the sentences too: "still the live document", "no longer exists", "has never been entered" carry no digit and are exactly where an unsourced fact hides. Point each at the line of the frozen input that says it. If no line says it, it is an inference — label it or cut it.
2. Every fact in the verdict carries a figure or appears in the gap list. Walk the verdict sentence clause by clause against the pooled gaps.
3. The trace's own guarantee is true as written. "Nothing else carries a number" is a claim about the merge, and it is false the moment a year, a chosen horizon or a tally appears untraced. A trace that overclaims completeness is worse than no trace, because it spends the reader's trust on the pass it did not run. State the guarantee you actually met.

Boundary

This skill is one member of the human- family. The routing table — which member handles which material — is references/human-contract.md section 1.

• Not /human-output — the contract for writing one piece. A merge is written to it and adds only the cross-report layer. Every style rule stays there.
• Not /human-rewrite — that repairs one existing text and explicitly refuses several separate reports, because it reads nothing beyond the one text it was handed. When it refuses, this skill is where the work goes. Call it first on any single input too broken to read into the merge.
• Not /human-draw — that owns what is inside the fence. Hand off when the source structure or the joint figure is relational or proportional; the prose around it stays this merge's output.
• Not /extract-signal — that gathers and vets signal from noisy sources. Merge reads the supplied finished reports and accompanying provenance briefs; it does not retrieve new source material.
• Not /double-check or /verify — merge never re-verifies a claim against the world. It grades and traces what it was given, and names the check that would settle a conflict for someone else to run.
• Not an aggregator. Concatenating reports under headings, or averaging their conclusions, is the problem this skill was written against.
• Not a substitute for the underlying work. Weak reports merged cleanly produce a confident wrong answer faster.

Human output. This skill's handler-facing output obeys the human-output contract (human-output/SKILL.md).

Next skills. On completion, run the Next Skills protocol (next-skills/SKILL.md): surface the next-skills recommendations from front-matter for the caller to pick. Offer only — never auto-invoke.

Self-improvement. Run the Self-Improvement Protocol only when its installed owner is available. Preserve its qualification, vetting and approval rules. If absent, skip that optional hook without claiming it ran, surfacing an unvetted fix or auto-editing; substantive work may complete. (self-improve/SKILL.md): apply its signal taxonomy and threshold to this run; surface any surviving proposal for handler approval. Conservative — silent otherwise. Never auto-apply.