Human Merge

/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 Name what is superseded — and whether a live decision already rests on the dead figure.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 Then ask what already rests on the superseded figure.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 the briefs before the reports. 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 what the example prints in italics at the end. 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 frozen inputs first, so every figure below can be checked against them.

The inputs, as received

  R1  Conservation, 2026-02-10
      Surveyed 4 of the 11 archive rooms, taking own readings:
      rooms 3, 5, 7, 9. Relative humidity above the 60 per cent
      ceiling in rooms 3, 5 and 7. The other 7 rooms were locked.

  R2  Finance, 2026-03-06
      Keep on site: new climate plant, EUR 610,000 to install and
      EUR 48,000 a year to run. Move off site: EUR 90,000 removals
      once, EUR 71,000 a year to lease. Lease price quoted by the
      store operator; both capital figures from the same 2025
      outline pricing note. Assumes the EUR 55,000 annual
      conservation grant continues.

  R3  Registrar, 2026-04-14
      The store operator states its halls hold 55 per cent relative
      humidity year round. Basis: the operator's brochure.

  R4  Reader services, 2026-05-08
      1,240 reader visits in 2025; 380 of those consulted archive
      material. Off-site retrieval adds 2 working days per request.

  R5  Building surveyor, 2026-05-20
      Water ingress through the north range roof recorded in rooms
      5 and 7 during February 2026. Own inspection.

  R6  Finance addendum, 2026-06-30
      The EUR 55,000 conservation grant was not renewed for 2027.
      The 2027 budget submission was filed on 2026-04-30, carries the
      grant as income, and is the standing submission; no revision
      has been filed. Basis for the non-renewal: not stated.

The merged surface

Moving off site is cheaper by EUR 290,000 over ten years but dearer for ever after year 22.6, and no report says how long the museum intends to hold this archive — so the ten-year frame may be the wrong one and the EUR 290,000 with it. Two of the three damp rooms also have a leaking roof that neither option fixes and that no report prices, and the standing 2027 budget still counts a EUR 55,000 grant that Finance now says was not renewed. Do not sign the lease on these reports.

Decision needed before the 2027 budget is reopened: refile the budget without the grant, and commission an independent humidity reading at the store. Both are prerequisites, not alternatives.

No longer true: the EUR 55,000 grant — on Finance's own unattributed word. R2 assumed the grant continues; R6 says it was not renewed for 2027. R6 also says the 2027 submission was filed on 2026-04-30, carries the grant as income, and is still the standing submission with no revision filed. The non-renewal was reported on 2026-06-30, so the submission was filed two months before the fact that breaks it was known, and the live document still counts money Finance says is gone. That is the verdict of this merge — and it rests on one source. (Estimate; the test is the next paragraph.)

Why that is an estimate and not "confirmed, R2 and R6". Two reports, so run the five keys. R6 is titled a Finance addendum and R2 is the Finance report — same author function and same corpus, so key 3 matches on the face of it; and an addendum to a corpus is written with that corpus in hand, which is key 4 (an inference from the title, not a stated fact, and it is not what decides this). The counterfactual: could R2 have been wrong about the grant while R6 stayed right? Not independently. They are one source, and their agreement adds nothing. R6 also states no basis for the non-renewal, and an unattributed claim reaches estimate at best however much else points the same way. Both rules bind and the weaker wins. Confirming it is one telephone call to the grant body, and it is the highest-value check in this set. (R2 and R6.)

R6 does not say the non-renewal falls differently on the two options, and this merge does not assume that it does. It bears on whether either option is affordable, not on which one is cheaper.

Together: the ten-year cost, which no report states. The ten-year window is this merge's own choice and appears in no report; it is here only because it is the conventional frame, and the crossover below is what actually decides. Keep on site is EUR 610,000 plus ten years at EUR 48,000, so EUR 610,000 + EUR 480,000 = EUR 1,090,000. Move off site is EUR 90,000 plus ten years at EUR 71,000, so EUR 90,000 + EUR 710,000 = EUR 800,000. The gap is EUR 1,090,000 - EUR 800,000 = EUR 290,000. (Arithmetic on R2.)

  Ten-year cost, one cell = EUR 50,000

  Keep on site   ######################   EUR 1,090,000
  Move off site  ################         EUR   800,000

  Staying is dearer until year 22.6, then cheaper for ever after.

The crossover, so the ten-year window is not doing the arguing. Moving is EUR 610,000 - EUR 90,000 = EUR 520,000 cheaper to start, and costs EUR 71,000 - EUR 48,000 = EUR 23,000 more each year. EUR 520,000 divided by EUR 23,000 is 22.6 years. If the museum expects to hold this archive longer than that, the cheaper option reverses. No report says how long it expects to hold it. (Arithmetic on R2; estimate — see the label below.)

Both capital figures rest on one costing note, so this comparison is an estimate. R2's EUR 610,000 and EUR 90,000 both come from the same 2025 outline pricing note — a primary record R2 did not produce but does name, so confirmed is reachable on access alone. Two figures from one document are one source, not two, so the merged cost conclusion is capped at estimate whatever its arithmetic. (R2.)

The damp is two problems, not one — but that split is itself an estimate. R1 measured rooms 3, 5 and 7 over the ceiling. R5 recorded roof ingress in rooms 5 and 7. Two of the three overlap; room 3 is over the ceiling with no ingress reported. A climate plant does not stop water entering through a roof, so the plant and the roof are separate spends. R1 and R5 each measured or inspected the thing themselves, but both use the same room numbering and neither says whether the surveyor had R1 in hand. Independence not established — counted as one source for the room-matching claim, which is therefore an estimate. (R1 and R5.)

The retrieval delay matters only because of the move — and it cannot be sized from these reports. R4's 380 archive consultations a year are unremarkable alone; against R2's off-site option they become the reader cost of moving. R4 is the load, R2 is what makes it heavy. But the two figures do not multiply. R4 counts 380 consultations and prices 2 working days per retrieval request — different units. No report says how many requests a consultation raises, or whether one request serves several, so 380 x 2 is a number with no parent. It becomes combinable when someone states the requests-per-consultation ratio, or R4 recounts in one unit. Until then the delay is real, unpriced, and not 760 days. (R4 counted its own visits; parked as non-combining.)

Does not combine: the store's 55 per cent. R3 passes on the operator's brochure — an interested party on its own condition, so judgement at best — about a different building; R1 measured this one itself. Different building and different basis: netting them would produce a number with no parent. It becomes combinable when someone takes an independent reading inside the store's halls.

Two reports carry a figure that favours the store; behind both is one source, and it is the party being paid. R2's lease price and R3's humidity figure both come from the store operator. Neither report states a view on whether the store is suitable — that reading is this merge's, and it is why the two figures are counted once.

  claim: "the off-site store is a suitable home for the archive"

  R2 finance    lease EUR 71,000 a year --.
                                           +-- store operator -> 1 source
  R3 registrar  55 per cent humidity    --'

  R1 conservation  own readings, 4 of 11 rooms -> 1 source, this building

  reports carrying a figure that favours the store: 2
  independent sources behind them: 1
  and that one source is the party being paid

No contradiction anywhere in the set. No two reports state incompatible values for the same subject, attribute and scope. The apparent humidity clash is a scope mismatch (two buildings) and the grant is vintage.

Nobody checked, pooled across all six: what it costs to repair the north range roof — the fact the verdict rests on, and no report gives a figure or was asked for one; whether the lease has a break clause; the reader cost of off-site retrieval, which cannot be sized from R4 as written; how long the museum expects to hold this archive, which decides whether year 22.6 falls inside the horizon or outside it; and the condition of the 7 rooms R1 found locked, on which no report in this set reports. Each author's own gap list omits all five, because none of the five sits inside one report's scope.

Merged 6 reports; source counts are per claim and given beside each — 2 reports collapse to 1 source on store suitability, 2 reports to 1 source on the grant, 2 capital figures to 1 costing note. 0 contradictions. 1 superseded figure with a live document resting on it. 2 items parked as non-combining. 3 claims left at estimate — the grant, the cost comparison, the room match. 5 pooled gaps.

The numeral trace

Every numeral above falls into one of four classes: verbatim from the frozen block; arithmetic on numbers that are; a count of the merge's own items or of this file's own lists; or — in exactly one case, flagged as such — a frame this merge chose. Every one of them is in the table. Shown longhand:

One number in the merge comes from no report: the ten-year horizon. It is a frame this merge picked, not a finding, and it is labelled as one where it is first used — which is why the crossover at year 22.6 is printed beside it rather than below it. Change the horizon and the EUR 290,000 changes with it; the crossover does not.

No date, deadline, headcount or threshold was invented, and no input's confidence was raised: the two figures resting on a single costing note are labelled estimate even though the arithmetic on them is exact.

Notice what the merge did not do. It did not average anything. It did not say "three reports agree the store is fine". It did not resolve the humidity question, because the reports cannot. It did not multiply two figures whose units differ merely because both were to hand. And it did not invent a fact to demonstrate a feature — the live decision resting on the dead figure is the 2027 budget submission, and each element of it (the filing date, the grant carried as income, the submission still standing) is stated in R6 rather than inferred from R6 and R2 together. Where this merge did reason across two reports, it said so and dropped the label to estimate.

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 is handed finished reports and reads nothing beyond them.
• 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. On completion, run the Self-Improvement Protocol (self-improve/SKILL.md): if this run surfaced a concrete, blocking-or- recurring weakness in this skill, propose a specific fix for the handler to approve. Conservative — silent otherwise. Never auto-apply.