--- skill: Contested Ground version: 1.0 source: The Mantle Institute, mantlecollective.com/playbook/contested-ground layer: school sfia_minimum: 4 model_maturity: L3 tags: [research, knowledge-work, sensemaking, writing, evidence] license: CC BY 4.0 --- # Contested Ground A method for turning an unreadable volume of articles into a small set of live debates, by mapping where trustworthy sources disagree with each other rather than what they are collectively about. The output is a list of contested questions, each with the sources arguing either side and the specific claims they make. That list is a writing outline: an unsettled question, the evidence on both sides, and the space where your own view belongs. --- ## How to use this skill Load this file as context at the start of a session. Tell your AI: "We are running Contested Ground over the following sources. Work through the phases in order. Do not cluster anything until every item has been distilled to a claim and an anti-claim." The order matters more than it looks. The value of this method comes from comparing claims, not documents, so any shortcut that clusters whole articles by topic produces the thing this method exists to avoid. --- ## Input modes **Reading list.** A set of articles you have already gathered. The most common starting point, and the one where results arrive fastest. **Standing feed.** An ongoing harvest that grows over time. Debates persist and shift across runs, so this mode benefits from the lineage step in Phase 6. **Single question.** A specific question you want to map. Gather sources deliberately from opposing camps rather than from one publication or one discipline. --- ## Workflow ### Phase 1: Harvest wider than you can read Collect more material than you can possibly consume, and accept that most of it will stay unread. You cannot detect a disagreement inside a single viewpoint, and the most valuable clashes tend to cross fields: a clinical study that undercuts a claim made about enterprise software, a policy paper that contradicts a vendor benchmark. The goal is not to read more. It is to make *not reading* something a decision you took rather than an accident of what happened to arrive. --- ### Phase 2: Distil each item to a claim and its opposition For every item you look at, produce exactly two sentences: 1. **The main claim.** What this source argues, stated so plainly that its author would recognise it. 2. **The anti-claim.** The strongest opposing position a rigorous critic would hold, stated as a claim in its own right rather than as a rebuttal. The anti-claim is what makes the rest of the method work. The main claim tells you what a source believes. The anti-claim tells you what it is arguing against, which is a direct pointer to where another source in your collection is standing. Discard anything that makes no real argument. Announcements, launch coverage and marketing copy yield claims so broad that they appear to conflict with everything, which buries the specific debates you are looking for. The Adversarial Distiller skill covers this phase in depth if you want stronger claims than a single pass produces. --- ### Phase 3: Propose candidate pairs cheaply Comparing every claim against every other claim does not scale: a thousand claims is roughly half a million comparisons. Use a two stage funnel instead. **Cheap proposal.** A lexical ranking method such as BM25 will surface candidate pairs at almost no cost. Retrieving the top twenty candidates per claim caught roughly four fifths of the genuinely related pairs in our own corpus. It is not exhaustive and it is not meant to be. **Expensive judgement.** Spend your real effort, human or model, only on deciding how each candidate pair relates. Know what the cheap stage misses. Two claims can contradict each other without sharing a single word. "A conformance test can certify any AI system" and "Non deterministic output makes certification meaningless" are in direct opposition and have almost no vocabulary in common. Word matching will never connect them. If a debate matters to you, go and look for its other side by hand. --- ### Phase 4: Judge the direction of each pair For each candidate pair, decide which of these applies: | Verdict | Meaning | |---------|---------| | `CONTRADICTS` | Both cannot be true. A direct clash. | | `CHALLENGES` | One questions the evidence, scope or framing of the other without asserting the opposite. | | `SUPPORTS` | One strengthens the other with independent grounds. | | `RESTATES` | Same claim in different words. Carries no new information. | | `UNRELATED` | The proposal was a false lead. | `CHALLENGES` is the verdict most methods leave out, and leaving it out is expensive. A great deal of genuine academic and professional disagreement takes the form "that may hold, but it is not demonstrated here", which is neither agreement nor contradiction. When we tested a version without it, accuracy against a hand marked set sat at 42 per cent. Adding it took the same test to 85 per cent, because most real disagreement had been getting forced into a box that did not fit. Record `UNRELATED` verdicts rather than discarding them. They are answers, and keeping them stops you paying to re-judge the same dead end on every subsequent run. --- ### Phase 5: Cluster on disagreement, not agreement Build your groups from `CONTRADICTS` and `CHALLENGES` edges. Treat `SUPPORTS` with suspicion as a grouping signal: agreement is easy to find, and following every weak agreement link collapses a collection into a single mass. Our first attempt produced exactly one cluster containing 226 items, which is not a map of anything. Two defences work. Set a minimum confidence for an edge before it can join anything. Then use a community detection method that finds densely connected neighbourhoods rather than anything merely reachable, so a single weak link cannot fuse two separate debates. You are aiming for many medium sized clusters. One cluster holding most of your collection is a failure of the method, whatever its size. --- ### Phase 6: Name each cluster by the question it argues about Name every group after **the question in dispute**, not the topic it sits under. - Good: "Do technical AI benchmarks predict real clinical performance?" - Bad: "AI in healthcare" The question names what is actually at stake and reminds you that credible people hold both positions. A topic label quietly hides that. For a standing feed, match each run's clusters to the previous run's by membership overlap rather than by exact composition. A debate that gains an article is still the same debate and should keep its name and its history. Record when a cluster was last seen as a fact; treat dormancy as something you derive from that, because a debate that goes quiet for a month has not necessarily ended. --- ### Phase 7: Write from the seam You now have unsettled questions with sources arguing each side and the specific claims each side makes. The blank page problem is gone, because the structure of the piece is already there: here is an open question, here is what each side holds, here is where I land and why. This is where the method stops and you start. The map shows where the lines are drawn. It cannot tell you who is right. That judgement comes from your own experience, your own mistakes and your own knowledge of the field, and it is the part no tooling can supply. --- ## Maturity Requirements **Human (SFIA):** Level 4 minimum. The method produces contested questions, not answers, and its output is only worth having if the practitioner can adjudicate a genuine disagreement between credible sources and hold a position under challenge. At SFIA 5 and above the practitioner can also judge which contested questions are worth their organisation's attention, which is a different and harder call than judging which are real. The binding constraint is rarely tooling. It is whether the practitioner will sit with an unresolved disagreement long enough to form a view, rather than reaching for the source that already agrees with them. **Model (L1 to L5):** L3 for standard use. The model distils claims, proposes pairs and suggests verdicts; the human decides what the disagreement means and what to write. At L4 the model may also propose which contested questions deserve attention, though the selection of what to publish should stay with the human. --- ## Gotchas **Thin sources poison the map.** A claim distilled from a headline, a podcast blurb or a paywalled first paragraph comes out vague, and vague claims appear to conflict with everything. In our corpus, source length below roughly 900 characters predicted a bad claim far better than any test of the claim's own wording. Set a minimum source length, and accept that a real share of what you harvest will not clear it. **Filtering for signal quietly filters out conflict.** Every step that removes noise also removes disagreement, and from the inside the two look identical. It is entirely possible to tune a collection until it agrees with you and to experience that as improved quality. Track the ratio of disagreement to agreement over time. If it drifts towards agreement, the honest question is whether the field settled or whether you stopped listening to the other side. **Look for disagreement and you will manufacture it.** Some pairs the funnel proposes are wording differences dressed as conflict. Keep every claim traceable to its source so any surprising clash can be checked against the original. Never normalise several phrasings into one canonical sentence, because that destroys the evidence you would need to audit it. **Topic labels are not claims.** Two pieces filed under the same heading may share nothing, while a piece on clinical software and a piece on hiring algorithms may make the identical argument about whether automated evaluation survives contact with reality. Cluster on what is claimed, never on where it was filed. **Absence of disagreement is a finding, not a gap.** A question that draws no conflict means either the matter is genuinely settled or your sources are too alike to argue. Both are worth knowing. Only one is good news. --- ## Business Area Impact The roles most directly affected are those whose output depends on knowing where a field is unsettled: strategy, policy, analyst relations, technical marketing and any research function that briefs decision makers. The usual failure in those roles is not missing information. It is confidently reporting consensus that does not exist, because a scan of the literature surfaced the loudest position rather than the contested one. The roles that gain most are subject matter experts who hold strong views but lack the time to survey a field broadly enough to know whether their view is contested, marginal or mainstream. This method tells them which, and points them at the sources that would argue. --- ## Handoff Produces a set of named contested questions, the sources on each side, and the specific claims in dispute. That structure feeds article outlines, briefing notes, debate preparation and literature reviews directly. Because each claim keeps its link back to its source, the output also serves as a citation list for whatever gets written from it. --- *From The Mantle Institute, mantlecollective.com/playbook* *Human/AI Skills Matrix, Contested Ground v1.0*