{"path":"research/incentives-analysis.md","content":"# The Incentive Structure of Deliberus: What Makes Legibility Pay\n\n*Aug 13, 2026. A full incentive analysis, for the destination and for the path to it. This composes with two existing docs rather than repeating them: [adoption-problem.md](adoption-problem.md) covers why twenty-plus predecessors died and what successful knowledge platforms did, and [epistemic-gamification.md](epistemic-gamification.md) covers reward mechanisms, badge taxonomies, gaming defenses and the Stack Overflow warning. Neither asks the question this document is built on: **under what conditions is being legible individually advantageous?** Everything else follows from the answer.*\n\n---\n\n## 1. The diagnosis has changed and the corpus has not caught up\n\n`adoption-problem.md` locates the failure of every predecessor in **friction**: structured argumentation demanded formal structure as input, which is a cognitive tax only academics would pay. The project's answer has been that LLMs absorb the structuring burden, so structure becomes output. That is correct and it is done — extraction works, and the tax that killed Debategraph and the Deliberatorium is genuinely gone.\n\nBut removing the labor cost does not make contribution individually rational. It removes one of three costs. Look at what a contributor actually pays once the typing is free:\n\n1. **Labor** — structuring the reasoning. **Solved by LLM extraction.**\n2. **Exposure** — every premise you state becomes a surface someone can attack. A person who articulates their argument has handed opponents a target list; a person who stays vague has not.\n3. **Social** — the cost of being *seen* engaging seriously with the other side, which in a polarized setting reads as defection rather than virtue.\n\nSo the honest post-2023 statement is: **LLMs solved the labor problem. Exposure and social costs are now the binding constraints, and neither is addressed by better tooling.** The red team already found the second and third in another vocabulary — \"constructive ambiguity,\" the observation that some agreements survive only unspoken — but the corpus has never named vagueness as a *defensive asset* whose surrender is what we are asking for.\n\nThis matters because it changes what growth work looks like. If friction is the constraint, you build better UX. If exposure is the constraint, you find or construct the conditions where exposure pays. Those are different projects.\n\n## 2. When legibility pays — eight conditions\n\nThis is the analytical key. Legibility is not universally advantageous or universally costly; it flips sign depending on the situation. Enumerating the situations where it pays gives both the target population and the growth sequence.\n\n| | Condition | Why legibility pays | Who is in it |\n|---|---|---|---|\n| **L1** | **You are being misrepresented and it costs you** | A ratified, citable statement of your actual view is a defensive asset | Holders of unpopular-but-defensible positions; researchers whose work is misdescribed; anyone routinely strawmanned who believes they would win on the merits |\n| **L2** | **You must persuade someone who does not trust you** | Legibility substitutes for trust — show your work and trust becomes unnecessary | Grant applicants, expert witnesses, consultants, anyone writing a safety case |\n| **L3** | **You must coordinate with someone you disagree with** | Not understanding them costs you the deal | Groups with a decision at stake and a disagreeing counterparty |\n| **L4** | **Your reasoning is your product** | Inspectability is a quality signal, not a vulnerability | Analysts, researchers, forecasters, reviewers |\n| **L5** | **You will be audited anyway** | Structuring now is cheaper than justifying later | Regulated industries, medical and legal decisions, institutional review |\n| **L6** | **You are arguing with yourself** | No adversary means **no exposure cost at all** | Anyone thinking through a hard personal question |\n| **L7** | **Legibility is a status move** | Understanding your opponent well reads as strength rather than concession | Contexts where an ideological Turing test is the scoring rule |\n| **L8** | **AI is the reader** | Structured reasoning is *distribution* in an AI-mediated information environment | Anyone who would rather be represented by their own reasoning than by someone's summary of it |\n\nThree observations that fall straight out.\n\n**L6 is the only condition with zero exposure cost**, which explains why \"personal thinking tool\" keeps recurring as a framing even though the founder has rejected it as the identity. The resolution: **solo use is incentive-clean and therefore the correct first stage, while remaining wrong as the destination.** The July reading — that product, personal tool and instrument were three successive legibility-shrinkages — was right about identity and should not be read as a prohibition on sequence. What makes the difference is whether the private stage has a path out of itself, which §6 takes up.\n\n**L5 is the strongest incentive available and the hardest to obtain.** When an institution requires showing your work, individual incentives stop mattering and compliance takes over. This is why the peer-review surface is worth the attention it got, and why the proposed Legibility Rule is an *incentive* mechanism as much as a constitutional one.\n\n**L8 is new, timely, already built, and undervalued.** The agent-readable surface has been treated as a courtesy toward AI readers. It is better understood as the reasoning layer's distribution channel: as AI mediation grows, whoever's reasoning is structured gets represented in AI-mediated discourse, and whoever's is not gets summarized by someone else. That is a self-interested argument for contribution that did not exist three years ago. §7 flags the reciprocity risk that comes with it. **L8 was collected into a strategy on 2026-08-17** — agents as a consumer class, with a second falsifier that needs no humans, and with the reciprocity question of §7 promoted from a values question to the survival question, since the graph's content still comes from people: [agents-as-a-consumer-class.md](agents-as-a-consumer-class.md).\n\n## 3. The actor ledger\n\nFor each party: what they get, what they pay, and whether it nets positive **without altruism**. That last column is the whole point — a commons that requires altruism from every participant is a commons that stays small.\n\n| Actor | Gets | Pays | Nets positive without altruism? |\n|---|---|---|---|\n| **Reader** | Understanding at falling marginal cost | Nothing — reading is free and unauthenticated | **Yes, trivially.** Which is why there is no conversion pressure at all |\n| **Contributor (own view)** | Accurate representation, a citable artifact | Exposure of every premise | **Only under L1–L5.** Otherwise no |\n| **Contributor (attacks own view)** | Nothing | Exposure plus advantage conceded | **No.** See §4.2 — this is the load-bearing failure |\n| **Decomposer** | — | Historically the binding cost | **Removed.** Machine work now |\n| **Maintainer** | Nothing | Ongoing pruning, re-typing, updating | **No, and no mechanism is even proposed.** See §4.5 |\n| **Institution** | Error reduction, legitimacy, audit trail | Process change, exposure of its own reasoning | **Yes under L5**, and its motive is its own legitimacy rather than the commons |\n| **Funder** | Evidence of a real instrument in a neglected niche | Money | **Yes**, with the bias noted in §4.6 |\n| **AI lab / AI reader** | High-quality structured reasoning | Nothing currently | **Yes** — and that asymmetry is §7's warning |\n| **Adversary** | Selective credibility (§5) | Almost nothing | **Yes.** Which is the problem |\n| **The public** | The commons | Nothing | Yes, and has no agency, which is what makes this a public good |\n\nThe shape is standard for a public good and worse than standard in one respect. Ordinarily the contributor pays a cost and produces a diffuse benefit. Here the contributor pays a cost, produces a diffuse benefit, **and increases their own vulnerability.** Triply negative, which is a better explanation of twenty years of failure than friction alone.\n\n## 4. Five structural asymmetries\n\n### 4.1 Cold start is inverted, and extraction is the answer\n\nReader value scales superlinearly with corpus coverage, because the marginal cost of understanding the *n*th holder of a position collapses once the premises are mapped (the corpus effect, [lowering-the-cost.md](lowering-the-cost.md) §4). Contributor cost is roughly constant. So early contributors pay full price for near-zero reader value, and the incentive to contribute is weakest exactly when the graph is smallest.\n\nNothing motivates early contribution except private value (L6) or the founder doing it himself. Which is why **the extraction pipeline is a cold-start solution and not merely a feature**: it accumulates coverage without requiring any contributor incentive at all, because the reasoning being mapped belongs to authors who already wrote it.\n\nThat works, and it has an unnamed cost. The people whose reasoning is mapped did not opt in and receive nothing. At twenty-one sources of public text, fairly analyzed, this is unremarkable. At scale it becomes a legitimacy problem — and **ratification is precisely the mechanism that converts an unconsented mapping into a consented one while delivering the L1 benefit to the person mapped.** So ratification is not only a trust primitive; it is what makes the cold-start strategy survivable as it grows.\n\n### 4.2 Supports are self-serving and attacks are altruistic — and the strength metric inherits the asymmetry\n\nThe sharpest finding here, and it is concrete.\n\nAdding a premise that supports your own claim serves you. Adding a strong attack on your own claim is pure altruism, and worse, it hands an opponent a weapon. So contribution will systematically over-collect supports and under-collect attacks.\n\nQBAF strength is computed from evidence energy — supports minus attacks — pushed through QEM. If the two inputs are collected under opposite incentives, **strength is systematically inflated, and the inflation is invisible because it looks like well-supported reasoning.**\n\nThis is not hypothetical. Run 5's hand extraction of a real referee report produced 22 claims, 14 support edges and **zero attack edges**. That was read at the time as a finding about referee practice — verdict and critique are orthogonal — which it is. It is also exactly what this asymmetry predicts.\n\nThree candidate defenses, none currently in place:\n- **Adversarial pairing at ingest.** Extract opposing sources together so attacks arrive from the other side's supports rather than from anyone's altruism. Already the informal practice for debate pairs; making it a rule turns it into a defense.\n- **Reward attack contribution specifically**, which `epistemic-gamification.md`'s Delta-system discussion is the closest existing thinking to.\n- **Report the asymmetry rather than hiding it.** A support-to-attack ratio per subgraph, published beside the strength, would make the inflation visible. This is the cheapest option and it fits the confession principle exactly: an instrument that cannot report its own bias will report health.\n\n**The third defense now exists at one layer** (Aug 13, 2026). `citation_balance` in the synthesis ledger reports the support-to-challenge mix of the claims an answer cited against the mix of what was retrieved, so a synthesis that quietly prefers supports is visible in the response body. Two limits worth stating rather than letting the fix look larger than it is. It measures the *synthesis*, not the graph — the per-subgraph ratio published beside QBAF strength, which is the version that addresses the metric itself, is still unbuilt. And it says nothing about actors: the ledger scores an answer's selection, and the selective-legibility move in §7 is a pattern across an actor's contributions that no per-answer instrument can see.\n\n### 4.3 Reading is free, so there is no gradient from reader to contributor\n\nCorrect for a commons, and it means nothing pulls a reader one step further in. The fix is not to charge for reading. It is to **make reading itself generate contribution as a byproduct**: votes, divergence marks, \"I did not follow this\" flags.\n\nThis composes with the position fingerprint proposal in a way that is almost too neat. The fingerprint — *you already agree with 14 of these 17 premises, here are the 3 that matter* — requires the reader's own votes to compute. And votes are the one contribution a reader will actually make, because they cost a click and immediately buy a better reading experience. **The feature that needs reader data and the only contribution readers will give are the same thing.** That is the cheapest growth mechanism available and it needs no altruism at any point.\n\n### 4.4 Beneficiaries are diffuse and future; cost-bearers are specific and present\n\nThe classic shape, stated for completeness. It is the reason the funding pipeline is not ancillary to the project but structural: someone has to pay the present cost of a future diffuse benefit, and in the absence of a market that is either a funder or a founder.\n\n### 4.5 Nobody's incentive is to maintain the graph\n\nContribution has eight candidate incentives. **Maintenance has none.** Pruning stale claims, re-typing termini as evidence moves, updating premises that have aged out, correcting a bad decomposition — all of it is cost with no return to the person doing it.\n\nNote that this is the same gap as the temporal rung, arriving from the incentive side. Strength computation is age-blind, so nothing *asks* to be maintained; and nothing rewards maintenance, so nothing would be maintained even if it asked. Wikipedia solved this with editor identity and status. Deliberus has no editor role, no status attached to curation, and no plan for either.\n\n### 4.6 The funder's incentive is to see progress\n\nStandard and worth naming rather than pretending away: funders reward legible artifacts, which biases work toward producible instruments over intractable problems. The honest guard already in place is that the artifacts are *falsifiable* instruments — a residue map that can score the wager as losing is not the same kind of object as a polished demo. That distinction is what keeps artifact-production from being a sophisticated form of avoidance.\n\n## 5. Adversarial incentives: selective legibility\n\nThe red team covered weaponized decomposition, coordination gaming and sybil exposure. One attack is missing, and it is the one the incentive structure actively rewards.\n\n**A rational actor is legible exactly where legibility helps and vague everywhere else.** Map your strong arguments in full detail; leave the weak ones unstated. The graph then shows you as well-supported and your position as thoroughly examined, because it can only measure what is in it. Meanwhile the honest actor who mapped everything, including their soft spots, looks worse.\n\n**Selective legibility is an attack that is indistinguishable from good citizenship**, and every existing instrument is blind to it. The completeness oracle measures a *claim's* exposure, not an *actor's*. There is no representation anywhere for strategic silence.\n\nThe defense follows from §4.1 and is elegant: **let opponents propose your premises, and let you only ratify or contest them.** You gain accurate representation and lose control of the agenda. That trade is positive for someone confident in their position and negative for someone who is not, which is exactly the selection the system should want. Stated as a design principle: **the incentive structure should advantage the confident and disadvantage the evasive.**\n\nTwo smaller ones worth recording. The completeness oracle is a rhetorical weapon in waiting — *\"your position has four unsupported value premises and mine has one\"* — which is Goodhart applied to the honesty instrument itself. And whoever pays for extraction acquires agenda-setting power over what gets mapped, which is the funding-layer version of the same problem; if AI labs ever become the paying customer for structured reasoning data, the incentive shifts from mapping reasoning faithfully to producing data that scores well.\n\n## 6. The growth path: which incentive carries each stage\n\nThe incentives that work at each stage differ, and some poison later stages. This is the sequence the eight conditions imply.\n\n**Stage 0 — now. Evidence, not users.** Founder plus a handful of collaborators. Their goodwill is social capital rather than incentive, it does not scale, and the DeepMind result confirms that their liking it proves nothing about quality. Stage 0's job is producing artifacts that persuade a stranger: the residue map, the hand-run extractions, the disagreement-preservation benchmark. Correctly, this is the current posture.\n\n**Stage 1 — solo use, the incentive-clean entry (L6).** Private thinking has no adversary and therefore no exposure cost. It is the only door with no toll. The danger is stalling there, and the guard is that **the private-to-public promotion path is the actual growth mechanism**, not a UX nicety. That layer is already half-built as deliberation drafts with staged maturity, and this analysis reclassifies it: it is the bridge between the only stage with clean incentives and everything after.\n\n**Stage 2 — dyadic (L3, L1).** Two parties with a decision at stake and a disagreement. The cost of *not* understanding is felt immediately and privately, which is the strongest near-term pull available. This is where ratification and the position fingerprint pay off, and it is the right shape for the friends round: not \"come look at the graph\" but \"you two disagree about something real, here is what you actually disagree about.\"\n\n**Stage 3 — institutional (L5, L4, L2).** Mandatory legibility, where compliance replaces motivation. Peer review is the identified surface, pre-submission self-review is the entry point that needs no institutional buy-in, and the institution's own motive is its legitimacy rather than the commons. This is the first stage where growth does not depend on individual willingness.\n\n**Stage 4 — the commons (L7, L8).** Status and AI-mediated distribution carry it, with network effects finally working in the project's favor because the corpus effect has accumulated. Note the ordering constraint: L7 requires a culture that scores understanding as strength, which is built rather than found, and L8 requires the reciprocity question in §7 to have an answer.\n\nThe sequence has one hard property worth stating: **each stage must produce the asset the next stage needs.** Solo use produces claims. Dyadic use produces votes and divergence data. Institutional use produces the coverage and credibility that make the commons worth reading. Skipping a stage means arriving without the asset.\n\n## 6b. The supply-side-first variant (founder-proposed 2026-08-20): fill the graph first, and where each stage's incentives then stem from\n\nThe founder's proposal: ingest the world's arguments at scale — the way AI labs ingest text — fill the graph, refine the ontology on real coverage, and let users arrive to an already-rich map, with incentives following coverage. This is §4.1's cold-start observation promoted from tactic to strategy, and the Romer point (value scales with coverage, not user count) is what makes it possible. The staged incentive-origin analysis:\n\n**Stage A — machine filling.** Actors: founder + pipeline. Incentive source: founder conviction + funder-facing evidence; deliberately requires nothing from anyone else. The incentive question INVERTS here — from \"why contribute?\" to **\"why come look?\"** — so the binding constraint shifts from exposure to ATTENTION (distribution, not extraction, is the scarce input). Per audience: funders (stems from artifact falsifiability, not usage — coverage metrics, never preference metrics, per the DeepMind discipline); agents (stems from dispute-structure data nobody else holds, activated by coverage × query fit; gated on the strength layer + provenance labels); drive-by readers (stems from the map answering an arrived-with question; requires discoverability); and the mapped authors, who at this stage get nothing and pay something — every source accrues **ratification debt**, unconsented mapping that is unremarkable at 25 sources and a legitimacy problem at 10,000.\n\n**Stage B — the mapped arrive, which is the payoff.** Mass ingestion **inverts the exposure calculus of §1**: for an author whose published argument is already mapped, the armor came off at publication, not here. The decision is no longer \"do I dare expose my premises?\" but \"is my mapped position accurate?\" — the L1 incentive running downhill. This is the Wikipedia-biography dynamic (nobody opts in; nearly everyone with a page cares that it is right), with the same pathology to design against: done badly it reads as surveillance, and ratification is what converts it to consent. Supply-side-first thus MANUFACTURES the project's strongest contributor incentive rather than waiting for it.\n\n**Stage C — readers with stakes.** Dyads with a live disagreement start from a mapped landscape instead of an empty canvas; incentive stems from the decision at stake plus the corpus effect. This does not violate the no-pre-crunched-workshop ruling: their own dispute is still structured live from their own words — the background corpus is the library the session happens inside, not a tour replacing contribution. Reader votes (the one zero-altruism channel, §4.3) begin compounding because the position fingerprint finally has a corpus to compute against.\n\n**Stage D — institutions and the commons**: as §6 stages 3–4, arriving earlier because coverage exists.\n\n**Three gates that decide whether Stage A produces a map or a pile:**\n\n1. **The connective tissue is the currently-failing layer.** Run 6 measured it: claim-level cosine finds vocabulary, not kinship — the Israel/Palestine pair added 8 SIMILAR_TO edges, all internal, zero to the other 23 sources. Mass ingestion today mass-produces islands. The honest unit of ingestion is therefore the **debate cluster**, not the source — adversarial pairs and constellations extracted together carry their own connective tissue, and §8's recommendation 2 already wants this rule for the attack-collection reason.\n2. **Corpus size is a liability while the ontology is moving.** The Data Freshness directive requires re-extraction on material pipeline changes, and the three-runs law says taxonomies break per REGISTER, not per volume — so refinement wants register diversity at moderate volume; mass volume comes after the ontology stabilizes, or every refinement multiplies in cost.\n3. **Quality ceiling + maintenance.** The crux of a real debate is typically implicit (run 3F, both batches), and implicit-premise recall is where cheap extraction is weakest — coverage without the load-bearing layer. And §4.5 stands: nobody's incentive is maintenance; a 10,000-source graph decays with no gardener (the temporal rung arriving at scale).\n\n**A fourth gate (added 2026-08-20 from the AI-slop analysis)**: source quality/provenance. Mass-ingesting a slop-contaminated literature mints phantom claims with fake evidence at scale — prefer sources with resolvable references and verifiable authorship, and run a deterministic reference-resolution check where the register allows ([ai-slop-and-the-reasoning-layer.md](ai-slop-and-the-reasoning-layer.md) §2c, §2f).\n\n**The lab disanalogy worth keeping**: labs ingest into weights — lossy, provenance-free, legitimacy debt hidden inside the model. Deliberus ingests into a public, attributed, contestable map — the debt is worn openly, which is both the product and the obligation. Corollary: an LLM's knowledge of the debate landscape is a SCOUT, not a SOURCE — it can enumerate debates, poles and canonical documents, but claims enter the graph from checkable sources, because a memory-derived claim has no provenance for the confession principle to audit. (Cost note: at paid tier, a thousand sources is roughly $300 of extraction — the gates above are the real constraints, not the money.)\n\n## 7. Two precedents that matter more than the rest\n\n`adoption-problem.md` surveys Reddit, Wikipedia, Stack Overflow, Hacker News and Quora for network effects. Two deserve re-reading specifically as *incentive designs*.\n\n**Community Notes has already operationalized the ideological Turing test at scale, and solved the incentive.** The reward is publication of your note. The gate is that raters who usually disagree must both accept it. So the contributor is rewarded precisely for writing something the other side signs off on — which is the L7 incentive, running in production, on a platform with hundreds of millions of users, embedded where the arguing already happens rather than at a destination people must visit. The corpus cites it repeatedly as a *bridging* precedent (`bridging.md`, `the-missing-layer.md`, the red-team docs). It has never been read as the answer to \"how do you make cross-tribal understanding individually rewarding,\" which is what it is. **Highest-value external study available for this question.**\n\n**Stack Overflow is the warning, and it now applies to a surface Deliberus has already shipped.** `epistemic-gamification.md` records the Stack Overflow lesson about gamifying the wrong thing. The newer and sharper version: Stack Overflow's contributor incentive collapsed when AI could extract the value without reciprocating — answers were consumed at scale, attribution and traffic disappeared, and contribution fell. The agent-readable surface is Deliberus's version of that exposure. It is a genuine L8 distribution win and simultaneously an invitation to extraction without return. CC-BY requires attribution, which is a floor rather than a solution. **The open question this raises: what does an AI reader owe the graph it reads?** Nothing in the design answers that, and the question will not stay theoretical.\n\n## 8. What this implies, ranked\n\nBy leverage over effort, and all of it composes with what is already registered.\n\n1. **Publish the support-to-attack ratio beside QBAF strength.** Cheapest fix for the most consequential asymmetry, and it is the confession principle applied to the project's central metric — an instrument that cannot report its own bias will report health. Run 5's zero attack edges is the first datapoint.\n2. **Make adversarial pairing a rule at ingest, not a habit.** Extract opposing sources together so attacks arrive from the other side's supports rather than from anyone's altruism. This is the structural version of the same fix.\n3. **Ship ratification.** It buys trust (§3), converts unconsented mapping into consent (§4.1), delivers the L1 benefit that makes contribution self-interested, and enables the opponent-proposes/you-ratify design that defeats selective legibility (§5). One mechanism, four returns.\n4. **Treat reader votes as the contribution channel.** They cost a click, they immediately buy the reader a better experience through the position fingerprint, and they require no altruism. The only growth mechanism here with no incentive problem at all.\n5. **Study Community Notes as an incentive design** rather than as a bridging precedent, and specifically its publication-gated-on-cross-tribal-agreement mechanism.\n6. **Reclassify the promotion path** from staged-maturity UX to the growth bridge out of stage 1, and design it accordingly.\n7. **Name a maintenance role**, or accept that the graph decays and say so. Currently neither.\n\n## 9. Open forks for the founder\n\n- **Does anything gate reading?** Total free-riding is right for a commons and removes every conversion gradient. The recommendation above is to keep reading free and harvest votes instead, but the alternative — light reciprocity, in the Community Notes shape — has not been considered.\n- **What does an AI reader owe the graph?** (§7.) Attribution is the CC-BY floor. Whether more is wanted is a values question, and it interacts with the whole L8 distribution bet.\n- **Should the support-to-attack asymmetry be reported, corrected, or both?** Reporting is honest and cheap; correcting means weighting attacks upward, which is a semantics change to the shipped QBAF and should not be done casually.\n- **Is \"advantage the confident, disadvantage the evasive\" an acceptable selection principle?** It is what the opponent-proposes design does, and it has an obvious failure mode: confidence and correctness are not the same thing, and the loud-and-wrong would be advantaged over the careful-and-uncertain. Stated plainly because it is a real cost of a mechanism this document otherwise recommends.\n\n## 10. What this document deliberately does not contain\n\n**The founder's own incentive position.** It is the most load-bearing incentive in the project — an unfunded solo maintainer faces a different structure than a funded team, and which paths are viable follows from it — and it cannot be written here, because the specifics are personal financial circumstances and everything under `docs/` is published at deliberus.com the moment it is committed.\n\nIts home is `.private/`, which a cloud session cannot reach. Registered as a TODO item for a local session rather than improvised into a home that would publish it (CLAUDE.md § Cloud-Session Notice). The structural half that *is* safe to state: the grant pipeline is not ancillary to this project, it is the mechanism that makes continued development individually rational for whoever does it, and the self-financing alternative competes for the same hours.\n\n---\n\n**See also**: [adoption-problem.md](adoption-problem.md) · [epistemic-gamification.md](epistemic-gamification.md) · [lowering-the-cost.md](lowering-the-cost.md) · [curiosity-as-growth-fuel.md](curiosity-as-growth-fuel.md) · [governance-and-financing.md](governance-and-financing.md) · [legibility-under-power-red-team.md](legibility-under-power-red-team.md) · [bridging.md](../bridging.md) · [the-missing-layer.md](../the-missing-layer.md) · [vision.md](../vision.md) § Opacity Is a Cost, Not a Mystery\n"}