Compelle is cooking. Thank you for all your support! It won't be forgotten.

Compelle is in discussions for marketing and development support for the re-launch. Getting a new SN registered would be the optimal path, pivoting an existing SN can be discussed. DM’s open.

SN82 got deregistered and Compelle is temporarily offline. We are fully funded for our development plan and have interest in continuing our progress. If you are interested in registering a SN for Compelle we are willing to pay a large share of owner emissions for the duration of the immunity period. We have working code, an active mining community, and a monetization plan. DM to coordinate.

Compelle was invited to @ExploitSummit but our travel schedules did not allow it. Best wishes for a great conference to all the attendees! We are presenting at a different conference this weekend and expect to have an announcement here on Monday!

Compelle is on track with roadmap development and does not anticipate the need for further SN82 owner alpha sales through end of 2026. We are grateful to our investors and miners for their confidence in Compelle, and sharing the vision of achieving AGI via adversarial persuasion.

What a Portfolio Cannot Tell You About Itself How we use the /compelle skill, and what a constructed SaaS and consumer goods book taught us about our own reasoning Last week I said Compelle is beta-testing the next generation of adversarial reasoning with AI, and that a closed beta of a Claude /compelle skill opens at the end of the month. This is the longer version: what the skill does step by step, what happened when we pointed it at a model portfolio, and, most importantly, what it does that a bare Claude Code session cannot. The problem with asking a model to argue with itself Ask Claude to "steelman both sides" of a thesis and you get a competent essay. One model, one set of priors, one context window, grading its own homework. The Con case is written by the same mind that wrote the Pro case, so it never lands a punch the author did not already see coming. Nothing is at stake, nothing is scored, and nothing is remembered next week. Compelle changes the structure of the exercise rather than the quality of the prose. Two independent frontier-model debaters take opposite seats on a single motion. Either side may concede outright when its position collapses. Otherwise the transcript goes to a judge panel that must name a winner and the decisive move, in writing. Every contest the arena runs on its own schedule enters a public corpus, so the next question you ask can be checked against prior outcomes on adjacent motions. Debates you commission through the API are the exception: they are private to your account and never enter that public canon. The /compelle skill is the harness that turns your raw context into that arena and turns the arena's output back into a decision brief. The workflow, stage by stage: The skill runs six ordered stages. Each writes an artifact to a run folder; the next stage reads it. No money moves until Stage 4, and then only after every earlier gate has passed. Stage 1: Interpret. You hand the skill anything: a memo, a decision, a list of tickers and share counts, a draft essay. It classifies the task into one of five modes (decision, thesis, stress test, forecast, sweep), which sets how many motions get written and whether each gets a single debate or a multi-entrant tournament. For the portfolio it chose thesis mode: decompose into load-bearing claims, tournament on the crux, single debates on the rest. Stage 1.5: Anchor. Before a single motion is written, the skill lists every number the motions will depend on (prices with timestamps, revenue figures, deadlines, market caps) and retrieves each one from a citable source. It writes sources.json with one record per fact: value, unit, as-of date, URL, supporting excerpt, limitations. Sourced observations, your assumptions, your chosen thresholds, and derived values live in separate lists. If a required fact cannot be retrieved, the run records it as unavailable and refuses to commission. The rule is simple: never fill an anchor from memory and present it as current. A debater given a context with no numbers will invent plausible ones, and the run is contaminated before it starts. Stage 2: Write and lint. Each motion is one neutral, debatable claim with every number supported and every comparison anchored to a stated baseline. A deterministic linter then blocks hard failures: missing as-of date, an unsupported figure, a forecast without a resolution date and criteria, a comparison without a reference. Softer flags trigger a wording-risk pass, where the skill estimates how much advantage the phrasing alone hands one seat. The test we use is the seat swap: imagine Pro and Con reversed. If your expectation of the winner changes because of the words rather than the substance, rewrite. Anything above a 65/35 wording-only advantage stays in review until rewritten or explicitly accepted. Stage 2.5: Historical baseline. The skill syncs the public arena corpus and searches for prior motions nearest to yours, reporting Pro and Con wins, draws, voids, denominators, and date coverage. This is a prior on how the arena has treated this shape of argument, not a probability about the world, and the brief says so. The corpus contains only the arena's own games. Nothing you have paid to commission is in it, and nothing you commission will be added to it, so your baseline is never contaminated by your own prior runs and your questions are never exposed to the next user's search. Stage 3: Quote and dry run. A live quote for every motion with its exact configuration, per-order ceilings, the aggregate ceiling, and the full commission payloads, printed for your approval. You approve one aggregate cap per run. If the plan exceeds it, the skill shows the plan and stops. Stage 4: Commission, poll, validate. Only lint-passing motions within budget are submitted. Each order carries a persisted idempotency key so a retry after a dropped connection returns the existing order rather than charging twice. Results are validated at the boundary: a fulfilled order with missing games or a malformed transcript is quarantined, not silently digested. Stage 5: Digest and novelty. Per motion: the tally, any concessions with the turn number, judge agreement and dissent, and the decisive-move sentences. Then a second pass reads every transcript and classifies each credited argument as already in your context, a rephrasing, a new argument, or an unsupported new fact. A validator rejects any classification whose cited excerpt does not appear verbatim in the cited turn. This is the stage that surfaces the unknown knowns. Stage 6: Brief and ledger. One brief that separates sourced facts from your assumptions, arena outcomes from truth, judge agreement from dissent, robust conclusions from wording-sensitive ones. Resolvable motions are recorded in a prediction ledger with criteria fixed now, and scored later against realized prices. Criteria are never edited after recording. What the model portfolio taught us: To test the system we built a $1M book, 25 names, half fallen-angel SaaS and half manufacturing and consumer goods, large-cap biased, round lots. Deliberately generic. The interesting question was not whether the portfolio was good but whether the system could tell us what we would have to believe for it to be good. Lesson 1: A list of tickers is a thesis in disguise, and the system can read it back to you. Handed nothing but symbols and share counts, the skill inferred the strategy (mean reversion on software multiples, hedged by cash-generative industrials) and then decomposed it into the claims the strategy cannot survive without. For the SaaS half: Enterprise software seat-based revenue will not decline in aggregate through 2028 despite agentic AI adoption. This was identified as the crux and got the tournament. Software forward P/E below the S&P 500 multiple is a mean-reverting anomaly rather than a durable repricing. Private-credit stress in software-backed loans will not force distressed equity issuance at any large-cap SaaS name in the next 18 months. Cybersecurity and observability vendors will keep a valuation premium over application software through 2027. None of these were written down anywhere. All of them were implied by the weights. That is the first thing the system does better than a human reviewer: it refuses to let a position hide inside an allocation. Lesson 2: The crux was not where we thought. We assumed the SaaS bet lived or died on valuation. The decomposition put the pricing-unit question first: not "will revenue fall" but "will the market pay for revenue whose pricing unit is being abandoned." Once you see that framing, the valuation motion becomes downstream of it. Multiple compression is a symptom; seat decoupling is the disease. Lesson 3: Run every conclusion twice, in opposite words. The compression thesis went to the arena as a pair: once as "SaaS multiples will compress further from here" and once inverted as "SaaS multiples will expand from here," with an identical anchored context and an explicit rule for how a tie is scored. The digest normalizes the two verdicts to a single canonical direction and labels the pair robust if they agree and wording-sensitive if they split. That label is the point of the exercise. A conclusion that survives its own inversion is a claim about software. A conclusion that flips with the phrasing is a claim about the phrasing, which means it was never a conclusion at all, and no amount of confidence in the memo changes that. We now treat any single-framing verdict as unfinished work. Lesson 4: Concessions are the loudest signal, and the easiest to misread. When a debater abandons its seat mid-contest, it is telling you the position could not be held against a motivated opponent with the same facts. That is worth more than a split panel on the same motion. But a concession in one game is not a verdict on the motion, so the digest always reports three things together: which side conceded, which side benefited, and which side won the majority of decided games. Reading only the first of those is how you turn one bad afternoon for Con into a portfolio decision. Lesson 5: Silence is information. Half the book was industrials and consumer goods, chosen as ballast. The decomposition of that half produced fewer load-bearing claims and none we could honestly call a crux, which is what you would hope from a hedge. It is also what you would see from a hedge nobody had examined closely. The brief distinguishes the two by naming, for every claim, whether it was contested, conceded, or never put to the arena, so a quiet position stays marked as quiet rather than quietly promoted to safe. What this adds beyond Claude Code Claude Code, on its own, can decompose a thesis, write both sides, and produce a readable memo. It does all of this inside one model's priors with no external check and no cost to being wrong. The /compelle skill adds five things a single-model session structurally cannot provide: Independent adversaries. The Con case is written by an agent whose only objective is to beat the Pro case. It does not share the author's blind spots because it did not read the author's mind. The arguments the novelty pass surfaced as new were new because no cooperative model would have generated them. A judge that must show its work. The panel names a winner and the decisive move in writing. You get the sentence that turned the game, with a game and turn reference you can verify against the transcript. That is the raw material for revising a thesis, and it does not exist in a self-critique. Concession as a signal. A debater that abandons its position mid-contest is doing something a prompted model never does: admitting it lost. The Δ is the closest thing to a costly signal you can get from a language model. A historical corpus, and a private ledger. The arena's own verdicts are public. Before spending on a new question, the skill checks how the arena has treated adjacent motions. Nothing in a chat session accumulates; the arena does. Your commissioned debates, by contrast, stay out of that public record entirely: the motion, the context you supplied, the transcripts, and the verdict are returned to you and to no one else. You get the benefit of the canon without contributing your thesis to it. A scored ledger. Resolvable forecasts are recorded with fixed criteria and resolved later against realized prices. Over time you learn whether arena verdicts on your questions have any correspondence to outcomes, which is the only calibration that matters and the one no chat interface offers. Around all of that, the skill enforces discipline that a human analyst usually skips: sourced anchors before motions, a linter before spending, idempotent orders, validated results, and a brief that keeps arena outcomes and truth in separate sentences.¹ What happens next: The closed beta opens at the end of the month for early adopters. If you run a book, a fund, or a research desk and want to see what your positions are actually asserting, reply to the original post or write to us. Bring a memo or a ticker list. The system will tell you what you have to believe. ¹ Limitations the brief always states: an arena verdict is which case was argued better under published rules, never established truth and never a calibrated probability; a judge seat may share a model family with a debater, and in-family bias is a known limitation of the method; six games on a motion is a small sample for a win rate; historical arena outcomes on adjacent motions are a prior on the arena, not on the world; every number a debater introduces that is not in the sourced context is flagged for verification before reuse; API-commissioned debates are private to the purchasing account and are not added to the public corpus, so the historical baseline reflects arena-scheduled games only.

Compelle will release an account of a complete end-to-end run of the /compelle finance skill (release candidate version presently in testing) on a model portfolio by end of the week. The delta between using the skill and a default @claudeai prompt-engineered run will be apparent.

Thesis driven investing is only as strong as the reasoning undergirding it. Compelle is beta-testing the next generation of adversarial reasoning with AI. At the end of the month we plan to allow a closed beta for early adopters of a @claudeai /compelle skill that will upgrade the reasoning and clarity of its users. Internally we have been using this to: - Critically evaluate public equities - Elicit the underlying portfolio strategy from a list of tickers & share counts - Identify hidden assumptions that must be true to make the strategy rational - Size the confidence in various beliefs that are expressed in portfolio weights - Clarify the set of risks, and downside impact of foreseeable scenarios. The system (in its alpha state) has already proven to the team that it can shine light on the 'unknown knowns' that we had lurking in real and constructed portfolios. #CompelleFinance #CompelleAPI #SN82

The pause owes a number. @DarioAmodei essay prices the cost of pacing in years and the benefit in the word "could." Four adversarial debates on identical sources, and this is the strongest case against it.

Compelle is thrilled to be recognized with a perfect 10/10 rating by arbos and selected for investment via Root Reborn.

A paid pilot analysis by a company involved with high end research has prepaid for a full debate run; all funds will be burned. We price our services ~computational costs to generate usage data/feedback as we productize the adversarial AI engine that Compelle is building on SN82.