Abstract
Most quantitative systems model the market as a noisy function. This paper also tests an author metaphor: price and volume are mapped to possible campaigns and a multi-agent game. The first engine reads observations as a sequence of scenes and estimates which modeled campaign, if any, best fits. The second maps the tape to game-state labels such as coordination, bait, and response. These are hypotheses inferred from observable data; they do not identify a participant, verify intent, or establish that a human or institution authored a move.
The most important lesson here is a failure we then repaired: a catalog of named plays decays for exactly the reason Papers I–III predict — naming a play makes it manufacturable — and the durable fix was to stop asking only "which known play is this?" and start asking "is this a real play I cannot yet name?"
1. From Judging Predictions to Reading the Opponent
Paper III separated prediction from judgment. This paper adds a third posture: modeling possible behavior on the other side of the trade. A price series is an observable residue of transactions; labels such as accumulation, distribution, bait, role, and victim are interpretations imposed by the model. If the observations follow repeatable sequences, a narrative representation can be tested against a feature-vector baseline. If several flows appear to interact, a game representation can be tested in the same way.
The two engines below are those two representations. They are deliberately not predictors. They are interpreters — they convert raw tape into a description of what the opponent appears to be doing, which the rest of the system can then reason over and, per Paper III, decline to trust.
2. The Playbook Decoder: Reading the Tape as a Script
The first engine tests whether some market sequences are usefully represented as choreography: a quiet build, a period of pressure, a release, and a hand-off. Each modeled campaign has a named structure, an implied counterparty role, and a possible counter-reading. Those labels summarize observed sequences; they do not prove that an institution planned a campaign or targeted a victim.
- Scene reading. It classifies each interval of tape into a small vocabulary of scenes — states of behavior rather than price levels.
- Sequence scoring. It scores the recent run of scenes against the known grammars of several campaigns, using a sequence model with recency weighting, to estimate which campaign is playing and how far it has progressed.
- Trap awareness. It watches for the tells of misdirection — isolated moves designed to look like a breakout or breakdown, inserted to bait a reaction before the real script resumes.
The output is not buy or sell. It is a reading: which script, how confident, how complete, and whether the last move looked like a genuine development or a planted decoy. We do not disclose the campaign catalog, the scene vocabulary, or the grammars — those are the engine.
3. The Law That Broke Our Own Engine: Named Plays Decay
Here is the most useful thing in this paper, and it is a failure of our own. A decoder built on a fixed catalog of named plays is, by construction, a set of learnable templates — and Papers I–III predict exactly what happens to learnable templates: they get manufactured and they decay. We observed this directly. Over time, the named-play readings from this engine lost almost all of their value to the downstream models.
The loss of value was not explained by the coding checks we ran. One interpretation consistent with the result was adversarial adaptation: later sequences no longer matched the fixed grammar and may have reflected changing behavior, crowding, or deliberate imitation. The data did not identify which mechanism caused the drift. The defensible conclusion was narrower: a static catalog of named patterns had become a snapshot of a moving target.
4. The Fix That Generalizes: Absolute Fit and Novelty
The repair is more interesting than the break, because it generalizes to any pattern-recognition system operating against an adaptive opponent. Our decoder had a subtle, fatal design property: it only ever measured relative fit — which of the known scripts is the least-bad explanation of the tape. Forced to choose, it would always nominate something, even when the honest answer was "none of these." It had no notion of absolute fit — does any known script actually explain this at all? — and therefore no notion of novelty — this is clearly a deliberate campaign, but it matches nothing I know.
- Absolute fit asks whether the best-matching script genuinely explains the behavior, rather than merely winning a contest among poor options.
- Novelty is high exactly where a real campaign is unmistakably running and the named catalog is blind to it.
The consequence is the important part. In the reported historical tests, information that had drained out of "which named play is this?" reappeared when the engine measured poor absolute fit and novelty. That result supported a model of structured but uncatalogued behavior; it did not prove authored intent. The durable research question became whether observed structure resists the current catalog, not whether the system can name an actor's plan.
5. The Chess Engine: The Market as a Game Between Desks
The second engine abandons the single-actor script frame for a multi-agent game. Its testable premise is that some tape behavior can be represented as interacting flows — labeled as coordinating, baiting, or defending — and that those labels may add context. The labels do not verify the number, identity, or intent of actual desks.
- Coordination. Whether bursts of activity are independent or clustered in a way that implies desks moving together — the difference between a crowd and a conspiracy.
- Bait versus genuine. Whether an aggressive move carries the cost signature of real commitment or the cheap signature of a feint meant to be faded.
- State of play. A compact classification of the adversarial situation — quiet, probing, coordinated, baiting, trapping — because the same surface move means opposite things in different game states.
- Anticipation. A forward read of the likely next response, conditioned on that state.
Again, the engine emits a description of the opponent's posture, not a trade. It is an interpreter of intent, and it is wired as an overlay whose influence must be earned.
6. Two Questions Almost No System Asks
- Is there a human behind the machine? Desk automation and a human overriding that automation leave different fingerprints — in the regularity of timing, the entropy of sizing, the consistency of conviction. Distinguishing "an algorithm is running" from "a person has taken the wheel" changes how much a move should be trusted, because a human override often signals genuine, discretionary intent a purely statistical view would miss.
- Is our own model being poisoned? An adversary who knows some model is watching can feed it deliberately corrupted inputs — artificial price texture, drifting features, planted bait patterns — to degrade or mislead it. A sensor that watches for the signature of manufactured inputs lets the system distrust itself in exactly the conditions engineered to fool it. This is the defensive twin of Paper III's Oracle: not "should I trust this prediction," but "is my input stream itself being gamed?"
We describe these as questions, not as mechanisms. How each fingerprint or poisoning signature is computed is not disclosed.
7. Offense Gated by Understanding, Again
The chess engine has an explicitly offensive mode — the point at which reading the game becomes acting on it. It is built on the same rule Paper III made structural: understanding gates action. The engine spends most of its time observing, accumulating conviction silently, and it is permitted to strike only after the state of play is genuinely understood. An opportunity that arrives before understanding is treated as bait by construction. A reader of the opponent that acted on every apparent opening would simply be the easiest desk at the table to trap; patience is not a temperament here, it is an architectural constraint.
8. Earning Influence: Engines as Overlays, Not Oracles
Neither engine is trusted on arrival. Both begin as zero-authority overlays, and any later influence must be earned from fresh, version-matched paper outcomes under the same champion/challenger discipline described across the series. An interpreter of an opponent is itself a hypothesis; elapsed time and sufficient resolved evidence, not architectural cleverness, decide whether it earns further testing.
9. What Elevated the Models (Shape Only)
- Novelty recovered edge that named-play recognition had lost. The migration from "which known play" to "authored intent that resists naming" restored an independent signal precisely where the rigid catalog had gone silent — the clearest example in our work of an evolving question outliving a static answer.
- Game-state and coordination reads were additive as context, not as triggers. They earned their keep by changing the interpretation of other signals across adversarial states, consistent with the cascade philosophy of Paper III.
- Human-behind-the-machine and model-poisoning reads acted mostly as vetoes. Their value showed up disproportionately in avoiding bad decisions in engineered conditions — risk avoided, not risk taken.
None of this claims the opponent has been solved or even identified. The historical result was that authored-behavior features added information under the reported tests. That finding did not establish the proposed mechanism, current-model lift, or live profitability.
10. What This Paper Deliberately Does Not Disclose
As throughout the series, we describe roles and shapes, not instruments. We do not disclose the campaign catalog, the scene vocabulary, or the sequence grammars; the absolute-fit and novelty computations; the chess engine's coordination, bait, game-state, or anticipation mechanics; the fingerprints that distinguish human from algorithm or the signatures that reveal model poisoning; the conditions that license the offensive mode; or how engine weights are learned from outcomes. The value of these engines is not that they read the opponent — it is how they read the opponent, and that is the moat.
11. Conclusion
A signal system sees a curve. These engines test whether an author-and-game metaphor adds useful context without pretending the metaphor is observed fact. The clearest lesson came from the decoder's failure: a fixed catalog went stale. The surviving discipline is not to claim knowledge of an opponent's mind, but to detect when observed structure no longer fits the catalog and send that mismatch back through falsification.