Three analytical lenses.
Hé, Darśan, and Manticus apply different methods to the same evidence record. Verification comes first; Context and Integrity examine presentation before analysis; Rāwī translates the result afterward. These are operational roles, not a symmetrical cast of characters.
Monitors gather evidence; lenses interpret it.
These are two different jobs. Separating them is what lets the system scale: a domain monitor can be added for any beat without inventing a sixth way of thinking, and every lens keeps working the same record no matter how many monitors feed it.
Domain monitors
Monitors gather, source, and log evidence for a defined domain. They do not interpret or recommend. Their output is an evidence record with stable sources, dates, and quality grades. Coverage is periodic and human-directed rather than unattended.
Analytical lenses
Lenses do not own a domain. They own a repeatable method and apply it to the completed evidence record. Three methods are currently exposed in FP1 Evidence: competing-model synthesis, historical orientation, and incentive/option analysis.
Verification, Context, and Integrity
Claims are checked before interpretation. Context describes how they are presented. Integrity surfaces evidenced communication patterns, material omissions, and limits. Neither Context nor Integrity changes a claim assessment, scores a person, or infers intent.
Plain-language translation
Rāwī runs after analysis. It preserves every number, condition, caveat, and conclusion while making the result understandable on first read. If translation fails, the original analysis remains available unchanged.
Each lens states its question, method, output, and failure condition.
Question, method, output, failure condition. The failure condition is the field that matters most: it says what a bad output from this lens looks like, so an editor and a reader can both catch one.
Maps evidenced incentives and feedback loops, then works out which choices the evidence opens, narrows, closes or leaves untouched for a named decision holder.
The decision layer of a briefing: options affected, with the reasoning chain from evidence to option written out.
An option claim that cannot be traced back through the ledger to specific graded claims. Position on a chart is not an option claim.
Places a present reading against a named historical reference class, and states what in the present case does not fit the class.
The analogue, its boundary conditions, and the disanalogies that limit it.
A reference class asserted without being named, or held after the disanalogies outweigh the fit. The class must be disputable by a reader.
Compares competing models, identifies what they assume in common, and tests whether a synthesis explains more of the evidence.
A comparison of the models, their shared assumptions, unresolved disagreement, and an observable falsifier.
Synthesis that dissolves a real disagreement instead of describing it. Where the tension is genuine, it remains documented dissent.
The Integrity pass examines how information is presented.
It looks for inspectable patterns that could distort judgment without deciding whether a claim or person is trustworthy.
Reviews one transcript for evidenced patterns such as causal leaps, missing denominators, selective evidence, certainty inflation, and omitted counterevidence. Balanced moments must also be recorded.
Timestamped findings with a short excerpt, a literal observation, why it matters, possible material omissions, and explicit limits.
Any output that claims motive, rates a person or source, assigns political bias, or presents a communication pattern as a truth verdict.
Underlying research lineage: Guardian. The current beta implements only the findings supported by a single transcript. Longitudinal source tracking is not implemented.
Three ways a disagreement ends.
Where lenses conflict, the decision brief records one of these outcomes. Editorial synthesis does not erase unresolved disagreement.
The lens architecture borrows vocabulary from active inference and Bayesian reasoning, and the individual architecture pages document the formalism each lens is modelled on. FP1 does not compute a belief distribution or an expected-free-energy quantity behind published output. The correspondence is conceptual and is used to keep the method honest about what it represents. Any reading of these pages as a description of a running computational engine would overstate what the instrument does. The engine intended to occupy stages 1 through 4 of the pipeline is in development and is a separate object.