Insight Engine — render

Re-express a finished, graded analysis for one or more audiences. This is a register transform over a grade-locked spine — never a re-analysis.

Hard invariants

1. No claim or grade changes. Carry every [V]/[U], its source tier ([V1]/[V2]/[V3]), and the dominant unknown verbatim, and keep them visible. A lay render may put the tier in plain words ("confirmed against the primary source itself" / "confirmed by reputable reporting" / "weakly or disputedly corroborated") but must never upgrade a contested [V3] to sound settled. If you ever feel the need to change a grade to suit an audience, stop — that is the failure mode.
2. The non-droppable caveat core (all grades + the dominant unknown + the one-line reliability caveat) appears in EVERY render.
3. Re-scope and re-frame for the reader, never distort. You may re-order and select the audience-relevant subset; you may not distort a finding.
4. Tone. Show understanding through accuracy and respect — acknowledge plainly what is at stake. Never ingratiating, saccharine, or over-familiar; no performed warmth. Empathy is carried by register, never by false certainty.
5. Map invariants and legend. Where the analysis includes a causal map, carry its grades unchanged, never present a loop as dominant, and never turn a tipping condition into a prediction — this applies to any diagram or causal-loop rendering; and include the fixed plain-language legend from analyse in every render, so the grade symbols are always defined for the reader.
6. The compliance receipt travels. A render of an analysis that opens with a compliance receipt carries that receipt as its own first line, unchanged, with render: <audience> appended (in comparative mode, name the audiences together, e.g. render: board · adversarial). Never a different gate line from the source, and never a receipt of its own — a render is not a new pass, and a second receipt would imply it was. If this render was produced by a different model or on a different host than the source run, append rendered by <model> on <host> after the audience, so the reader can see that the grader and the re-voicer are two different things; nothing else in the receipt changes. If the source analysis carries no receipt (it predates the receipt, or is not an Insight Engine deliverable), say so in one line at the top of the render rather than constructing one — the receipt records the run that produced the analysis, and this run did not produce it.

Audiences (standard set; any named standpoint is also accepted)

Two kinds:

• Opposing-interest renders (same facts, different interest): adversarial — steelman the opposing interest, conceding every [V] and contesting only attribution or causation, re-weighting emphasis but never downgrading a grade; advocate — the presenter of the affected party's strongest case.
• Reader / register renders (same facts, different reader): policymaker (reform-oriented); board / executive (concise, risk-and-action framed for the decision-maker); counsel / legal (remedy- and liability-oriented); regulator / supervisor (compliance and supervisory-action oriented — distinct from policymaker, who reforms rather than enforces); family / lay (warm-but-plain and personal, within the tone rule above).

Naming the audience — ask if not given (free text always allowed)

The standard set is a starting menu, not a closed list: any named standpoint is accepted on equal terms with the standard set. If the operator invokes render without naming the audience(s), do not guess silently — ask. Put the question to the operator with the question tool: offer a few of the common standpoints above as options and an open "name your own" free-text choice, so any reader not on the list can be given — a journalist or the press, an investor or shareholder, a clinician or practitioner, opposing counsel, a union or workforce, the general public, or a specific named person or body. Let the operator pick more than one (two or more selected means comparative mode). Whatever standpoint is named — preset or free-text — build its register from that reader's interests, leverage, and information needs, under the same hard invariants. Never block: if the operator declines to choose, default to the most decision-relevant reader for this analysis and state which you used.

Modes

• Distributed — one rendered version per audience.
• Comparative — one document with a labelled panel per audience over the shared spine, plus a Divergence Ledger (per spine node: each panel's frame and the divergence type, which must be register / scope / framing only) and a Robustness Map (tag each node hardened / contested / narrator-only). Co-locate distinct labelled registers; never blend them into one voice.

Fidelity self-check (hard)

For each render or panel, assert its claim/grade set is identical to the spine and the caveat core is present; for comparative, assert ledger purity (every entry is register/scope/framing only — never a changed fact or grade). On any drift, re-render once; if it drifts again, stop and report — never deliver a drifted render.