Ti Report

Reference files for rendering styled evaluation reports via scripts/render_report.py. Not directly invoked by Claude — consumed by the script layer only.

Files

Dual-Environment Rendering Contract

scripts/render_report.py builds two Jinja2 environments sharing the same FileSystemLoader:

from jinja2 import Environment, FileSystemLoader, select_autoescape

loader = FileSystemLoader(str(TI_REPORT_DIR))

html_env = Environment(
    loader=loader,
    autoescape=True,  # force escaping for this env
    trim_blocks=True,
    lstrip_blocks=True,
)

md_env = Environment(
    loader=loader,
    autoescape=False,
    trim_blocks=True,
    lstrip_blocks=True,
)

Pitfall 2 warning: Do NOT use select_autoescape(["html", "htm"]) — template.html.j2 ends in .j2 and would not be matched. The html_env must use autoescape=True directly to guarantee HTML escaping regardless of filename.

The CSS is passed as a template variable css:

data["css"] = (TI_REPORT_DIR / "styles.css").read_text()

The HTML template inlines it: <style>{{ css | safe }}</style>.

|safe is used only for the css variable. The CSS contains > combinators (.rh-tabs > input), " attribute selectors ([data-panel="user"]), and & characters; without |safe, Jinja2's autoescape mangles them into > / " / & and breaks the CSS-only tab switcher plus many other selectors. The CSS comes from a static file on disk loaded by render_report.py — it is never user-supplied, so the |safe bypass is secure.

All other interpolated prose (idea text, narrative fields, assumption text, research text, key findings, score explanations) flows through either autoescape or the linkify filter (which escapes first, then wraps URLs). The only raw HTML beyond css is the inline SVG icon sprite and theme-toggle script, which are trusted template literals, not data.

Data Contract

Template variables assembled by render_report.py from the run artifact family:

Key lookups used in templates:

• idea.codename, idea.icon, idea.subtitle — all optional; template renders sensible fallbacks when absent (codename → "Evaluation", icon → 📋, subtitle → idea.problem).
• numbers.rankings[i] — each entry now carries weight (0–1 fraction of registry weight), used to render {weight*100}% in the dim-grid and the per-dimension collapsibles.
• numbers.rankings[i].slug → key into dimensions dict for dimensions[r.slug].key_finding and dimensions[r.slug].score_explanation.
• numbers.verdict_bucket → one of GO|PIVOT|STOP — used to select the verdict pill color (green, yellow, red) and the .verdict-word class.
• numbers.verdict_label → constant string in the form "BUCKET — descriptor" (e.g. "PIVOT — Promising, address weak areas"). The hero-stats meta line slices off the bucket prefix and renders the descriptor tail next to the styled .verdict-word. There is no LLM-authored verdict rationale.
• numbers.rankings[i].evidence_strength → {both_confirmed, research_only, founder_only, assumed, grade}; grade is one of A+|A|A-|B+|B|B-|C|D|F.
• numbers.overall_grade → single letter grade summarizing evidence strength across all dimensions (same enum as per-dim grade). Drives the hero-stats .grade-big badge and the properties grid.
• numbers.evidence_totals → aggregate tier counts {both_confirmed, research_only, founder_only, assumed} summed across all valid dimensions. Drives the hero-stats evidence breakdown bar and its legend counts.
• numbers.reach_potential_blocks → per-dimension grouping {dim, slug, from, to, weighted_uplift, assumption_texts[]}, sorted by weighted_uplift desc. Drives the Reach Potential section.
• numbers.assumption_impact_math → [{assumption_text, dim, score_delta, weighted_uplift}]. Kept for the markdown fallback and downstream tooling; the HTML template prefers reach_potential_blocks.
• research.available → boolean gate for the Research Highlights tabs.
• research.clusters.{user,competitive,market}[i] → each item is now {text, source_url?, source_title?}. Rows with source_url render a clickable chip; rows without (pure synthesis) render a grey "synthesis" chip.
• strengths_weaknesses — the HTML template renders only the first strength + first weakness as risk callouts; the markdown template matches. Upstream ti-narrative still produces top-3 of each, so future designs can show more without a data change.