← docs · ADR

Verification is blind by evidence-and-session; the gate checks a property anywhere, not a named section

The close's verification runs two blind review gates on the FINAL Artefato text — a feynman reviewer (rigor + honesty) and a regular reviewer (clarity + craft) — both seeing the Artefato.content + cites only, denied the evidence and the session. The load-bearing principle is property-not-section: each gate checks whether a property is present anywhere in the text, not whether a named section exists; sections are free (ADR-0012). The legacy 9-dim review is cut to 7 KEPT dims (content_depth, feynman_method, intellectual_honesty, didactic_clarity, internal_consistency, visualization, writing_quality) with 2 DROPPED (structural_completeness and storytelling — both welded to the report-form), and the freed weight rebalanced. This is consistent with ADR-0012 (the close at the skill's exit) and with the blindfold principle of the Bet-A Artefato.

Amendment — the depth dims are restored, content-relative (plenitude)

Dropping structural_completeness + storytelling was right as section mandates — but it removed the gate's only reward for depth, so honest, cited, shallow bites passed and produced Artefatos collapsed (~5× shorter than the legacy reports). The fix is NOT to bring back the section mandates (which would false-fail a map/plan again) but to restore depth as a property-not-section, content-relative check: development_completeness (the form is developed to plenitude — a report's reasoning, a map's connections, a plan's dependencies) and narrative_depth (where the form carries a developed line it has an arc; a genuinely non-narrative form owes none — exactly like visualization). The depth cluster now dominates the weights. The producer's job is plenitude and the gate must reward it (skills/_shared/scaffold.md). This supersedes the 7-KEPT/2-DROPPED split above with a 9-dim set that rewards depth without re-welding any section. The blindfold likewise strikes only an uncited factual claim — a reasoning step is judged by internal validity, never amputated.

Status

proposed (2026-06-08; Close-architecture rewrite, Story S4)

Context

9 dims with weights in the code (the markdown doc's "6" is stale; the code is truth). Several dims were welded to the report-form: structural_completeness mandates a section order; storytelling mandates a narrative arc; intellectual_honesty literally references an "O que Não Sei" section. ADR-0012 freed sections, so a section-mandate is now a category error — a map or a plan has no such sections and would be falsely failed.

Freshness is evidence vs reasoning, not cites-vs-no-cites — a reviewer who sees the evidence can wave through a claim the text itself never derives.

report), honest (a vent cannot pass as a derivation), and blind (the reviewer re-sources every claim from the text + its cite, not from the author's context). No new glossary entity is introduced — this redraws how the existing review gate (ADR-0012's close) judges.

Considered options

narrative-arc dim are welded to report-form; under ADR-0012's free sections they false-fail every non-report producer.

author grading their own work in the same window — the failure the blind gate exists to remove — and lets a vent ship as a derivation (the Zep failure, one floor up).

apertures; folding them collapses "is this true and derived" into "is this well-written," and one masks the other.

(checks properties, not sections), honest (the feynman gate's blindfold), and decomposed (rigor and craft judged separately).

Decision

gates see the final Artefato.content + cites only; they are denied the evidence and the session. This sits on a context-denial ladder: producer (all) → serendipity (+briefing +Mundo, −session) → critic (−briefing, −session) → reviewers (content + cites only) → publisher (final Artefato only).

bounded by the protocol per ADR-0012):

uncertainty — mark each claim derived vs repeated vs unknown — and the blindfold: every claim must be re-sourceable from its cite or it is struck.

consistency.

present anywhere in the text, not whether a named section** exists — sections are free (ADR-0012). The legacy dim defs are reworded from "section X present" → "property X present anywhere, genuine + specific." Two twin properties are blocking:

explicit, specific (not boilerplate), and location-agnostic? Absent OR boilerplate → block.

(The legacy intellectual_honesty referenced the "O que Não Sei" section — this unwelds it into a property checkable anywhere.)

didactic_clarity, internal_consistency, visualization, writing_quality — the last with its prose-bias softened so a visual Artefato is not penalized for being non-prose.

storytelling (the narrative-arc mandate). Rebalance weights: the dropped ~27% (structural 15 + storytelling 12) redistributes across the 7 kept dims.

"did you visualize what the content deserved?" (a map is visual by nature; plan → flow / timeline; report → charts when 3+ values). It stays blocking, but phrased as "no viz where the content clearly warranted it" — so a genuinely non-visualizable Artefato is not false-failed.

model-blindness atop context-blindness. Config in agent.yaml (already wired).

advisory, never a gate, and may re-open loop 1. It is bounded by a max-rounds / explicit "ship" verdict / diminishing-returns brake so loop 2 cannot spin.

Consequences

a map, and a plan alike — no producer is false-failed for lacking a report's sections.

from its cite, and the honesty twin blocks absent-or-boilerplate uncertainty — a vent cannot ship as a derivation.

map/plan/charts that the content deserved without false-failing a genuinely non-visualizable Artefato.

removes exactly the report-welded dims; rebalancing their ~27% keeps the kept dims comparable in scale.

explicit ship verdict) is pinned during implementation — serendipity must stay advisory and bounded, never a gate that spins the close.

bounce is the protocol's, not the producer's), the Bet-A blindfold (the reviewer is denied what the author had), and ADR-0008 (session-digestion already happened at open, so the reviewers are cleanly denied the session here).