User: "shouldn't mention the names of the mind-clones, should merge all
analyses and write like a best-seller author would, about what happened."
Voice rewrite (prompts/case-writer.md):
- Reference voices: Erik Larson, Sam Kean, John McPhee, Mark Bowden.
Plainspoken non-fiction, scene-driven, fascinated.
- One narrator. NEVER say "Sherlock Holmes argues" / "Sun-Tzu builds
the case" / "the team concluded". No internal-process names reach
the reader.
- Hook the first paragraph. Open in a scene with a date, place, and
person doing something specific. NOT "This case investigates..."
- Show, don't argue. Verbatim quotes stay source-language in
blockquotes; the narration around them is the narrator's voice.
- Every claim cites a chunk with [[doc-id/pNNN#cNNNN]].
- Forbidden ceremony: "In summary…", "Em suma…", "Ultimately…",
"It is worth noting…", detective names, probability tables,
hypothesis tournaments.
- The honest unknown is the subject, not a failure: "Whatever was in
the sky over Sandia in December 1948, the government never said."
- 4-6 numbered scenes, each title-cased specifically ("The Green
Sphere Over Highway 60" not "Background").
- Bilingual EN + PT-BR per CLAUDE.md §3 — sections alternate, no
mid-paragraph language mixing.
- Refusal: emit INSUFFICIENT_ARTEFACTS rather than padding when the
corpus is thin.
Raw-material pipeline (src/detectives/case_writer.ts):
- hybridSearch(topic, lang, top_k=18) gives the narrator real corpus
scenes with verbatim text + chunk_id citations + bbox metadata.
This is what was missing — v1 only saw pre-digested hypothesis
artefacts, which is how the academic prose got there.
- Dropped the hypotheses + contradictions queries from the loader.
They were skeptic-framing scaffolding that doesn't belong in the
raw material a best-seller narrator works from.
- New buildPrompt sections: "Primary-source scenes", "Curated
verbatim quotes", "Anomalies and surprises", "Named witnesses".
Anomalies (Taleb's outlier gaps) reframed: drop dominant_model
skeptic baseline, keep title + why_surprising as gold material.
- Refusal floor: < 4 scenes from hybridSearch → skip with reason.
Co-Authored-By: Claude Opus 4.7 (1M context) <noreply@anthropic.com>
Live failure surfaced by user feedback: Poirot wrote a low-credibility
verdict on J. Edgar Hoover (W-0002) based on 1 actual chunk and 11
entity_mentions false positives where 'DIRECTOR'/'DIRETOR' was linked to
him by mistake. Poirot's own bias_notes correctly identified this — yet
still produced a verdict. Published on a 'Disclosure Bureau' site, that's
libellously misleading.
Deleted W-0001 (Donald Keyhoe) and W-0002 (J. Edgar Hoover) from
public.witnesses + their .md files.
Prompt rewrite (prompts/poirot.md):
- New "What counts as testimony" section up front, before discipline.
Direct testimony = the person AUTHORED, was QUOTED verbatim with
attribution, or GAVE testimony in a recorded hearing. Not: third-
party mentions, generic title appearances ('Director'/'Diretor'
that entity-extraction speculatively linked), CC lines.
- HARD FLOOR rule: emit `direct_testimony_chunk_ids[]`. If < 3, refuse
with INSUFFICIENT_TESTIMONY. For famous historical figures
(Wikipedia-worthy public figures) the floor is 5.
- Bias claims MUST cite a specific chunk; ungrounded bias claims drop.
- Tone: "careful prosecutor preparing a brief, not debunker scoring
points."
Defense in depth (poirot.ts):
- Detective enforces the same floor before calling writeWitnessAnalysis,
using a FAMOUS slug list (j-edgar-hoover, donald-keyhoe, j-allen-
hynek, curtis-lemay, vannevar-bush, eisenhower, truman, kennedy,
ted-bloecher, ...).
- When the floor isn't met, emit `poirot_refused_floor` audit event +
skip with reason like `insufficient_direct_testimony_1_of_5`.
- Sentinel parser now also catches INSUFFICIENT_TESTIMONY when it
appears on the first line of an otherwise-prose response.
Co-Authored-By: Claude Opus 4.7 (1M context) <noreply@anthropic.com>
Two complaints in one wave:
(W4.1) User: "Não pode ter vícios de IA como uso excessivo de '-' que a IA
coloca geralmente no lugar de vírgulas por exemplo. Isso deve fazer parte
do prompt geral."
- New prompts/_house-style.md banning the 9 most common AI prose tells
in both EN and PT-BR:
1. Em dashes as comma replacements (—)
2. Rule-of-three lists ("concrete, rigorous, and grounded")
3. Conjunctive openers ("Moreover", "Notably", "Ademais")
4. Superficial -ing analyses ("marking a shift", "destacando")
5. Inflated symbolism + AI vocab (tapestry, navigate, delve,
underscore, robust, multifaceted, marco histórico, ...)
6. Negative parallelisms ("Not just X but Y")
7. Vague attribution ("Some scholars say...")
8. Summary closers ("In summary...", "Em suma...")
9. Hedging fluff ("It's important to note...")
Verbatim chunk quotes are explicitly exempt; preserve as-is.
- claude.ts callClaude() lazily loads _house-style.md once per process
and PREPENDS it to every detective's system prompt:
composedSystem = houseStyle + "---" + detective.systemPrompt
This means all 7 detectives + future ones get the rules without any
per-prompt change.
(W4.2) User: "Quando entra em uma página da investigação não tem como
voltar! UX terrível!"
- New <BureauNav> sticky topbar with explicit "← home" + "🔎 bureau"
buttons + clickable breadcrumb trail. Always visible at the top of
every bureau page so the user can escape in one click.
- Wired into /bureau, /h/[hypothesisId], /c/[slug], /jobs/[id]. Each
page passes its sensible parent crumb (/bureau#hypotheses,
/bureau#reports, /bureau#jobs).
- Replaces the previous plain-text "disclosure.top / hypothesis /
H-0004" line which had no visual affordance.
Co-Authored-By: Claude Opus 4.7 (1M context) <noreply@anthropic.com>
User flagged that the bureau was emitting English-only output, violating
the project's bilingual rule. Every narrative field now ships in both
languages: stored in sibling DB columns + rendered as adjacent markdown
sections per CLAUDE.md §3.
Migration 0007 (apply as supabase_admin):
- public.hypotheses +question_pt_br, +position_pt_br,
+argument_for_pt_br, +argument_against_pt_br
- public.contradictions +topic_pt_br, +notes_pt_br
- public.witnesses +access_to_event_pt_br, +bias_notes_pt_br,
+verdict_pt_br
- public.gaps +description_pt_br, +suggested_next_move_pt_br
- public.evidence: unchanged (verbatim_excerpt stays source-language)
- JSONB siblings inside contradictions.chunks + gaps.scope handled at
runtime (statement_pt_br, title_pt_br, dominant_model_pt_br,
why_surprising_pt_br, what_it_implies_pt_br).
Detective prompts (all 7) rewritten with explicit bilingual JSON contract:
- Output protocol section names every EN field + its _pt_br sibling
- "Bilingual is mandatory" warning in the task instruction
- Sentinel skip-states unchanged (NO_HYPOTHESES, NO_CONTRADICTIONS,
INSUFFICIENT_TESTIMONY, INSUFFICIENT_HYPOTHESIS, NO_OUTLIERS,
NO_NEW_EVIDENCE, INSUFFICIENT_ARTEFACTS)
- Schneier: parallel arrays — hidden_assumptions[i] matches
hidden_assumptions_pt_br[i], lengths must match
- Case-Writer: interleaved §1 (EN) / §1 (PT-BR) per act in the body
Writer-side validation (all 7 tools):
- Reject INSERT if PT-BR sibling missing when EN field is set
- Persist both languages atomically in one INSERT (no half-updates)
- Markdown renderers write adjacent EN+PT-BR sections in case files
(## Argument for (EN) followed by ## Argumento a favor (PT-BR), etc.)
Detective parse layer (all 7 detectives):
- Coerce both keys from JSON output
- "incomplete_bilingual_*" skip reason when either side missing
- Defensive: PT-BR fields trimmed + length-capped same as EN
Orchestrator propagates question_pt_br + topic_pt_br through job payload
to runHolmes / runCaseWriter, mirroring the chat-tool entry point.
Web (UI):
- /api/jobs/[id] hydrates _pt_br siblings from pg
- job-status-poller HypothesisCard: PT-BR primary, EN in <details>
fallback when both exist
- ContradictionCard: PT-BR statement primary + secondary EN quote
- WitnessCard: PT-BR verdict primary + secondary EN quote, panels in PT
- GapCard: PT-BR title/why/implies primary
- /bureau hub: SELECTs both columns, renders PT-BR primary
- /h/[id]: ArgumentPanel renders PT-BR primary with collapsible EN
fallback when both exist
- BureauSnapshot homepage: position_pt_br / topic_pt_br / verdict_pt_br
primary
- DocBureauPanel /d/[doc]: same primary-PT-BR pattern
- New web/lib/i18n/pick.ts helper (unused yet by chat/agents — kept
for future locale-driven switching when both languages are equally
full; current rule is PT-BR-first since the user is brasileiro)
Co-Authored-By: Claude Opus 4.7 (1M context) <noreply@anthropic.com>
Brings the bureau from 4 → 8 detectives. All eight run as Bun + claude-CLI
subprocesses against the same Supabase + investigation_jobs LISTEN/NOTIFY
queue, sharing search.ts hybridSearch and writer-side validators that
gate writes against schema + FK.
New detectives:
Poirot (witness_analysis)
- prompts/poirot.md — credibility / access / bias / corroboration /
verdict; uses entity_mentions JOIN chunks to pull 12 chunks per
person; resolves corroboration_refs chunk_ids defensively (accepts
bare cNNNN even when the model emits pNNN/cNNNN).
- INSERT into public.witnesses with W-NNNN naming.
- Tone: purple (#9b5de5).
Taleb (outlier_scan)
- prompts/taleb.md — "surprise is relative to a model"; at most 3
outliers; each requires explicit dominant_model + why_surprising +
what_it_implies; fan-out into public.gaps with scope.kind="outlier".
- Same unscoped-fallback as Dupin (Pass 1 with doc_id, Pass 2 widens
to corpus if hits < 3).
- Tone: yellow (#ffd23f).
Tetlock (calibrate_hypothesis)
- prompts/tetlock.md — honest Bayesian update; emits new_posterior +
Δ + recommended_action ∈ {keep, downgrade, upgrade, supersede}.
- write_calibration UPDATEs public.hypotheses + APPENDS a
"## Calibration history" section to the H-NNNN.md case file
(calibration is append-only — each datapoint matters). Posterior
band auto-corrected to match Tetlock thresholds.
- NO_NEW_EVIDENCE sentinel handled; pure 'keep' with |Δ|<0.005 only
touches updated_at + reviewed_by.
- Tone: teal (#26d4cc).
Case-Writer (case_report)
- prompts/case-writer.md — Dr. Watson assembles all artefacts
(E-NNNN, H-NNNN, R-NNNN, W-NNNN, G-NNNN) into a five-act narrative.
ILIKE filter on topic; doc_id optional scope.
- Larger budget cap (≥ $0.50) + longer timeout for prose generation.
- Writes case/reports/<slug>.md with frontmatter (topic + counts);
no DB table for v0.
- New page /c/[slug] renders the report via MarkdownBody + stat chips.
- Tone: gold (#e0c080).
Hardening across the bureau:
- Sentinel parsing now accepts backticked AND prose-trailing forms
(Holmes NO_HYPOTHESES, Dupin NO_CONTRADICTIONS, Schneier
INSUFFICIENT_HYPOTHESIS, Poirot INSUFFICIENT_TESTIMONY, Taleb
NO_OUTLIERS, Tetlock NO_NEW_EVIDENCE, Case-Writer
INSUFFICIENT_ARTEFACTS). Avoids the failure mode where the model
refuses honestly but the runtime treated it as a parse error
(observed live with Poirot+Hoover identifying the DIRECTOR
false-positive disambiguation issue in entity_mentions).
Chat tool extensions (web/lib/chat/tools.ts):
- request_investigation now accepts 7 kinds. Each routes to its
detective with appropriate validation (hypothesis_id regex,
person_id kebab-case, topic non-empty, doc_id for evidence_chain).
- ETA per kind: Holmes/Dupin 60s, Poirot 45s, Schneier/Tetlock 30s,
Taleb 50s, Case-Writer 180s (longer prose), Locard 30×n_chunks.
UI integration:
- chat-bubble inline card paints each detective in its tone color.
- /jobs/[id] page header swaps name/subtitle/tone per detective;
question label adapts ("Topic" / "Hypothesis under attack" /
"Witness under analysis" / "Topic to outlier-scan" / "Hypothesis
under recalibration" / "Case to assemble").
- job-status-poller renders: case-report link card (gold), outlier
cards (yellow), witness cards (purple) — alongside existing
hypothesis, evidence, contradiction cards.
- /api/jobs/[id] hydrates witnesses (JOIN entities for canonical_name)
+ gaps (with scope JSONB).
- /c/[slug] page reads /data/ufo/case/reports/<slug>.md and renders
with MarkdownBody, frontmatter parsed for stat chips.
Co-Authored-By: Claude Opus 4.7 (1M context) <noreply@anthropic.com>
Adds the fourth AI detective in the Investigation Bureau runtime: Bruce
Schneier, who attacks an existing hypothesis as a red-team operator.
Runtime:
- prompts/schneier.md — discipline (don't disprove, just attack;
structured output with hidden_assumptions, failure_modes,
alternative_explanations, recommended_tests, verdict_one_sentence;
severity ∈ {low, medium, high}; emit INSUFFICIENT_HYPOTHESIS when
the input is too thin)
- src/detectives/schneier.ts — reads the hypothesis row + evidence
chain (joined via evidence_refs FK), feeds Claude with the
arguments + verbatim quotes, parses strict JSON object
- src/tools/write_red_team_review.ts — UPDATEs hypotheses.reviewed_by
+ updated_at; APPENDS (or replaces if re-reviewed) a structured
"## Red-team review (Schneier · X severity)" section to
case/hypotheses/H-NNNN.md. Caps each list at 5 entries × 240 chars,
validates verdict ≤ 280 chars.
- orchestrator: new `red_team_review` kind dispatching to runSchneier
Chat + UI:
- request_investigation gains kind=red_team_review + hypothesis_id arg
(validated against H-NNNN regex); detective auto-resolves to schneier
- chat-bubble inline card paints Schneier in red (#ff3344)
- /jobs/[id] page swaps title/subtitle/tone per detective; the
"Question" label becomes "Hypothesis under attack" for red_team_review
New /h/[hypothesisId] page (hypothesis dossier):
- Server-rendered from public.hypotheses + public.evidence (joined
via evidence_refs FK + chunk lookup)
- Header: ID + creator + reviewer (highlighted when Schneier has
visited), position as headline, question subtitle, Tetlock band
- Prior + posterior bars with Δ-delta indicator
- Argument grid: argument_for (green) vs argument_against (pink)
side-by-side with [[wiki-link]] auto-linking to source chunks
- Evidence chain: each E-NNNN with Grade A/B/C badge, verbatim
blockquote, link to source page
- Red-team review panel: parses the markdown section in the case
file (severity badge, verdict, 4 bullet panels for
hidden_assumptions / failure_modes / alternative_explanations /
recommended_tests). Empty state when not yet reviewed.
RedTeamRequestButton client component + POST /api/h/[id]/red-team —
authenticated user can trigger Schneier in one click; UI swaps to
"acompanhar" link to /jobs/[id] once queued.
Co-Authored-By: Claude Opus 4.7 (1M context) <noreply@anthropic.com>
Adds the third AI detective in the Investigation Bureau runtime: C. Auguste
Dupin, who scans a corpus shortlist for pairs (or small groups) of chunks
that cannot both be true under any ordinary reading.
Runtime:
- prompts/dupin.md — discipline (no contradiction without ≥2 distinct
chunk_ids; reject same-vocabulary near-misses; FEW high-confidence
over MANY weak ones; emit `NO_CONTRADICTIONS` when corpus is silent)
- src/detectives/dupin.ts — hybridSearch with k=18 (more chunks than
Holmes because contradictions emerge from comparing dispersed
claims), strict JSON-array parsing, AT MOST 3 contradictions per call
- src/tools/write_contradiction.ts — validates topic + ≥2 positions
drawn from ≥2 distinct chunks, resolves chunk_pk via DB lookup
(rejects positions citing unknown chunks), INSERTs into
public.contradictions + writes case/contradictions/R-NNNN.md
- orchestrator: new `contradiction_scan` kind dispatching to runDupin;
payload { topic, doc_id?, lang?, context_chunks? }
Chat + UI:
- request_investigation gains kind=contradiction_scan + topic arg;
triggered detective auto-resolves to dupin
- chat-bubble inline card renders dupin in orange (#ff8a4d) to
distinguish from holmes (cyan) and locard (green)
- /jobs/[id] page swaps title + subtitle + tone per detective;
"Question" label becomes "Topic" for contradiction_scan
- /api/jobs/[id] hydrates public.contradictions when outputs[] surfaces
contradiction_ids
- job-status-poller renders ContradictionCard: topic + N positions
(verbatim statements quoted, stance label optional, link to source
chunk) + optional notes panel, with resolution_status badge
(open/resolved/irreconcilable)
R-NNNN shares the contradiction_id_seq slot with relation per
CLAUDE.md naming — same conceptual class (a connection between two
pieces of evidence in tension).
Co-Authored-By: Claude Opus 4.7 (1M context) <noreply@anthropic.com>
Adds the second AI detective in the Investigation Bureau runtime: Sherlock
Holmes, who builds 2-3 rival hypotheses with calibrated priors + posteriors
against a corpus shortlist.
Pipeline:
1. hybridSearch() grounds Holmes with 8-15 chunks via the same
hybrid_search_chunks RPC the web uses (BM25 + dense + RRF). Default
max_dense_dist=0.55 (runtime favors recall over precision; web's
/api/search/hybrid stays at 0.40 for chat).
2. claude-sonnet-4-6 emits a strict JSON array with position +
argument_for + argument_against + prior + posterior + confidence_band
+ evidence_refs. Citations use [[doc-id/pNNN#cNNNN]] wiki-links.
3. writeHypothesis() validates posterior ∈ [0,1], auto-corrects the
Tetlock band from the posterior (high ≥0.90, medium 0.60-0.89,
low 0.30-0.59, speculation <0.30), checks evidence_refs FK against
public.evidence, INSERTs into public.hypotheses + writes
case/hypotheses/H-NNNN.md.
Discipline guarantees (prompts/holmes.md):
- posteriors across rivals sum to ≈1.0
- no claim without chunk citation
- prefer lower band when ambiguous (anti-inflation)
- declarative one-sentence position, no hedging
- emit `NO_HYPOTHESES` when corpus is silent (refuses to fabricate)
Smoke test (Sandia green fireballs 1948-49):
- H-0001 prior 0.5 → posterior 0.2 (speculation): natural meteoric
- H-0002 prior 0.3 → posterior 0.4 (low): classified weapons / tests
- H-0003 prior 0.2 → posterior 0.4 (low): genuinely unidentified
Bayesian update visible: "natural meteoric" prior dropped 60%; both
rivals climbed. 4 unique chunk citations across the 3 hypotheses.
orchestrator dispatches `hypothesis_tournament` kind via runHolmes;
job marked `failed` if all rivals error, `complete` otherwise.
Co-Authored-By: Claude Opus 4.7 (1M context) <noreply@anthropic.com>
Luiz Gustavo