Triangulate Phase 0 Retrospective: 5-Topic Disconfirming Experiment, May 2026

Internal synthesis. The full working document with per-topic detail lives at ~/Projects/personal/triangulate/phase-0-findings.md and the per-topic articles, claim graphs, and audit logs are in the same repo under articles/phase-0/, claim-graphs/phase-0/, and audit-logs/phase-0/. This source is the stable summary of the lessons.

What Phase 0 was

A disconfirming experiment to test the triangulate newsroom protocol before any code was written. The protocol applies the operational checklist from [[JOURNALISTIC_STANDARDS]] via four phases (Reporter → parallel checkers → render-with-House-Style + sequential gates → audit log) to produce publication-grade articles with reconstructable editorial reasoning.

The experiment ran five representative topics manually through Claude, varying both editorial register and vault substrate density to probe the protocol's failure modes.

The five topics

1. Topic 1: California AB 2624 status report (May 2026) — politically-charged + vault substrate. Event-status shape. Political-news register. Required a v2 House-Style pass; surfaced the architectural lesson that House-Style must be baked into render, not run as a post-render gate.
2. Topic 2: Minnesota childcare-funding freeze, four-month status — politically-charged + vault substrate. Multi-event status shape with parallel litigation. Political-news register. Highest-value F3 catch: case-conflation between S.D.N.Y. (Broderick) and N.D. Cal. (Thompson) parallel cases.
3. Topic 3: TSMC → Intel re-rate mechanism (May 2026) — non-political + rich vault substrate. Causal-thesis shape. Financial-thesis register — new register class. Highest-value F1 catch: Samsung 2nm yield conflict between vault's 70% (January) and TrendForce April 14 report of 55%. Real-time calibration event surfaced.
4. Topic 4: Pennsylvania Medicaid fraud (thinner-substrate test) — politically-charged + zero PA-specific vault substrate. Political-news register. Highest-value F3 catch: OIG "unallowable capitation payments" ≠ fraud (administrative-error-vs-fraud distinction). The critical test of whether the protocol survives a vault-thin topic.
5. Topic 5: Bronze Age Collapse, current scholarly state — outside-domain + production-pipeline shape. Serious-history feature register — new register class. Used /autoresearch --auto to build substrate first, then ran the newsroom protocol on the populated vault. Highest-value F3 catch: revisionism-overshoot (Millek's 61% misdated destructions ≠ "the collapse didn't happen").

The four go/no-go questions resolved

Q1 — Do F1/F3 checks catch real errors? YES, across all five topics. Total: 27 F1 + 24 F3 = 51 documented catches.

Q2 — Is the audit log reconstructable by a second reader? YES, across all five topics. The audit-log architecture (per-claim catch tables, source-class bias matrix, fetch-outcome appendments, gap list, standards-checklist trace) does not need to change before any UI work.

Q3 — Does the draft clear "send to a friend"? YES (with one v2 pass on Topic 1; single-pass news-coded output for Topics 2–5). House-Style discipline baked into render, not run as a post-render gate. Three editorial registers validated.

Q4 — Does the brain vault have enough substrate? Depends — and the question reframes. Vault-carries percentages across the five topics: 30% → 50% → 80% → 0% → 95%. Protocol quality stayed flat across that range.

The architectural finding (Q4 reframing — load-bearing)

The vault has two separable functions:

1. Substrate — topic-specific knowledge (entity / concept / mechanism / question pages) that strengthens an article when it exists.
2. Methodology — [[JOURNALISTIC_STANDARDS]], [[SOURCE_RELIABILITY]], triangulation posture in SCOPE.md files, typed page schemas, the parallel-checker discipline. The methodology layer generalizes to any topic and is load-bearing on Q1–Q3 quality even when topic-specific substrate is zero.

The Plan's explicit pivot-point question — "If the brain vault is too thin per topic → newsroom becomes an autoresearch wrapper" — is confirmed false. The vault-as-methodology layer is what differentiates the newsroom from an autoresearch wrapper. The vault-as-substrate layer is what strengthens a given article when it exists. Both contribute; they contribute differently.

The user-articulated cross-domain-pattern thesis

After Phase 0 completed, the user articulated the conclusion the experiment was designed to surface:

"Going through the process of building brain piece by piece has led to the ability to actually publish very rich useful articles on the other side. It feels like the more dense we make the brain across a wide range of topics the more we will be able to pull out for our journalistic purposes. We can analyze causal chains not just for the purposes of the stock market but also for mapping events through history. The patterns we are instantiating are cross applicable across domains."

The brain's typed primitives (entity / concept / mechanism / question per vault/_meta/RESEARCH.md) and standards discipline are domain-general. The same mechanism page format that captures [[tsmc-saturation-to-intel-anchor-stack]] captures [[lba-multi-stressor-collapse-chain]]. The same F3 misinterpretation discipline that caught case-conflation on the MN childcare freeze caught revisionism-overshoot on Millek's destruction-horizon audit.

Brain density across topics compounds into newsroom drafting power.

Transferable F3 misinterpretation patterns

Five recurring F3 patterns surfaced across the experiment. Each is transferable across topics; each should be encoded in the Misinterpretation Editor agent's prompt when Phase 2 builds it.

• Case-conflation (Topic 2) — treating parallel processes (e.g., two separate federal cases in different courts) as a single process. Verb-strength tell: "the court ruled" when it was a court, not the court.
• Administrative-error-vs-fraud (Topic 4) — collapsing an OIG integrity finding ("unallowable payments") into "fraud." Critical for any state-Medicaid coverage; the federal government distinguishes integrity findings from MFCU prosecution data, and so should the wiki.
• Revisionism-overshoot (Topic 5) — treating a scholar's revision of a phenomenon's scale ("61% of claimed destructions are false") as a dismissal of the phenomenon ("the collapse didn't happen"). Transferable to any topic where revisionism is mistakable for repudiation.
• Preliminary-vs-chose (Topic 3) — verb-strength control on stage-of-commitment claims. "Apple chose Intel" is materially different from "Apple and Intel reached a preliminary agreement." Highest-frequency F3 pattern for financial-thesis register.
• Anchor-roster collapse (Topic 3) — presenting multiple commitments of different formality as homogeneous. Intel's four named anchors (AWS multi-billion 18A / Microsoft Maia 2 18A-P / Apple preliminary / Terafab 14A) are at materially different stages; flattening them is an F3 collapse.

Three editorial register classes validated

• Political-news (Topics 1, 2, 4) — short attribution sentences, AP-style hedging, "documentary vs punitive" framing for federal-pressure stories, careful R/D bias-flagging, methods sentence + link at end.
• Financial-thesis (Topic 3) — thesis-statement ledes, explicit conditional reasoning, falsification-triggers in-body (not in audit log alone), verb-strength control on stage-of-commitment claims, position-disclosure as publication-blocker for any external publication.
• Serious-history feature (Topic 5) — lede framing the scholarly state of the field rather than the dramatic event, "settled vs open" structure, popular-vs-scholarly framing flagged, primary-source naming with dates.

The House-Style Editor will need exemplar corpora for each register class. Register selection should be driven by topic class, not produced as a single style across all contexts.

The four-phase Application Protocol (validated)

The protocol as applied:

1. Phase 1 — Reporter agent. Reads topic + brain-vault substrate. Identifies gaps. Optionally invokes /autoresearch for missing primary documents. Produces a typed claim graph: array of records {id, text, sources, evidence_class, subjects, inferences}.
2. Phase 2 — Parallel checkers on the claim graph. F1 (objective error), F3 (misinterpretation), Section C (bias framing), Section E (proportionality), Section H (citizen-journalism specific-claim verification) run independently on claim IDs. Each emits findings: {claim_id, finding_type, source_text, draft_text, action_required}.
3. Phase 3 — Render prose from approved claim graph WITH House-Style baked in, then run sequential gates. Section B (right-of-reply) on full draft. Section D (transparency/methodology) on methods section. Section G (harm minimization) on framing. House-Style is render-time, not post-render. Topic 1's v2-after-v1 result is what revealed that House-Style must be a render-time constraint, not a post-render gate.
4. Phase 4 — Audit log. Single deterministic pass from prior phase outputs. Per-claim catch tables (F1 / F3), source-class bias matrix, fetch-outcome appendments, gap list with follow-up actions, standards-checklist trace.

Architecture decisions confirmed

• Parallel-checker model justified; F1/F3 separation with per-claim emit catches errors single-shot drafting would not.
• Audit log schema correct; no schema rewrite needed before UI work.
• House-Style Editor is required and is render-time, not post-render gate.
• Autoresearch → newsroom handoff works end-to-end (Topic 5 validation): /autoresearch --auto builds substrate → clipping-promote routes → ingest-pending creates typed wiki pages → newsroom reads from the populated vault.
• Typed wiki pages are a higher-value drafting surface than raw text.
• Vault has two functions (substrate + methodology); both should be first-class in the packages/vault/ typed reader.

Source-reliability domains surfaced

Across the five topics' fresh fetches, the vault's [[SOURCE_RELIABILITY]] tracker gained:

• Reliable (new): oag.ca.gov, ag.state.mn.us, dcyf.mn.gov, govinfo.gov (Federal Register HTML mirror), localnewsmatters.org, trendforce.com, attorneygeneral.gov (PA AG), dailyitem.com, broadandliberty.com.
• Periodic-failure (new): factually.co (occasional 403), penncapital-star.com (one 403 today).
• Persistent-failure (new): acf.gov, www.hhs.gov, courthousenews.com, naeyc.org, www.cnbc.com.
• Hard-blocked (new): federalregister.gov (CAPTCHA redirect; mirror via govinfo.gov confirmed).

What this retrospective enables next

• Phase 1 of the Plan: build packages/vault/ typed reader in brain. The two-functions finding directly informs the API surface — substrate access (sources/wiki) AND methodology access (standards/source-reliability/scope) as separable concerns.
• Phase 2 of the Plan: newsroom MVP at ~/Projects/personal/triangulate/. The four-phase protocol is the implementation target; the transferable F3 patterns are inputs to the Misinterpretation Editor's prompt; the three register classes are inputs to the House-Style Editor's exemplar corpus.
• Ongoing: brain density across topics compounds into newsroom drafting power per the cross-domain-pattern thesis. Vault densification is itself a newsroom investment.