Transferable F3 Misinterpretation Patterns

Five recurring F3 (misinterpretation) patterns surfaced across the triangulate Phase 0 experiment. Each is transferable across topics; each should be encoded into the Misinterpretation Editor agent's system prompt when Phase 2 builds it. F3 catches were the highest-value class of error the protocol surfaced — most failure modes that distinguish triangulate's output from single-shot strong-prompted drafting fall here.

The patterns

Case-conflation

Treating parallel processes as a single process. Surfaced on Topic 2 (MN childcare freeze) — the protocol distinguished two separate federal cases: S.D.N.Y. Judge Vernon Broderick's preliminary injunction (the 5-state AG case, Feb 6 2026) from N.D. Cal. Judge Trina Thompson's preliminary injunction (the parallel labor-coalition case, April 1 2026). A single-shot draft would have treated the April 1 ruling as a continuation of the February 6 ruling.

Tell: "the court ruled" when there were multiple courts. Look for parallel-process indicators (e.g., separate filings, separate plaintiffs, separate jurisdictions).

Rule for the agent: Treat each judicial / regulatory / corporate process as distinct until proven to be the same proceeding. Use explicit phrasing like "a separate parallel case" when distinct.

Administrative-error-vs-fraud

Collapsing an OIG / GAO / inspector-general integrity finding into "fraud." Surfaced on Topic 4 (PA Medicaid) — the HHS OIG found Pennsylvania made $8.78 million in "unallowable capitation payments" because the state paid managed-care organizations as if duplicate Medicaid IDs were separate enrollees. That is an administrative-handling error, not a fraud scheme. A single-shot draft would have written "PA had $8.78M in Medicaid fraud" — which would be wrong.

Tell: OIG / GAO / IG language ("unallowable," "improper payments," "integrity finding") in the same paragraph as criminal-prosecution numbers.

Rule for the agent: Distinguish federal program-integrity findings from MFCU criminal-prosecution data. The federal government tracks them separately; the wiki should too. Use the OIG's own term ("unallowable payments") rather than substituting "fraud."

Revisionism-overshoot

Treating a scholar's revision of a phenomenon's scale as a dismissal of the phenomenon itself. Surfaced on Topic 5 (Bronze Age Collapse) — Jesse Millek's 2023 audit found that 61% of the 153 LBA destruction events traditionally attributed to ~1200 BCE were misdated, assumed on thin evidence, or false citations. A single-shot draft could collapse this into "the Bronze Age Collapse didn't happen" — which would be wrong. The collapse is real; its physical-destruction footprint was overstated.

Tell: A study showing X is overstated should not become "X is fake." Watch for "61%," "most claims," or "much of the literature" findings about scale.

Rule for the agent: When a revisionist source critiques scale, specify what the revisionist confirms AND what they debunk. Use language like "scale has been overstated" rather than "did not happen."

Preliminary-vs-chose (verb-strength control)

Treating preliminary commitments as final commitments. Surfaced on Topic 3 (TSMC-Intel re-rate) — the Apple-Intel deal was reported (Wall Street Journal, May 8 2026) as a preliminary agreement with no purchase orders. A single-shot draft would routinely produce "Apple chose Intel." This is the highest-frequency F3 pattern for financial-thesis register.

Tell: Subject-verb-object constructions where the verb implies more finality than the source supports.

Rule for the agent: Use the verb the underlying source uses. Default to "preliminary," "reported," "reached a preliminary agreement." Reserve "chose" / "selected" / "finalized" for final purchase orders or executed contracts.

Anchor-roster collapse

Presenting multiple commitments of differing formality as homogeneous. Surfaced on Topic 3 (TSMC-Intel re-rate) — Intel's four publicly-named anchor customers are at materially different stages: AWS (multi-billion 18A, most concrete), Microsoft (Maia 2 on 18A/18A-P, program-named), Apple (preliminary, no PO), Terafab (14A forward node). Flattening these into "four customers" loses load-bearing information.

Tell: Lists of commitments / agreements / partnerships presented in parallel grammatical structure without stage descriptors.

Rule for the agent: When listing commitments, include stage descriptors (preliminary, multi-billion, program-named, forward-node, etc.). Walk the list individually, not as a flat roster.

Why these are load-bearing

Single-shot strong-prompted LLM drafting makes these errors at high rates because they're each a small "tidying up" of underlying nuance toward something cleaner-sounding but materially wrong. The parallel-checker model surfaces them because the Misinterpretation Editor's narrow brief — check whether the prose says more than the source supports — is much more reliable than asking a single agent to "be careful."

Phase 0 surfaced these five patterns across five topics. There are almost certainly more. Future post-mortems should add to the catalog as new patterns emerge from operational catches.

Evidence

• From 2026-05-13-phase-0-retrospective: "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."
• From 2026-05-13-phase-0-retrospective: "Single-shot strong-prompted LLM drafting makes these errors at high rates because they're each a small 'tidying up' of underlying nuance toward something cleaner-sounding but materially wrong."
• From 2026-05-13-phase-0-retrospective: per-topic catch counts: Topic 1 (3 F3), Topic 2 (5 F3), Topic 3 (5 F3), Topic 4 (5 F3), Topic 5 (6 F3). 24 F3 catches total across Phase 0.

Related

• editorial-register-classes — verb-strength patterns vary by register
• triangulate-4-phase-application-protocol — where the Misinterpretation Editor sits in the pipeline
• vault-two-functions-substrate-and-methodology — these patterns are part of the methodology layer the wiki maintains