Company-dossier reader-test kit

This is the falsification harness Gate 2 needs to flip from building (25/25 counter target reached 2026-05-30) to passing. The gate's test, as stated in docs/strategy/gates.md, is:

> a dossier surfaces ≥3 actions a senior sector analyst missed in their own > scan, AND the dossier's exposure score holds up under a domain expert's > review.

Counter coverage alone does not flip the gate — the 2026-05-31 04:47Z tick note explicitly named this: "the next flip is external reader test, not more dossiers." This kit is the analog to Gate 5's reader-test kit at docs/intelligence/audits/causal-chain-reader-test.md. Same shape: blind reading by an unaligned domain reader, structured answer capture, administrator-side scoring against a per-dossier key, falsification rules that determine whether the gate flips or whether the dossiers need work.

The kit is fully self-contained. The administrator (the user) hands a domain reader the dossier-minus-action-list plus the Reader instructions below; the reader returns their own action list; the administrator scores against the dossier's actual list. The result is recorded in the Result log at the foot of this file.

Why this test, in one paragraph

The product claim Gate 2 carries is that a MacroLens company dossier surfaces non-obvious policy exposure — actions a sector specialist would not have caught in their normal workflow (FactSet alerts, Bloomberg NSE, sell-side notes, Politico Pro). If a senior semis analyst reading our SK Hynix dossier finds nothing in it they didn't already know, the dossier is a marketing page, not intelligence. The "≥3 missed" bar is what makes the dossier defensible as paid research rather than a directory entry. The test cannot be self-administered by the wake: only a reader with their own independent watchlist can produce a credible miss-count.

Reader profile we need

Stricter than Gate 5's profile, because the test is exposure-specific:

• Domain-active. Reader works (or recently worked) inside the sector

being tested — sell-side analyst, in-house exposure / strategy team, sector PM, sector M&A diligence, or sector regulator/policy desk. A generalist macro reader does NOT qualify; their miss-count is uninformative because they would miss everything.

• Maintains their own watchlist. They must be able to answer "what

policy actions touched <company> in the last 24 months?" off the top of their head or from their own notes. If their honest answer is "I'd have to look it up", they are not the reader we want.

• No prior MacroLens exposure. A reader who has seen our register

or any prior dossier draft is disqualified — they have already been primed. (Ask: "Have you ever opened macro.ciconialabs.com or seen any of our IPTM material?")

• 45–60 minutes of attention. That is enough for one dossier including

pre-test watchlist write-down and post-test discussion.

We need 1 reader per dossier × 6 dossiers = 6 readings, each by a different domain person (no single reader covers two dossiers — sector crossover is too narrow to assume independence). If one sector cannot source a qualified reader, drop that dossier from the V1 sample and promote the replacement from the V1-backup list below.

The six dossiers (V1 sample)

Stratified one per primary sector cluster from the 25-dossier corpus. Bias toward dossiers where the miss-count signal is highest — flagship mega-cap dossiers (NVDA / ASML / TSM) are intentionally excluded because every sector analyst follows them and the miss-count would be near-zero for reasons unrelated to dossier quality.

V1-backup list (use if a V1 sample reader cannot be sourced): NDA.DE (Nordex, onshore-wind alternate to Vestas), MP (MP Materials, US REE alternate to Eramet), BA.L (BAE Systems, UK defence alternate to Hanwha), BASF (DE chemicals alternate to Solvay), ALB (Albemarle, US lithium alternate to Eramet). Backups are sector-matched so a swap does not change the cluster coverage.

Reader instructions (paste this verbatim to the reader)

> I am running a structured test on a company-exposure write-up we built. > The test is not of you — it is of whether our write-up surfaces > material policy exposure that a sector specialist would not have > independently catalogued. I need you to do the following in order; > please do not skip step 1. > > 1. Before I send you the write-up, write down — from memory or your > own notes — every government policy action you can recall that has > touched <COMPANY NAME> in the last 24 months. Be specific: > issuer (US BIS, EU Commission, MOFCOM, etc.) + month/year + one > sentence on what the action did. Aim for 5–15 entries. Stop and > send me this list before reading anything else. > 2. After I receive your list, I will send you the full write-up. > Read the "Policy actions touching them (last 24 months)" > section in full. > 3. For each action in the write-up, mark it: > - K — knew it (was on your step-1 list, or was so obvious you'd > have included it if asked another day). > - M — missed it (not on your list, you genuinely did not know it > touched this company at this severity). > - A — argued it (you knew the action exists but disagree that it > touches this company in the way described). > 4. Finally, answer: > - E1 — exposure scoping: Is the dossier's framing of what > material/policy exposures matter for this company correct? (Same > three options: K / M / A — argued = you disagree.) > - E2 — strategic alternatives: Is the dossier's framing of what > this company can do about its exposure defensible?

Send the reader only the dossier file with the action-list section redacted (instructions for redaction in the administrator checklist below). Do not send this kit file. Do not send the dossier title in advance of step 1 — the company name only.

Administrator redaction protocol

The reader needs the dossier's frame (what they do, material exposures, strategic alternatives) but NOT the action list (which is what the miss-count tests). For each V1 dossier:

1. Copy the dossier to a working file (e.g. /tmp/dossier-reader-X.md). 2. Replace the entire ## Policy actions touching them (last 24 months) section content with [REDACTED FOR READING TEST — administrator will reveal after your step-1 list arrives]. 3. Send the redacted file via the channel the reader prefers (PDF / email body / shared doc — anything not the live MacroLens site, which the reader must not have seen). 4. After the reader returns their step-1 list, send the unredacted action-list section as a follow-up message. Do not send the rest of the dossier again — they already have it. 5. Score against the grader's key below.

Scoring rubric (administrator-only — do not share with reader)

For each reading, compute:

• `missed_count` = number of dossier actions marked M by the

reader (genuine misses).

• `argued_count` = number marked A. Argued actions do NOT count

as misses — the reader knew them but disputes the scoping. Track separately because high argued-count is a dossier-quality signal, not a coverage-depth signal.

• `reading_passes` if missed_count ≥ 3 AND **the exposure scoping

question E1 is not argued** (E1=A on a reading means the reader thinks the dossier is mis-framing the company; that override blocks pass even with high miss-count, because the miss-count is measuring the wrong exposure set).

The gate flips to passing when:

• At least 5 of the 6 readings pass (≥3 misses each, scoping not

argued), AND

• At least 1 reading per sector cluster passes (no sector sweeps a

blanket E1=A — meaning at least one reader per axis confirms our framing of that sector).

If 4 of 6 pass, the gate moves to state: near-passing — single-reader shortfall and the administrator sources one replacement reader from the V1-backup list to retest the failed dossier's sector. If ≤3 of 6 pass, the gate stays building and the failure-mode taxonomy below points to the next action.

Argued-actions threshold (separate from gate test): if any single reading produces argued_count ≥ 3, that dossier triggers a refresh spec independent of the gate result — the reader is telling us the dossier scoping is wrong even if it cleared the miss-count.

Grader's key (administrator-only — do not share with reader)

The grader's key here is lighter than Gate 5's because the dossier action list is itself the answer key — any action listed in the dossier and not in the reader's step-1 list is a candidate miss. The grader's judgement work is around what counts as "obvious" (i.e. the reader's "K so obvious you'd have included it" claim).

S — SK Hynix

• Obvious-floor: US BIS chip controls Oct-2022, Oct-2023 expansion, Dec-2024 HBM-included package; KR Semiconductor Special Act Jan-2026. A reader who lists none of these is not a credible semis analyst — disqualify the reading and source a replacement.
• High-miss candidates: 2025-09 VEU revocation, MOFCOM Ann.11/2026 (first PRC instrument naming a non-US target), KR-JP HBM supply-chain bilateral framework.

M — Eramet

• Obvious-floor: Indonesia hilirisasi nickel ore ban (2020); EU CRMA listing of Mn / Ni / Li. A miner-coverage analyst who has neither is uninformed.
• High-miss candidates: Gabon mining-code revision, NC (New Caledonia) provincial royalty, Indonesian PP 28/2025 nickel-smelter moratorium tier, AR (Argentina) RIGI capacity-buildout.

D — Hanwha Aerospace

• Obvious-floor: KR–PL K9 framework, UAE Cheongung KM-SAM-II export auth.
• High-miss candidates: AU Redback IFV LAND 400 Phase 3 selection nuance, Saudi NCMI co-prod offset, EU EDIS K9 add-on.

W — Vestas

• Obvious-floor: US IRA offshore-wind tax credits, EU CBAM steel passthrough.
• High-miss candidates: China REE counter-controls touching permanent-magnet supply, US Section 301 wind-blade and tower component decisions, Danish state-equity injection clauses.

C — Solvay

• Obvious-floor: EU CBAM (chemicals scope expansion), US IRA fluoropolymer / fluorochem sourcing language.
• High-miss candidates: EU CRMA strategic-project listings for Solvay-named facilities, US TSCA PFAS-related listings touching specific fluoropolymer plants, Belgian state-aid Rhodia rare-earth restart authorisation.

R — Umicore

• Obvious-floor: EU CRMA, EU Battery Regulation Annex VI (recyclate content), US IRA FEOC battery-CAM rules.
• High-miss candidates: EU CRMA Strategic Project listings for Olen / Hoboken / Nyrstar interactions, US BIS PGM / catalyst-related entries, KR–EU CAM JV bilateral framework.

The grader's job is to score each reader-marked K-so-obvious claim against the obvious-floor list — if the reader's claim falls inside the obvious-floor, accept K-without-listed; if it falls in the high-miss list, the reader cannot retroactively claim K-so-obvious (they had the chance to list it in step 1 and didn't). High-miss-list items unlisted in step 1 score M regardless of the reader's protest.

Failure-mode taxonomy

If the test fails or partially fails, the failure pattern points to the next action:

• Same dossier fails across multiple readers → that specific dossier

is the problem. Refresh it: add 3+ non-obvious actions, re-test with a fresh reader, keep others' results.

• Same action class missed by every reader (e.g. every reader missed

the consultation-status / pending tier) → the dossier template is weak on that section. Update docs/intelligence/dossiers/README.md to require richer pending-action coverage; re-spec all 25 dossiers for that section.

• Argued-count high across multiple dossiers → the **exposure-scoring

methodology** is over-claiming. File a project to tighten scoping rules. Do not flip the gate even if miss-count passes — over-scoped miss-count is a Pyrrhic pass.

• One sector sweeps 0/1 pass → that sector's dossiers are

systematically under-researched (or the reader for that sector was miscast). Retest with V1-backup; if backup also fails, file the sector-cluster for a research refresh.

• All 6 pass cleanly → flip gate to passing. Public-flip readiness

advances 3/7 → 4/7 — threshold reached, the indexed-vs-private decision comes back to the user.

Distribution & administration checklist

Before sending the redacted dossier to a reader:

• [ ] Confirm the reader has not previously seen MacroLens. (Ask the

explicit question — "Have you ever opened macro.ciconialabs.com, or seen any IPTM-branded material, or read a MacroLens dossier draft?")

• [ ] Confirm the reader fits the domain-active profile for the dossier

they will read (semis specialist for S, miner specialist for M, etc).

• [ ] Prepare the redacted dossier file (tmp/dossier-reader-X.md) per

the redaction protocol above.

• [ ] Send the reader only: (a) the company name + sector frame in

one sentence, (b) the Reader instructions block above. Do NOT send: this kit, the dossier action list, any case-study URL, any /companies/ page link.

• [ ] Wait for the reader's step-1 list. Do not nudge — the unprompted

size of the step-1 list is itself signal.

• [ ] After step-1 list arrives, send the redacted dossier + the

unredacted action-list section.

• [ ] Record the reader's K/M/A markings verbatim. Compute miss-count.

Optionally record qualitative reactions (surprise, push-back, disagreement) — qualitative signal worth keeping even outside scoring.

Caveats

The same caveats Gate 5's kit acknowledges apply, plus two specific to Gate 2:

1. Administrator-and-grader are the same person (the user). The grader's-key calls about K-so-obvious vs M are unavoidably judgement-laden. To reduce administrator bias, the grader's key fixes the obvious-floor and high-miss-candidate lists in advance — the administrator is matching reader claims against a fixed list, not inventing the list on the fly. 2. Domain-reader supply is the real bottleneck. Six qualified domain readers, one per sector, with no prior MacroLens exposure is harder than the four-to-six casual readers Gate 5 needs. If the user's network cannot source these readers within a few weeks, the correct next move is to narrow V1 to 3 sectors (semis + minerals + defence — the strongest commercial frames) rather than dilute the reader profile. A reader-of-convenience produces an uninformative miss-count. 3. Reader-of-convenience drift. If a reader claims domain coverage but their step-1 list is implausibly short (e.g. a "semis analyst" who lists 0–2 actions for SK Hynix), disqualify the reading and source a replacement. The miss-count from an under-active reader is noise. 4. Sector-cluster substitutability. Solvay (C) and Umicore (R) both sit at the specialty-chemicals / critical-materials junction; their readers may overlap. If one reader covers both honestly, accept it but only count it once toward the per-sector-pass requirement.

Result log

Aggregate result (administrator computes after all 6 readings):

• Readings passed: ___ / 6
• Sectors with ≥1 pass: ___ / 6
• Gate-flip criterion (≥5 pass AND ≥1 per sector): _pending_
• Argued-count flags (any dossier ≥3 argued, triggers refresh): _pending_

Once the aggregate is computed, the next strategy tick either flips Gate 2 to passing (criterion met), moves it to near-passing — single-reader shortfall (4/6 with one sector retest needed), or keeps it building with the failure-mode taxonomy pointing to the specific dossier or section refresh required.