16 — Accuracy, Trust and Research Risks

16 — Accuracy, Trust and Research Risks

Phase 1 deliverable · research date 2026-09-15

16.0 Why this document is evidence-led rather than aspirational

Most editorial policies are written in advance of the problems they describe. This one is written after them. Every risk in §16.1 is grounded in something the Phase 1 research actually encountered across 25 sectors, 500 trend records, 2,580 evidence items and 603 registered sources — including several cases where the research programme was itself the target of the failure and recorded it rather than absorbing it.

The rules in §16.2 are therefore written as enforceable controls with a named enforcement point (ingestion, record, render, review, or publish), not as principles. A principle with no enforcement point is a marketing claim about ourselves, and this platform's whole proposition is that it labels those.

16.1 The risk register

16.1a Summary

16.1b The six that need more than a table row

1 · Vendor-captured evidence bases. This is the most serious risk in the register because it is invisible at the article level and only visible at the corpus level. Sector 04 is the worst case and documented itself as such: a deliberate September 2026 search for independent adoption or efficacy data on the autonomous AI SOC returned exclusively vendor blogs, vendor-sponsored buyer's guides and consultancy content marketing — no neutral study, no regulator dataset, no peer-reviewed evaluation. T-04-19 is recorded unverified with evidence_quality 1 and a composite of 50.4. The dossier states the absence as a finding rather than filling it. Two structural aggravators sit underneath: vendor-published alert-reduction percentages measure workload inside the tool rather than detection efficacy and are trivially gamed by suppressing alerts; and only 24% of H1 2026 US breach notices disclosed an attack vector, the lowest share ITRC has recorded, which hands the narrative to the only remaining source of vector data — vendors. Sector 04 additionally refuses three figures that saturate the trade press: any "cost of a data breach" average, any "unfilled cybersecurity jobs" count, and any "$X billion market by 2032" projection from an unnamed methodology. The control is a corpus-level test, not an article-level one: if a category's entire evidence base shares one commercial interest, that fact is the lede.

2 · The statistic sold by the company selling the remedy. Sector 24's defining number — AI-generated tracks exceeding 50% of daily new uploads, about 90,000 tracks a day — comes from Deezer, which is the only platform publishing AI-upload and AI-fraud data and which licenses the detection product those numbers justify to collecting societies. There is no second series to triangulate against. T-24-04 and T-24-19 are marked single_source with evidence_quality held to 2, setting a score ceiling of 70.0 on both — a ceiling neither raw score reaches, which is the correct outcome: the cap is a guard rail, not a penalty. The same disclosure contains the demand figure — fully AI-generated music is 1–3% of total streams, of which Deezer identified up to 85% as fraudulent in 2025 against platform-wide fraud of 8% — and only the supply number circulates. The lesson generalises past this record: when a single vendor is both the sole measurer and the seller of the remedy, the measurement is an estimate with a disclosed interest, never a fact, however primary the publisher.

3 · Date laundering. Recorded in the macro addendum as a confirmed, cross-sector statistical trap: searching for a current metric often surfaces a two-year-old article in the top results, because search ranking rewards authority and link age rather than recency. The research programme's countermeasure was to check the publication date of every result and prefer the primary issuer's own current page. This is why every source record carries a date and undated pages are Tier C at best and flagged undated: true. A related trap: rounded cumulative disclosures cannot be differenced — YouTube's "$100bn paid to creators over 4 years" was reported at the same $100bn a year earlier, so no annual growth series exists, and any article deriving one has invented it.

4 · Non-English and non-Western absence. The agents recorded this honestly and at cost to their own dossiers, which is exactly the behaviour the policy must protect:

Three findings show the gap is a retrieval and budget problem with a known fix, not an unknowable. Sector 13 met the regional-source requirement through the Shanghai Stock Exchange carrying Unitree's audited prospectus figures. Sector 03 used Korean-origin primaries directly and found Seoul Economic Daily reporting HBM4 pricing weeks ahead of equivalent English-language trade coverage. And sector 03 identified Taiwan's MOPS monthly revenue disclosure — mandatory for every TWSE-listed company, and the highest-frequency hard data in the global semiconductor industry — as "systematically underused outside Taiwan because the MOPS interface is Chinese-language." The best source in the sector is free, public, and behind a language barrier rather than a paywall.

5 · Overrepresentation of wealthy markets, underrepresentation of small companies. These are the same bias viewed from two angles: coverage follows disclosure, and disclosure follows capital. Americas took 66% of Q2 2026 VC funding; OpenAI and Anthropic alone took $217B, 43% of all H1 2026 funding; the top 10 deals were 46% of Q2. A source base built on filings and IR pages will mirror that concentration precisely. Sector 11 supplies the correction and it must be applied everywhere: report medians and counts alongside sums. Aggregate dollars look euphoric while 2021-vintage DPI sits at 0.05x, net LP cash flow since 2022 is −$202B, and seed→Series A graduation has fallen from 55%+ to 16%. A second structural aggravator is shrinking public disclosure: EA completed its take-private on 2026-08-04, removing a top-five publisher's quarterly reporting. Disclosure shrinkage is itself a trend the platform must track, because it degrades the evidence base for everything else.

6 · Algorithmic bias in our own scoring. The platform's ranking and scoring are opinions expressed as arithmetic, and pretending otherwise would be the exact failure we document in others. Two defences are already in place and both must be maintained: the weights are published with their rationale, and a sensitivity check against equal weights showed 24 of 25 sectors moving two places or fewer with a maximum movement of three — so the ranking is not an artefact of the weighting. Three residual biases are recorded rather than solved: data_availability is deliberately weighted lowest (6%) because poor data availability is also the opportunity, which systematically favours sectors we can research over sectors that matter; speed_of_change and intelligence_demand at 13% each favour fast, well-funded, English- language sectors; and the composite is explicitly a coverage-priority score, not an importance score — publishing under a different name would itself be a form of bias.

16.2 The editorial policy — enforceable rules

Each rule states its enforcement point and what happens when it fails. "Blocks publication" means the CMS refuses the render; it is not a reminder.

R1 · Source transparency

Enforcement: ingestion + render. Every claim carries source_idx resolving to a sources[] entry with title, publisher, url, date and tier. Every source in the registry carries reliability, bias_notes, cost and legal_use_restrictions. Tier C may never be the sole support for a claim, and Tier C entries render with a warning label. Where a source has a commercial interest in the claim it supplies, bias_notes must state the interest and the interest renders alongside the claim in prose (see 18-editorial-formats.md §12.5.2). Seed baseline across trend records: 1,037 Tier A, 793 Tier B, 66 Tier C. Failure: unresolvable source_idx, missing date, or Tier-C-only support blocks publication.

R2 · Independence and corroboration

Enforcement: record. Two outlets syndicating one wire story are one source. Derivative coverage of a primary report does not corroborate it. Any load-bearing numeric claim without two genuinely independent sources is marked verification_status: "single_source" and its evidence_quality is capped at 2, which propagates into the composite through the evidence cap. Seed baseline: 347 triangulated, 146 single-source, 7 unverified. Failure: mislabelled verification status is a correctable error under R14 and a repeat pattern is an analyst performance issue.

R3 · Date stamps

Enforcement: ingestion + render. Every evidence item, every source and every record carries a date. last_verified renders as an absolute date plus an age in days on every article. Undated web pages are Tier C at best and flagged undated: true. Records older than 90 days render a staleness banner; older than 180 days are removed from ranked feeds until re-verified. Regulatory claims carry the four-stage marker — proposed / enacted / in force / enforced — because vendor marketing routinely collapses them, and the EU AI Act high-risk employment deferral from 2026-08-02 to 2027-12-02 is the standing worked example. Failure: a missing date blocks publication.

R4 · Confidence levels

Enforcement: render. confidence, verification_status, evidence_quality, source_diversity, composite_score, raw_score, evidence_cap and capped render as a block on every article, never as a single adjective. Where capped is true the article states the raw score, the cap and the reason. Twenty-five of 500 seed records are capped. Confidence is never paywalled at any tier, including anonymous readers. Failure: a render missing the confidence block blocks publication.

R5 · Fact versus forecast separation

Enforcement: record + render. The six-type ladder is enumerated and closed: fact, estimate, opinion, forecast, signal, marketing. Ingestion rejects any other value — including survey, which appears once in the seed set at T-11-11 and must be re-typed as estimate or opinion. A forecast requires stated assumptions, a stated horizon and an identified falsifier in the form specified by 15-forecasting-framework.md; a future-tense claim without a falsifier is a prediction and this platform does not produce predictions. Others' predictions are recorded as opinion, attributed and dated, because the track record of predictors is itself data. A marketing claim may never render without its label — no exception, in any format, including headlines, summaries, social posts and AI-generated abstracts. Failure: an unlabelled marketing claim or a falsifier-free forecast blocks publication.

R6 · Sponsorship, advertising and affiliate disclosure

Enforcement: publish.

1. No sponsored trend records, ever. Sponsorship may attach to formats (newsletters, events, reports as a whole) and never to a record, a score, a classification or a contradiction.
2. Sponsored placements are visually distinct, labelled "Paid", excluded from ranked feeds, and excluded from any AI-generated summary the platform produces.
3. No affiliate links in editorial. If commerce links are ever introduced they are confined to a separate, labelled surface and disclosed at the point of the link.
4. No vendor pre-review. Subjects may be offered a factual accuracy check on specified claims with a stated deadline; they never see the classification, the score or the headline.
5. Analysts declare holdings and outside income; see §16.4.
6. Anyone whose product is covered may not sponsor coverage of their own sector for the quarter of publication or the quarter preceding it. Failure: an undisclosed commercial relationship is a retraction-grade event under R14.

R7 · Human review of sensitive content

Enforcement: review gate. Specified in full at §16.3.

R8 · The attention firewall

Enforcement: scoring model. Search interest, funding volume, media coverage and social volume are inputs to velocity only, and velocity is 7.5% of the composite. Attention alone caps velocity at 3. A trend under twelve months old cannot score above 2 on persistence. The evidence cap is arithmetic: cap = 40 + 60 × mean(evidence_quality, source_diversity), so weak evidence caps a trend at 40 regardless of how loud it is. No editorial override exists, and none will be built. Failure: n/a — this rule cannot be violated by a human, which is the point.

R9 · Data licensing

Enforcement: ingestion. No scraping behind logins, no paywall circumvention, no ToS violations, no robots.txt violations. Official APIs, RSS, public pages and openly published data only. Every source record carries legal_use_restrictions, and rate limits and User-Agent requirements are honoured (SEC's declared-contact and 10 req/s limit is the reference case). Blocked sources are recorded as licensing needs rather than worked around — congress.gov, FDIC QBP, NPCI, Banco Central do Brasil, Box Office Mojo, CISA KEV JSON and cve.org are already on that list. Where a paid source is required it is bought; 106 of 603 registered sources are non-free and Phase 2 licensing priority is WARC first, IAB second. Failure: ingestion from a restricted source blocks publication and triggers removal of derived records.

R10 · Privacy

Enforcement: record. Person entities (113 of 994) hold only professional information: role, employer, public statements, public filings. No personal contact details, no home address, no family information, no non-public biographical data, no inferred characteristics. Public statements are quoted with date and venue. Individuals named in litigation are recorded with procedural status stated precisely and never beyond what the filing supports — sector 17 explicitly recorded case names while declining to assert procedural statuses it could not disambiguate, and that is the standard. Removal requests route to the review gate.

R11 · Methodological transparency

Enforcement: publish. The scoring formula, the 15 dimensions and their anchors, the sector weights and the equal-weight sensitivity check are published and versioned. The composite is described everywhere as a coverage-priority score, not an importance score. Rankings are re-scored annually, and a sector's speed_of_change is published as the indicator of how fast its own ranking will go stale.

R12 · Geographic and linguistic coverage disclosure

Enforcement: record + render. Every record carries geography[]. Any record claiming global scope whose sources are entirely English-language and Western renders a coverage limitation. Each sector dossier states its non-English gaps explicitly (§16.1b table), and those statements render on the sector page rather than sitting in an internal file. Phase 2 targets: at least one non-English primary source per sector where the sector's centre of gravity is non-Western, and the sector-13, sector-16 and sector-05 gaps closed first because they are the largest.

R13 · Aggregate-versus-median discipline

Enforcement: render. Any sum reported across companies renders with a count and, where available, a median. Concentration is stated whenever it exceeds a third of the total — "$217B to two companies, 43% of all H1 2026 funding" is the reference construction. Cumulative rounded disclosures may not be differenced to produce a growth series.

R14 · Corrections

Enforcement: publish. Four severities, all logged, all dated, all public:

Corrections are never made silently to a claim that a reader may have acted on. A correction is published within one business day of confirmation, and the confirmation clock starts when a reader reports it, not when we agree. The corrections log is public, permanent and not paywalled. Reader-reported errors are acknowledged with the reporter's name where they consent — a large non-expert readership is the cheapest audit available (17-audiences.md §10.4).

R15 · Historical versioning

Enforcement: storage. Every record is versioned on every write. Each version stores the full record, the diff, the actor (analyst ID or pipeline job), the timestamp, and the trigger. Published articles are permalinked to a version: /t/T-13-19 serves current, /t/T-13-19@2026-09-15 serves the version as it stood. Score history is retained so an evidence-cap demotion is a visible event rather than a silent change. Nothing is deleted. Withdrawn records are tombstoned with the reason and remain resolvable, because a citation that breaks is a citation that cannot be checked. Retention: indefinite for records and versions; 24 months minimum for ingestion logs.

16.3 The human-review gate

Automation renders articles; it does not clear them. Six categories always route to a named human reviewer before publication, regardless of confidence, source tier, or how mechanical the render was.

16.3a Categories, triggers and reviewers

16.3b Flags the system must preserve

Phase 1 agents set editorial flags by hand, in prose, in two sectors. Those flags are data and the pipeline must carry them, not re-derive them.

• Sector 06 tagged 36 health claims. The convention: "Every clinical, efficacy or safety claim in this dossier and in the accompanying trend records is prefixed [HEALTH-CLAIM] so downstream editorial review can find it with a single grep." The count today is 30 occurrences inside the trend records, spread across 12 of the sector's 20 records — T-06-01, -05, -07, -09, -10, -11, -12, -13, -14, -16, -18, -19 — plus 6 in the dossier text: 36 in total. Migration requirement: each [HEALTH-CLAIM] prefix becomes a structured review_flags: ["health"] entry on the evidence item, and the prefix is retained in the claim text until the structured flag is verified present on all 36. A migration that strips the prefixes before the flags exist would silently drop a sector's entire review surface.
• Sector 17 flagged three records for safety review. T-17-09, T-17-17 and T-17-20 concern AI companion apps and minors' safety, including statutory suicide-prevention and crisis-referral requirements. The dossier states: "No claim about the existence, causation or severity of specific harms is asserted by this dossier. These records are flagged for editorial review before any external publication." All three carry low-to-medium confidence and two are single_source. Migration requirement: review_flags: ["safety","minors"], a publication block until a named human clears them, and preservation of the dossier's non-assertion language in the rendered article.
• Generalisation. Any flag an analyst sets in prose is a defect in the schema, not a defect in the analyst. Phase 2 adds a first-class review_flags[] array to the trend and evidence schemas with the six categories above as its closed vocabulary, and a backfill job that reads the existing prose flags before any text normalisation runs.

16.3c How the gate operates

1. Blocking, not advisory. A flagged record cannot render to a public URL without a recorded clearance: reviewer identity, timestamp, and the version cleared. Clearance is version-specific — a material edit re-opens the gate.
2. Named reviewer, logged. No anonymous clearances. The log is internal but auditable and is disclosed to a regulator or court on request.
3. Service levels. Health, legal and safety: 2 business days. Financial, political, reputation-sensitive: 1 business day. A breach of the SLA delays publication; it never auto-clears.
4. Escalation. Reviewer disagreement escalates to the editor-in-chief; a reviewer-versus-analyst disagreement on a factual claim is resolved by returning to the primary source, not by seniority.
5. Quarterly audit. A sample of cleared records is re-reviewed cold. The audit measures flag recall (flagged records that should have been) and flag precision (records flagged that needed no specialist judgement); a precision below ~0.5 means the triggers are too broad and are costing review capacity that the genuinely sensitive records need.

16.4 Conflicts-of-interest register

Two registers, one design. Internal covers our people and the company. External covers the interests of the sources we cite. Both are queryable at render time, and both drive disclosures that appear in the article rather than in a policy page nobody reads.

16.4a Table sketch

conflicts_of_interest
---------------------
coi_id              TEXT PRIMARY KEY      -- COI-0001
scope               TEXT NOT NULL         -- internal | external
holder_type         TEXT NOT NULL         -- staff | contributor | company | source_org | source_author
holder_id           TEXT NOT NULL         -- staff ID, or entity_id / source_id
subject_type        TEXT NOT NULL         -- entity | sector | trend | product
subject_id          TEXT NOT NULL         -- E-01-01 | 04 | T-13-19
interest_type       TEXT NOT NULL         -- equity | options | debt | employment | board_seat |
                                          -- advisory | consulting_income | sponsorship |
                                          -- vendor_of_remedy | licensing_revenue | family |
                                          -- litigation_party | grant_funding
direction           TEXT                  -- long | short | neutral | unknown
materiality         TEXT NOT NULL         -- de_minimis | disclosable | disqualifying
description         TEXT NOT NULL
declared_on         DATE NOT NULL
effective_from      DATE NOT NULL
effective_to        DATE                  -- NULL = current
disclosure_string   TEXT NOT NULL         -- exact text rendered to the reader
renders_on          TEXT NOT NULL         -- article | sector_page | source_row | internal_only
verified_by         TEXT NOT NULL         -- reviewer ID
verified_on         DATE NOT NULL
review_due          DATE NOT NULL         -- max 12 months
evidence_url        TEXT                  -- filing, registry, contract reference
status              TEXT NOT NULL         -- active | expired | waived | breached

16.4b Operating rules

• Declare on joining, on change, and annually. Analysts declare holdings, options, prior employment (24 months), board and advisory roles, consulting income, and immediate-family equivalents.
• Disqualifying interests bar coverage. An analyst holding a disclosable position in a company does not write about that company; a disqualifying position bars the sector. Sector assignment reads the register before it assigns.
• The company declares too. Investors in this platform, and any sector in which they have a material position, are disclosed on that sector's page.
• External interests attach to sources. source_org rows are the structural control behind R1's interest-disclosure requirement and are populated from the registry's existing bias_notes. The two standing cases: Deezer as sole measurer and vendor of the AI-detection remedy (vendor_of_remedy + licensing_revenue), and the cybersecurity vendor corpus (vendor_of_remedy at category scale). A vendor_of_remedy row forces the interest into the prose sentence carrying the claim.
• Sponsorship is a COI row, not a sales record. Every sponsorship writes a row with renders_on: article for the duration plus the preceding quarter (R6.6).
• The register is auditable and partially public. Disclosure strings and the fact of a declaration are public; monetary amounts are not.

16.5 Market-manipulation policy

The premise. This platform's coverage can move prices — most acutely in private markets, where a mark is set by a small number of observers and a reclassification from emerging to cooling is itself information. Sector 11 established that four companies account for 93.5% of 2026 exit value and that marks are rising while cash is not; sector 13 recorded humanoid valuations deflating from an August 2026 peak (T-13-16) while the category's disclosed revenue is single-digit millions. A platform whose product is calibrated scepticism about valuations has a short position in narrative whether it wants one or not. The policy exists to ensure that position is never a financial one.

Rules

1. No trading on coverage. Staff and contributors may not trade securities, tokens or private interests in any entity they cover, for the duration of coverage and 30 days after publication. No short positions in covered entities at any time. Index funds and blind trusts are exempt; single-name positions are not.
2. No pre-publication disclosure, to anyone. Not to subscribers, not to investors in this platform, not to sources. Every tier receives a given article at the same moment. Tier differences are about workflow — API, export, alerting, history — never about timing of a market-relevant classification.
3. No differential embargo. Where an embargo is agreed with a source, it binds all readers equally.
4. Private-valuation claims carry mandatory furniture. Date, observer, and the basis (primary round, secondary trade, tender, model). The reference construction is the entity record for OpenAI: "Last verified valuation $730B (Axios, 2026-05-28). Tier C claims of a later $852B mark after a $122B raise are unverified." A secondary-market price is not a valuation and is never rendered as one.
5. Short-seller and activist sourcing is labelled at the point of use. Such research may be cited — it is often the only adversarial analysis available — but always as opinion with the position disclosed in the same sentence, and never as the sole support for a reputation-sensitive claim.
6. Reclassification is treated as market-sensitive. A move into or out of overhyped or cooling, or a composite-score change of 5 points or more, publishes on the public feed simultaneously with any alert, and is never previewed to the subject beyond the factual accuracy check permitted by R6.4.
7. Coordinated-promotion detection is a coverage duty. Where the evidence for a trend is dominated by promotional activity from parties with positions, that finding goes in the article. The Gotbit judgment establishes wash trading to "create the false impression of market interest" as prosecuted conduct in sector 21; the equivalent in sector 13 is "deployment tracker" pages counting letters of intent as deployments; the equivalent in sector 16 is automated traffic contaminating audience models. The promotion is the story whenever the promotion is the evidence.
8. We do not take positions in what we cover, as a company. No corporate venture arm, no revenue share on covered products, no equity-for-coverage in any form.

16.6 Measuring whether the policy works

Published quarterly on the methodology page. A policy with no metric is a promise.

More research from the library · Explore trend records