The Reporting-Edge Standard: Confirmed, Corroborated, Inferred, or Unplaced

This standard grades every reporting line in a competitor org chart you have rebuilt from public profiles, filings, and press releases. It is for talent-intelligence analysts and strategy teams who have to tell an executive which boxes are fact and which are guesses before anyone makes an org-design or hiring decision from the chart. It gives you four grade definitions, the minimum evidence each grade needs, span-of-control and staleness checks that catch impossible trees, and an agreement test that shows two analysts would grade the same chart the same way.

What the four reporting-edge grades mean

Every reporting edge gets exactly one grade: Confirmed, Corroborated, Inferred, or Unplaced. A reporting edge is one line from a manager's box to a direct report's box. The grade records the strongest evidence attached to that edge, not how sure the analyst feels.

Most public advice on building a competitor org chart from LinkedIn stops at "titles are noisy, so attach a confidence level." A confidence level with no written criteria is just an opinion with a label on it. The grades below are criteria, written so that a skeptical executive asking "how do you know she reports to him?" gets the same answer from any analyst on the team.

Definitions written so two analysts grade alike

• Confirmed. A company-published source (a press release, a bio on the company's site, or the text of a filing) contains explicit reporting language such as "reports to" or "reporting to". It names the child and names either the parent or a role only one node holds, such as the CEO. The quote, URL, and capture date are logged, the capture is inside the shelf-life cap, and nothing newer contradicts it. What it proves: the company said so on that date. How it lies: the release announced an appointment and the structure has changed since.
• Corroborated. No company source states the edge, but a person-level source does: the child's or the parent's own profile line, or a conference or speaker bio. A second, independent signal agrees with it: the other person's profile, a matching function and business unit on the title ladder, or a job post that names the role the new hire reports to. What it proves: a person claimed the line and the surrounding evidence fits. How it lies: the claim is out of date or was edited after the fact.
• Inferred. No explicit reporting language exists anywhere. A title-ladder rule returns exactly one candidate parent, and the rule ID is logged on the edge. A title-ladder rule is a written rule that predicts a parent from function, business unit, and title level. What it proves: the titles are consistent with the line. How it lies: titles are inflated, or a "manager" title belongs to someone with no team.
• Unplaced. The person is evidenced at the company, but no source places them. The title-ladder rule returns zero candidates or several, the sources conflict, or all the evidence is past the cap. The box sits under the lowest ancestor you are sure of, with no line drawn. Unplaced is a finding, not a failure, because it shows the strategy team exactly where the map is blank.

What each public source proves about a reporting line

Each type of source proves something different. Filings prove a person holds a top-level role. Company statements prove an edge. Profile lines prove that someone made a claim. Title ladders prove only a hypothesis. Grade on what the source proves, not on how authoritative it feels.

Securities filings

Item 401 of Regulation S-K requires companies to set forth the names and ages of all directors and executive officers. That disclosure appears in Form S-1, proxy statements, and the 10-K, where the section heading now reads "Information about our Executive Officers". Item 401(c) adds separate disclosure of significant employees. So filings are the best source for which people sit at N-1, the people who report directly to the CEO. They rarely say who reports to whom. A filing that lists both a CFO and a Chief Accounting Officer does not tell you the second reports to the first.

I have not verified what officer-appointment filings usually say about reporting lines. If you use one, grade it on quoted language only, the same as any other company source.

Press releases and company bios

Look for the words themselves. "Will report to the Chief Executive Officer" in a company press release makes the edge Confirmed. "Joins the leadership team" does not, because it places a node, not an edge.

Self-stated profile lines and conference bios

A profile line such as "reporting to the CTO" is a claim made by one person. One guide that ranks for this topic says employees often list their manager and direct reports right on their profile, but it gives no evidence. Your own tally will show how often that holds at the company you are mapping. Profiles are also not fixed records. NBER Working Paper 35546, "Time Travel on Professional Profiles", found that 19.7 percent of established US LinkedIn users retroactively edit the title or description of a job they have already left.

Job posts and title ladders

A job post that names the role a hire reports to is a corroborating signal, not a placement. Inferring reporting lines from job titles produces hypotheses. Use it last, and only on edges that still have no evidence.

The depth floor

Nobody has published the depth (N-2, N-3, N-4) at which each type of source runs out. The case study that ranks for competitor org charts got its reporting lines from a series of in-depth interviews with current and former employees. It used profiles, job posts, and press announcements only for an initial framework. Treat the floor as something you measure. Tabulate grades by depth for each company: the first depth with no Confirmed edge is your desk-research floor for that company. Below it, Confirmed requires primary research, which this standard does not cover. Report the floor to the executive along with the chart.

How to grade one edge: decision order and tie-breaks

Take each edge down the evidence ladder and stop at the first rung it clears. That rung is the grade. Conflicts and expired evidence can only push an edge down, never up.

Tie-break rules

• Two dated sources conflict: the newer one wins if it is the same tier or higher. Record the loser in the notes.
• Two sources conflict and either one is undated: the edge is Unplaced until the conflict is resolved.
• Evidence is past the shelf-life cap: drop one grade. An expired Inferred edge becomes Unplaced.
• The title-ladder rule returns two candidate parents: the edge is Unplaced. Never pick the likelier one.
• A filing is the only source: the node is evidenced, and the edge cannot be graded above Inferred.

Edge types

No published standard covers how to record dotted-line, matrix, or dual reporting, so this part is house policy. Record a type on every edge, separately from its grade:

• Solid: the primary line, the one that owns the person's performance review and pay.
• Dotted: a secondary line that the evidence actually states.
• Regional: a country or region head reporting to a global head.
• Functional: a specialist embedded in a business unit reporting to the head of their function.

A dotted, regional, or functional edge can never be Inferred. Title ladders only predict solid lines, so a non-solid edge needs at least Corroborated evidence or you don't draw it.

edge_id      = E-[company]-[number]
parent       = [name, title]
child        = [name, title]
edge_type    = solid | dotted | regional | functional
depth        = N-[1, 2, 3...]
grade        = Confirmed | Corroborated | Inferred | Unplaced
source_1     = [type] | [URL] | captured [YYYY-MM-DD] | snapshot [archive path]
quote_1      = "[exact reporting wording]"
source_2     = [type] | [URL] | captured [YYYY-MM-DD] | snapshot [archive path]
quote_2      = "[exact wording, or the independent signal described]"
title_rule   = [rule ID, Inferred edges only]
expires      = [YYYY-MM-DD]
graded_by    = [analyst] / second grader [analyst or none]
notes        = [conflicts, downgrades, span flags]

The grading procedure, step by step

Nine steps take an ungraded chart to a published chart whose edges expire. No source publishes how long each step takes, so I have left durations out rather than guess. Published sources also disagree on the order. One ranking case study runs primary interviews alongside desk research, and another guide starts from LinkedIn and verifies afterwards. I put filings first because they are cheap and they pin down the N-1 layer before anyone draws a line.

Step 5 is where most of the hours go. Profiles that state a reporting line in so many words are scattered and few, and keyword search on titles brings back mostly noise. Refolk lets you ask for exactly those people in plain English, across LinkedIn and the open web, so you start from a list to snapshot instead of a trawl.

A search result is a lead, not evidence. Each line it surfaces still needs step 5's URL, snapshot date, and exact wording before it counts toward a grade.

Span-of-control benchmark band for a reconstructed chart

Flag any subtree whose reconstructed number of direct reports falls outside a band built from level-specific benchmarks. Span of control is the number of direct reports one manager has. No published source defines an "implausible" band for a reconstructed chart, so the band is a policy choice and must be labelled as one on the deliverable.

The Gallup shares sum to 101% because of rounding. These benchmarks are refreshed from time to time (Pave says its post has been updated with data over time), so re-check the source pages each time you re-baseline the band.

Level-specific benchmarks are much tighter than national ones. Pave's directors average 4.6 direct reports. Gallup's US mean of 12.1 mixes in frontline roles, where 13% of managers oversee 25 or more people. A software-org subtree judged against the Gallup mean would almost never be flagged.

House default band (adjust it, then write it down):

• Wide flag: any Manager through Senior Director node with more than 10 reconstructed direct reports. Two-thirds of managers manage fewer than 10, and 10 is more than double Pave's averages for those levels.
• Empty flag: any node with a manager title and zero reconstructed direct reports. Either it is an individual contributor, or there is a team you have not found yet.
• Shift flag: any parent whose span changes after you deduplicate people and normalise employer names.

A flag is not a verdict. It sends the subtree back through the failure-mode checks in the next section.

How reporting-edge grading goes wrong

Most false Confirmed edges and invented manager nodes come from eight repeatable mistakes, and each one has a check that catches it. Run these checks on every flagged subtree and on a sample of the unflagged ones.

The biggest trap is assuming title counts mirror the org. In Refolk's index of professional profiles for US computer software, seniority bands do the opposite:

Ratios and shares other than Microsoft's are derived from the index counts.

The inversion comes from title inflation combined with "manager" titles held by individual contributors. Inferring layers from title counts is unsafe for this reason, and so is any title-ladder rule that treats "manager" as proof of a team.

The eight failure modes

1. Grading a filing as proof of an edge. Item 401 lists officers but never says who they report to. False positive: a Confirmed edge between two officers whose relationship the filing never describes. Check: is there quoted reporting language from the filing itself?
2. Treating an individual-contributor title as a manager. False positive: a manager node invented on top of a team that does not exist. Check: does the person have any self-stated direct reports, or any child edge graded above Inferred?
3. One person in two bands. In Refolk's index, the same title, "Director of Platform Content" at the same employer, appears in both the Experienced Manager and the Director results. False positive: one person counted twice, which inflates the span. Check: deduplicate by name plus employer before counting any span.
4. Duplicate employer entities. Among Director profiles in Refolk's index, one company appears under two capitalisations of its name (5 profiles plus 1). False positive: one company's subtree split in two, with each half looking thin. Check: normalise employer names before building any subtree.
5. Using a mean as the plausibility band. False positive: inflated subtrees pass because the national mean is double the median. Check: use level-specific figures and medians where they exist.
6. Assuming profiles are fixed records. 19.7 percent of established users rewrite past jobs. False positive: a historical title, edited later, used as evidence of a past edge. Check: archive a dated snapshot of every profile line you rely on.
7. The kappa paradox. In one practitioner example, 96.5% raw agreement came with a kappa of only 0.57 because one grade dominated the sample. False positive: rewriting rules that were actually working, or trusting a high raw agreement that hides disagreement on the rare grades. Check: report percent agreement alongside kappa, and consider Gwet's AC1.
8. A provenance gap. Trade-secret law rejects the "could have" defense. False positive: an edge that is true but cannot be shown to come from proper means, which makes it a legal exposure. Check: no edge without a URL and a capture date.

Agreement test: do two analysts grade the same chart alike?

The standard holds only if two analysts who grade a sample independently reach a Cohen's kappa of at least 0.61 and raw percent agreement of at least 80%. Kappa is an agreement statistic that discounts the agreement two graders would reach by chance alone.

Cohen's original bands treat 0.41 to 0.60 as moderate, 0.61 to 0.80 as substantial, and 0.81 to 1.00 as almost perfect agreement. McHugh's stricter reading in Biochemia Medica holds that any kappa below 0.60 shows inadequate agreement between raters, and notes that many texts recommend 80% agreement as the minimum acceptable. That is where the house thresholds come from: substantial kappa, plus the 80% floor.

How to run it:

• Grade blind. Each analyst grades from the same evidence records without seeing the other's grades.
• Compute both numbers. Report kappa and percent agreement side by side, always.
• Diagnose before rewriting. If agreement is high but kappa is low, check whether one grade dominates the sample, and compute Gwet's AC1 before you touch the rules.
• Fix the rule, not the analyst. Every disagreement points to a definition that allowed two readings. Rewrite that sentence in the policy, then re-grade.

I found no published rule for how large the sample should be for this use. My practice is to double-grade the whole of the first chart a team builds under this standard, then sample on later charts, making sure every grade appears in the sample more than a handful of times.

REPORTING-EDGE GRADING POLICY   version [v1], adopted [date], owner [name]

1. Grades, one per edge, strongest evidence wins
   Confirmed    = company source quotes reporting language naming both ends
   Corroborated = person-level reporting claim + one independent signal
   Inferred     = logged title-ladder rule returns exactly one parent
   Unplaced     = evidenced at the company, no unique placement
2. Filings evidence nodes only. They never raise an edge above Inferred.
3. Dotted, regional, and functional edges need Corroborated or better.
4. Conflict with an undated source = Unplaced. Expired evidence = drop one grade.
5. Span band (policy) = flag Manager to Sr Director nodes with 0 or more than [10] direct reports.
6. Shelf-life cap (policy) = [12] months from capture date, every source.
7. Agreement gate = kappa at least [0.61] AND raw agreement at least [80%], both reported.
8. Depth floor = first depth with no Confirmed edge, stated per company.
9. Every edge carries a URL and capture date. No source, no line.

Staleness caps and lawful provenance

Every Corroborated and Inferred edge carries an expiry date, and every edge carries a logged source. Profiles drift, and the provenance log is also your legal record. Nobody has published the median lag between a job change and the matching profile update, so any cap is a policy you derive, not a published finding.

Deriving the shelf-life cap

• Edges break faster than jobs. One data-freshness study puts median job tenure at 3.9 years. An edge breaks when either person moves, or when the parent's own manager reorganises the team, so its life is shorter than either person's tenure.
• Edits come before moves. NBER Working Paper 35546 found that workers who later change employers are more than twice as likely to have made a retroactive edit a full year before the move. Edits reach roughly double the baseline right around the move itself.
• The default follows from that. A 12-month cap from capture date matches that one-year lead time. Treat a freshly edited profile as a sign the edge may be about to go stale, and re-check it, not as proof that it is current.

A claim that "67% of LinkedIn profiles are out of date" circulates widely on a ranking page without any source. Don't cite it.

Lawful sourcing

Under the US Defend Trade Secrets Act, 18 U.S.C. 1839(6)(B) says improper means "does not include reverse engineering, independent derivation, or any other lawful means of acquisition." Independent derivation can rest on government filings, judicial and administrative proceedings, social media, and professional networking sites. But evidence that a defendant "could have" obtained the information that way is not a defense if proper means were not actually used. The evidence record is how you show they were.

Privacy law applies too. The privacy regulators' concluding joint statement on data scraping says publicly accessible personal information is still subject to data protection and privacy laws in most jurisdictions. It also says mass scraping incidents that harvest personal information can be reportable data breaches in many jurisdictions. The earlier statement in that series went to the parent companies of platforms including LinkedIn.

Keeping the graded chart current

A graded chart is a dated snapshot, not a living record. Schedule re-sweeps by expiry date and by known disclosure events, and keep the depth floor and span band visible on every version.

• On expiry: re-run steps 4 to 7 for every edge whose date has passed. Downgrade anything you cannot refresh.
• On new filings: when a new 10-K or proxy comes out, re-run step 3 and re-match every N-1 node.
• On company announcements: any leadership press release triggers a re-check of the edges it touches.
• On profile edits: if a person at either end of a Corroborated edge edits a past job, treat it as an early sign they may move.
• On each new chart: re-measure the depth floor. It differs by company, and it is the honest answer to "how far down can we trust this?"

For later re-checks, Refolk can pull the current people for a given role and employer in a single plain-English request, which makes the expiry sweep a short diff instead of a rebuild.