What we measure, how we measure it, and what we refuse to do.
This page states the rules the analysis follows and the vocabulary it uses. Where it cannot yet publish something real, it says so in that section.
Six rules for how a judgment is made and shown.
Seven words. Here is what each one asserts, and what has to be true before we use it.
Every commitment on a report resolves to exactly one of these seven words. The word is the finding — never a rank, a letter, or a number.
What each of the seven words asserts
- Major misalignment: The clearest conflict with the commitment’s own text.
- Significant misalignment: A serious conflict with the commitment — a distinct word from “major,” one degree below it.
- Some misalignment: A conflict with the commitment, at the narrowest of the three graded degrees.
- Within commitment: No distance between what the commitment states and what the holding does.
- Advances: The holding actively supports the commitment’s own text — not just avoids conflicting with it.
- Cannot be graded: The evidence available does not clear the floor a judgment in either direction would need. Stated as a gap, never defaulted to neutral.
- Counter to preference: Runs against a stated preference. A preference states no bound to breach, so this word has no degrees: more adverse evidence never escalates it.
The evidence floor a judgment about a company’s conduct has to clear
No word against a holding is said on thin evidence. Where the judgment rests on what a company did, a ratified evidence gate decides how far an adverse word may go: a Regulated source, or Researched evidence — two or more independent underlying sources, or a licensed dataset. Researched permits at most significant misalignment; only a Regulated record permits major. One credible secondary source is surfaced as a single-source observation, named and quoted on the commitment row, and never converted into a word.
A name or a sector pattern alone never produces a negative judgment. Where that is all the evidence there is, no claim that the holding works against your mission is asserted at all — the row says what is missing instead.
Source: the evidence gate in the ratified severity methodology the analysis engine reads. It is enforced in code, not only asked for in the prompt.
A judgment is a word from a ratified vocabulary, never a number. Where your commitment carries its own bound — a dollar cap, a percentage floor — the comparison is made under the comparator your own ratified measurement spec fixes. How far the word may then go is governed by a versioned rulebook the analysis engine reads.
That rulebook is real and versioned, and it is not on this page. We do not yet publish it here, and typing its contents in by hand would create a second copy free to drift from the one your report was graded under.
The same gap closes the version question. A report state is stamped with the method version current when it was saved, where an analysis saves one. What is missing is the way back out: nothing publishes that stamp to a page like this one. So this page cannot show you a version to match against your own report, and does not claim one.
Every source we use, grouped by how much we trust it.
Four tiers, highest trust first. A holding’s tier is never picked by hand. It follows from which source produced the classification.
Regulated
6 sourcesholdings with Regulated sources (SEC filings, AYS, PRI, BOCC, verified risk-factor database)
- bond_impact_db
- database_flagged
- database_verified
- impact_registry_certified
- nport_enriched
- third_party_screened
Researched
3 sourcesholdings with Researched sources (web research with citations, curated fund metadata)
- batch_web_researched
- metadata_known
- web_researched
Classified
6 sourcesholdings AI-Classified (Haiku classification or AI inference — no external source)
- ai_extracted
- ai_inferred
- classified
- opus_escalated
- pe_detected
- web_researched_unverified
Unknown
1 sourceholdings with Unknown source (requires additional research)
- unknown
This registry states each source’s tier — what governs how strongly a classification counts. It does not yet state how often each source is refreshed. Some sources stamp a date on each holding: a filing date, a screening vintage, a research date. None of them publish a schedule for the source itself. We will not invent one.
How provenance is labeled
Three tiers, read in order — never blurred into one line.
Cited outranks model-assessed, and model-assessed outranks unknown. Every finding on a report states which of the three it sits at.
1. Cited
A receipt you can open — the evidence itself, complete on its own. A finding is built from this tier directly, never paraphrased into it.
2. Model-assessed
The engine’s own classification, labeled as such. It is never blended into a cited dollar figure — a model-assessed number stays its own number, and never borrows the confidence of a receipt.
3. Unknown
A named absence. When we do not know, the report says so — never a blank, and never a zero standing in for silence.
Five kinds of missing, because they do not mean the same thing.
A missing value is a finding. Collapsing all five of these into one grey dash would hide the difference between a company that declined to publish and a measure that does not apply to its sector.
A missing value is a finding. There are five of them, and they do not mean the same thing.
Every cell in a report is either a receipted value or exactly one of five named absences — enforced in the data layer, not left to editorial discretion. A company’s silence, our unsuccessful search, a measure that does not apply, a gap in the public record, and a vehicle we cannot see into are five different facts about the world. Collapsing them into a single grey dash discards the most useful thing we know.
These states are the analysis working, not failing. Read as a distribution, a full column of named absences is a disclosure record — often the largest single finding in a report.
The company publishes no such figure in any filing, report, or registry submission we searched. This is the only absence that carries weight in a judgment: a firm choosing not to publish what its peers publish is conduct, and it aggregates — "11 of 14 holdings do not disclose" is a finding, not a data problem.
We searched the canonical sources for this measure and found nothing verifiable. This state is about us, not the company — so it always carries the search date and the venues searched, and it expires: a stale search is re-run rather than re-asserted.
The measure does not apply to this kind of business — a software firm has no produced-water disposal. Correctly excluded, not missing. This is the quietest state by design: it recedes almost entirely, and it is removed from every absence denominator, so it can never inflate a disclosure-gap figure.
No registry, filing or dataset covers this measure — for any company, not only this one. Matter-of-fact: nobody could answer it today. Where the commons roadmap has an answer, the state links to it, so the reader learns what would close the gap and who is building it.
A private fund or aggregated bond line with no public look-through. This absence is not about a company at all — it belongs to the coverage contract, sits outside every grade, and is the one state never rendered without its action: the look-through request, addressed to a named manager, with a date.
One column, six holdings, five different facts. The receipted value sits highest — a figure with a citation one hop away. Every absence below it descends in visual weight, and the two absences that are about us rather than the company carry a date or an action instead of a judgment.
Share of each issue’s measured positions, dollar-weighted. The shape of the bar is the finding.
When most holdings in an issue publish nothing at all, that silence is the sector’s disclosure record — not a data problem to fix before a report can say anything.
A report never treats a wall of the same absence as a gap to fill in before it can speak. The silence itself is stated as the position the portfolio currently holds, with the proportionate next step attached — never a divestment question reached for first.
Four cells, four different facts, one grey dash. It hides the sector’s disclosure record, hides our own search dates, and makes an opaque fund look like a company that simply had nothing to say.
Only one band of a report is open to a committee’s correction.
Every report splits into receipted fact, your own ratified commitments, and a model’s draft classification. Only the third is open to correction — and that is the honest size of what you are asked to take on our word.
Yours to correct
- A model-classified judgment behind any commitment row — recorded in the disagreement ledger, attributed and dated, and carried everywhere the holding appears.
- Your own mission and IPS text, through ratification. Amending a ratified commitment is a governance act, not a silent edit, and it is recorded with attribution and a date.
- Where a step on the ladder stands — climbed, deferred, or revised.
Not yours to change
- A receipted fact from a public source — verbatim, dated, and not editable by anyone, including us.
- A judgment that already clears the Regulated or Researched evidence floor — only the model-classified band of a report is open to override.
- The methodology itself: the bin vocabulary, the evidence-class hierarchy, and the rules on this page.
Every correction is entered as its own record — attributed, dated, and never edited or removed once made. A reversal is its own new entry, not a deletion of the old one.
A dated history of this method, prior versions included.
This section should list every prior version of this method, dated, each still readable at its own address. It does not. The versions themselves are recorded: each report state is stamped with the method version current when it was saved, and those stamps are kept. But nothing publishes them back out, and no address serves a prior version of this page.
The exact wording your compliance team would review — block DISC-3.2-A.
A distancing disclosure block, versioned and identified as DISC-3.2-A, is part of this method’s design. It is not shown here in the exact wording a compliance reader would check. The product does not yet keep a versioned disclosure-block registry to publish it from, and a sample would create the impression of a stable, citable text we cannot yet guarantee.
The advisors page states the same gap, for the readers it most concerns.