Signal Ledger

What this is

This page is the master record behind the thesis series. Seventeen signals are tracked here, each with a definition, a named recurring source, the cadence that source publishes on, two thresholds — one that would confirm it, one that would weaken it — and its latest dated reading. The theses argue; the ledger keeps score.

Signals belong to the ledger, not to any single thesis. A publication updates the signals it brings evidence for and may add new ones; nothing here obliges a publication to appear. The obligation runs the other way and it is dated: the nine staked calls in section 04 are graded on their resolution dates, whether they land or miss.

• 17 Signals tracked, all currently active, across seven clusters Opened by Thesis 01, 17 August 2026; five added by Thesis 02, 18 August 2026; six added by Thesis 03, 24 August 2026
• 9 Staked calls open, each with a fixed grading date 15 November 2026, 30 November 2026, 30 June 2027, 31 December 2027 and 30 June 2028
• 5 Registries behind the ledger: outcome prices, pricing models, agent acquisitions, adjudicated claims, data deals The data-deals registry went live with Thesis 03, 24 August 2026

How this ledger updates

• Publications update what they touch. A thesis or a note updates the subset of signals it has evidence for. Partial updates are the normal case: a signal a publication did not move is not a signal that was ignored.
• The bi-monthly Scorecard covers the rest. Every two months each live signal is re-read and re-scored, and “no change” is a legitimate result that gets written down. First Scorecard: October 2026.
• Staked calls are graded at their resolution dates. This is the only hard obligation in the system. A call is graded in public on its date, including when it misses.
• Retirement is written down. A signal leaves active status only with a dated line of reasoning. Silent deletion is not permitted, because a ledger you can quietly prune is not evidence of anything.

Lifecycle. Active — read at every Scorecard and by any publication with evidence for it. Dormant — retained, but its source has stopped publishing or the question is parked; no score is issued while dormant. Retired — closed with a dated reason, kept visible in the reading history. Resolved — a threshold was crossed and the question is settled; the crossing is recorded with its date.

Provenance. Every reading carries a date, the publication that logged it and a link to its source. Registry rows carry a disclosure column stating how the number is known: published-primary (the vendor's own rate card or pricing page), company-announced, press-reported, or vendor-undisclosed. Where only secondary reporting exists, the row says the price is unverified instead of repeating it as fact. Scores run −2 to +2 and are mine.

The signals

One table per cluster, one row per signal, each showing its latest dated reading. Net score across the seventeen is +1: pricing and infrastructure move with the theses, the consumer-transaction signal does not, bundling moves against the claim that these layers are independently investable, and the authority signal has been re-scored to neutral with its reasoning printed in the row. The six signals in the new supply cluster net zero between them, which is why the whole-ledger figure does not move: the pricing and disclosure signals there point one way, the institutionalisation signals point the other, and nothing in that cluster is yet a trend.

One convention is worth stating before the tables, because a reader who misses it will read half the ledger backwards. Every score runs against that signal's own confirming threshold, not for or against a thesis. R02.5 sits at −1 because no frontier lab has annexed canonical business state — which is exactly what CALL C (Thesis 02) predicts will not happen. A minus there supports the argument; a minus on R01.1 undercuts one. R03.5 and R03.6 work the same way and their rows print the sign note in the cell. The threshold columns say which is which.

Commerce

Pricing

Margins

Authority

Structure

Energy

Supply

Staked calls

Nine dated calls are open. Two were stated in Thesis 01, section 10, three in Thesis 02, section 10 and four in Thesis 03, section 10. All nine are reproduced here word for word, with the citations renumbered to this page's source list.

Thesis 02 and Thesis 03 each lettered their calls A, B and C, and they are different bets on different signals. On this page the letter never travels alone: every lettered call carries the thesis that staked it.

CALL 1 — agentic checkout share exceeds 5% of agentic activity by Q2 2027 (HUMAN series). Confidence ~55%. A bet against my own tracked data: the series is flat to down across four readings while total agent volume grew. I make it because the authorization infrastructure announced between June and August 2026 has not yet had a shopping season to run through. If the share is still under 3% by Q2 2027, the consumer-side machine-transaction thesis is behind schedule and section 4's ranking is wrong — authority-to-act would be binding harder and longer than I assume.

Signal R01.1 · resolution date 30 June 2027 · graded in the Scorecard that follows it.

CALL 2 — the share of AI builders offering outcome-based pricing as their primary model stays under 30% through end-2027 (ICONIQ series, currently 23%). Confidence ~70%. Unpredictable exception costs push vendors toward hybrid pricing — platform access plus usage, with safeguards — already the pattern ICONIQ observes. The unit is ICONIQ's, not mine: no public series measures the share of contracts signed on outcomes, so the call is stated in the vendor unit that is actually published and will be scored against it. The uncomfortable part, on the record: that series has run 2% to 18% to 23% in four quarters, so the call needs a steep climb to flatten roughly where it stands ^[2]. It is the least comfortable bet here.

Signal R01.2 · resolution date 31 December 2027 · graded in the Scorecard that follows it.

CALL A (Thesis 02) — by 31 December 2027, at least four of the five named harnesses (Anthropic, OpenAI, Google, Microsoft, Meta) will have a generally available shared agent living in a team surface with persistent group state and its own scoped identity. Confidence 75%. Ledger signal R02.1. The baseline makes it demanding: on the strict reading the count today is zero, because Anthropic has not cleared its own bar. Two base rates cut against me — parity in collaboration features has historically taken years rather than quarters, and the convergence was not independent, since a widely read investor essay told the category to build exactly this in December 2025 ^[21]. July's open licences do not resolve it; none of those parties is in the cohort.

Signal R02.1 · resolution date 31 December 2027 · graded in the Scorecard that follows it.

CALL B (Thesis 02) — through 31 December 2027, no member of the fixed cohort (Salesforce, HubSpot, Atlassian, monday.com, Asana, Smartsheet) will report a separately disclosed non-seat agent revenue line exceeding 10% of total revenue. Confidence 65%. Ledger signal R02.2. The rationale is not that vendors are silent, which would be false ^[6]. It is that disclosure exists and scale does not: $1.2 billion against roughly $44 billion annualized is about 2.7%. This is a bet on a reporting convention as much as on economics.

Signal R02.2 · resolution date 31 December 2027 · graded in the Scorecard that follows it.

CALL C (Thesis 02) — by 30 June 2028, no frontier lab will have shipped a first-party, general-purpose store of canonical business objects of customer-record or resource-planning class as a product. Confidence 70%. Ledger signal R02.5. The test of whether the harness annexes coordination capital or keeps renting it. Calls are graded in public at their resolution dates, misses included, at equal visual weight.

Signal R02.5 · resolution date 30 June 2028 · graded in the Scorecard that follows it.

CALL A (Thesis 03) — Reddit's contracted licensing revenue for calendar 2027 will still be under $45m in the third-quarter 2026 Form 10-Q. Confidence 80%. Ledger signal R03.2. The filing is expected around 30 October 2026 ^[16], which falls after the October Scorecard's cutoff, so this call is graded on its own date rather than in a Scorecard. Being wrong would mean the archive is being re-signed at durable term and the amortization reading is too strong.

Signal R03.2 · resolution date 15 November 2026 · graded on its own date, with a dated log entry and an appended ledger reading.

CALL B (Thesis 03) — through 30 June 2027, no United States restructuring adviser, investment bank or claims agent will publish a distinct sell-side offering for estate sales of corporate data corpora. Confidence 70%. Ledger signal R03.6. Graded from firm websites, restructuring-association conference programmes and named press comment ^[20]. Being wrong would mean the asset class institutionalised a year earlier than this reading expects, and it is the single fastest way Thesis 03 becomes obsolete.

Signal R03.6 · resolution date 30 June 2027 · graded in the Scorecard that follows it.

CALL C (Thesis 03) — through 30 June 2027, no frontier lab will name first-party agent-product usage data as a training-data category in a published model or system card. Confidence 70%. Ledger signal R03.5. Graded from cards published in the window ^[19]. This one runs against the popular version of its own argument: if a lab discloses it, “usage is the specification, not the corpus” needs rewriting.

Signal R03.5 · resolution date 30 June 2027 · graded in the Scorecard that follows it.

CALL D (Thesis 03) — the Spirit deidentified-data sale will be approved by 30 November 2026 with at least one employee-data condition attached — a carve-out, a review protocol, or a use restriction — that is absent from the sale agreement filed on 14 August 2026. Confidence 60%. Ledger signal R03.1. Graded from the signed sale order on the case's public docket ^[14]. This bets on the terms, not the outcome: if the sale is approved as filed, or is not approved by that date, it is a miss and will be graded as one.

Signal R03.1 · resolution date 30 November 2026 · graded on its own date, with a dated log entry and an appended ledger reading.

All nine calls are scored publicly in every subsequent publication, including if they are wrong. The first grading date is 15 November 2026.

Registries

Five registries sit behind the ledger. Each was a byproduct of writing a thesis rather than a separate data project, which is why it can be kept honest: the rows accumulate with time, and none of them can be faked retroactively without breaking a dated source link.

• Outcome price registry10 rows

  What vendors charge for a completed outcome — a resolved conversation, a qualified lead, a handled procedure — with each vendor's own definition of “resolved”, and the per-conversation and per-action meters kept alongside as counter-examples.

  Published rates and press-reported estimates are marked separately; unverified numbers say so.

• Pricing model registry38 rows

  How agent vendors meter, one row per product: access seats, consumption credits, per action, per conversation, per outcome, per session-hour, per container-hour — and which of those a single vendor runs at the same time. The wider view of the same question the outcome price registry asks in one category. The registry runs across the frontier harnesses, their APIs, two work-management vendors and four open-licence shared-agent projects. Three more were added on 19 August 2026, all from one infrastructure vendor's launch week: a usage-reporting endpoint with no price, agent tracing free in beta with a dated rate behind it, and a retrieval service carrying preview rates.

  Every row carries a price-disclosure column: published-primary, press-reported or vendor-undisclosed. The working file also holds a dated re-verification of the Claude Enterprise row against its 17 August seed; the ledger renders the later observation, not both, which is why the table shows 38 rows where the file holds 39.

• Agent M&A registry11 rows

  Acquisitions in the agent stack, tagged by layer — authority, assurance, execution-outcome, execution-context, rail-routing — with announced date, status and status date, so the bundling signal can be read from rows rather than from memory. Five of the eleven rows sit in the authority layer, and between them they hold the registry's whole range of price evidence: CrowdStrike/SGNL, the only row whose price comes from a filing rather than a release; Cisco/WideField Security, whose price no party has ever named and no outlet has ever reported; and Okta/Permiso, the cleanest fit anywhere here to R01.5's own wording. Two rows were added on 25 August 2026: Palo Alto Networks/Portkey, which gives the rail-routing layer a second buyer three months ahead of Stripe/OpenRouter and at a price no party has published; and Dynatrace/Arize, which opens the assurance layer and is deliberately not counted in R01.5, because Dynatrace is neither an IAM nor a hyperscaler incumbent.

  Prices are labelled company-announced, press-reported or undisclosed; nothing is stated as a fact the buyer never confirmed.

• Verdict registry6 rows published

  Public claims about this territory, adjudicated with the claim's owner named and quoted: who said it, where, on what date, what the verdict is — yes, no, refine or retired — and the reasoning behind it. Opened by Thesis 02 on 19 August 2026, and live from that date.

  A claim cannot be adjudicated here until its owner is quoted, because arguing with an unattributed claim is arguing with a strawman. The working file holds eight rows; two are earlier drafts whose owners could not be established and they are recorded as superseded rather than deleted.

• Data deals registry5 rows

  Transactions in which a body of data changed hands — one row per deal, sorted by channel: acqui-data, salvage, brokerage, licensing. Each row carries what was bought, what was paid, how the price is known, where the transaction stands, and the condition attached to using the data. Opened by Thesis 03 on 24 August 2026.

  Every row carries a use restriction, because a price without its use restriction is not a price: the same corpus, offered to the same bidder on different de-identification terms, cleared at different numbers. Status is a column, not a caveat in prose — a contested sale is never rendered as a completed one.

All five are printed below as they currently stand. In the four registries that carry numbers, each row carries a disclosure label saying how the number is known, and one evidence badge that follows from it: MEASURED where the vendor publishes the price itself, COMPANY CLAIM where a company announced a figure it never published as a rate, PRESS-REPORTED where the figure is known only from journalism — the outlet and how it sourced the number sit beside it — and INFERENCE where the number is my own arithmetic or judgement, including a row that records what a vendor has not published rather than a rate. An unpublished price is written as unpublished, not filled in from reporting and left to look like a rate. The verdict registry carries no prices: its rows carry the claim's owner, the evidence class of the claim itself, the verdict and the reasoning.

Outcome price registry

Two of these ten rows were taken by a larger vendor within five months: Forethought closed into Zendesk on 26 March 2026, and Fin sits under a $3.6B Salesforce agreement announced 15 June 2026 that has not yet closed. The category is being priced and bought at the same time.

Pricing model registry

Twenty-eight of the thirty-eight rows publish a rate; ten do not. Eleven run two or more meters at once, which is the finding this registry exists to hold: outcome pricing is arriving as an added meter rather than as a replacement for the old ones. Three of the twenty-eight publish a rate of zero — the shared-agent surface under an open licence — and one hosted vendor publishes the credit value sitting under each seat, which no frontier harness does.

Agent M&A registry

Six of the eleven are completed; the other five are announced, agreed or pending close. Three carry a price the buyer itself put its name to, and only one of those three comes from a filing rather than a release. Signal R01.5 reads this table, not the other way round — and reads only part of it: the rail-routing, execution-context and assurance rows sit outside its two-sided test, so neither row added on 25 August 2026 moves its absorption count.

Verdict registry

The fourth registry — public claims adjudicated with their named owners, verdict and reasoning — went live with Thesis 02 on 18 August 2026. The rule that kept it unpublished until then still governs every row: a claim cannot be adjudicated here until its owner is quoted, because adjudicating an unattributed claim is arguing with a strawman. Verdicts are yes, no, refine or retired, and a retirement is written down with its date and its reasoning rather than deleted.

Six rows publish; the working file holds eight. The two that do not publish are earlier drafts of the third and second rows above, written before an owner could be established for either — one an ownerless first-claim, the other a constructed composite with no single speaker. Both are recorded as superseded with their dates rather than deleted, which is the same rule that keeps the retired row visible here.

A second candidate for this release cleared its dating condition on 24 August 2026 and was still not run: the published thesis carries one verdict block, so no row is opened for a claim that was not adjudicated in print.

Data deals registry

The fifth registry — transactions in which a body of data changed hands — opens with Thesis 03 on 24 August 2026. It exists because there is no other public record of what a corpus sells for: accounting does not recognise internally generated data, so until a company is sold there is no observed price for the thing it produced while working. Each row carries the channel it came through, the price and how the price is known, where the transaction stands, and the condition attached to using the data. Channels are acqui-data (buying into the manufacturer), salvage (an estate sale), brokerage (buying the intermediary) and licensing (a contract with a living seller).

Five rows publish and the working file holds five: this registry opens with everything it has. Two disciplines govern it. The first is that a price is meaningless without the condition attached to it, which is why the use-restriction column is a column. The second is that status is printed in the table rather than explained in prose, so a contested sale is never rendered as a completed one.

Open core

The current state of this ledger is public and stays public: definitions, thresholds and the latest reading for every signal, free to read and to cite with attribution. That is the core, and the core is never closed.

What may eventually be sold is depth rather than access: the full reading history behind each signal, machine-readable exports of it, a queryable interface over it, alerts when a threshold is crossed, and work applied to one organization's own question. If that layer ever exists, this page keeps working exactly as it does today.

Corrections, better sources and threshold arguments are welcome on LinkedIn.
The Agents & Margins briefing is launching — until then, follow on LinkedIn

Sources

1. 1.HUMAN Security — State of Agentic Traffic, June 2026 edition, published 6 July 2026, with the April and May 2026 editions. No July 2026 edition had been published as of 17 August 2026.
2. 2.ICONIQ — 2026 State of AI Report: The Builder's Economy, July 2026 (~300 executives), and State of AI: Bi-Annual Snapshot, January 2026 (N=297 for pricing, N=269 for margins).
3. 3.Growth Unhinged — The 2026 State of B2B SaaS and AI Monetization Report, published 13 May 2026; more than 230 self-selected B2B software and AI companies, fielded April–May 2026.
4. 4.Published vendor pricing pages and product documentation for the five vendors tracked by R02.3, all read 18 August 2026 — Anthropic (Claude Tag documentation) and Anthropic pricing, Asana, Microsoft, plus the Google Workspace and OpenAI pricing pages, which returned a non-USD currency and HTTP 403 respectively to the research workstation and are marked author to verify.
5. 5.Bessemer Venture Partners — The State of AI 2025, 13 August 2025.
6. 6.Salesforce — Q1 FY2027 results, 27 May 2026, taken as the anchor release for R02.2; the other cohort disclosures behind the reading are dated in the row itself, from 30 April to 10 August 2026.
7. 7.NIST NCCoE — concept paper, Accelerating the Adoption of Software and AI Agent Identity and Authorization, posted 5 February 2026; comments closed 2 April 2026.
8. 8.Agentic AI Foundation (Linux Foundation) — Model Context Protocol specification 2026-07-28, with authorization hardening; Microsoft Entra Agent ID reached general availability in April 2026 and Agent 365 on 1 May 2026.
9. 9.Notion — release notes of 1 July 2026, external agents at general availability, taken as the anchor release for R02.4, with the Microsoft Work IQ general-availability note of 16 June 2026 and the Agent2Agent membership announcement of ~9 April 2026 behind the same reading.
10. 10.Zendesk — completed acquisition of Forethought, 26 March 2026, taken as the anchor release for the bundling signal; the other transactions behind the reading are recorded row by row in the agent M&A registry above, each with its own source and disclosure label.
11. 11.Anthropic — Claude Tag launch post, 23 June 2026, taken as the anchor release for R02.1; the other four harnesses in the census are read from their own newsrooms, changelogs and product documentation and dated in the row. The two later readings, dated 26 and 27 August 2026, rest on two further primaries: the founder statement posted on X on 26 August 2026, and the deepagents-code release notes, read together with the release notes and product documentation of the other fifteen harnesses in the sixteen-harness census.
12. 12.Anthropic — Claude for Small Business, 13 May 2026, taken as the anchor release for R02.5, read together with the Claude Tag documentation on per-channel service accounts and the vendor-hosted sandbox.
13. 13.Gartner — worldwide AI-optimized IaaS spending forecast, press release 10 August 2026.
14. 14.ECF 1463, Notice of Auction Results for the Deidentified Data, filed 14 August 2026 in In re Spirit Aviation Holdings, Inc., et al., No. 25-11897 (SHL), United States Bankruptcy Court for the Southern District of New York — the results table ($10,000,000 successful, $7,500,000 alternate), the mirror alternate sale agreement carrying the identical referential-integrity clause, and the exhibit naming the alternate bidder's own upload facility. Filings are free at dm.epiq11.com/case/spirit. The sale was not approved; the hearing is set for 9 September 2026.
15. 15.SimpleClosure — Asset Hub launch, 16 April 2026, the source of the roughly 100 transactions, the $10,000–$100,000 band and the description of the buyers. The company's H1 2026 shutdown report of 20 August 2026 carries no Asset Hub transaction figures at all, so those three data points remain sourced to April. Company claims.
16. 16.Reddit, Inc. — Form 10-Q for the quarter ended 30 June 2026, filed 31 July 2026, and the Form 10-K for FY2025, filed 6 February 2026: the remaining-performance-obligation balances behind R03.2, the note that the balance consists primarily of long-term content licensing, and the renewal risk language.
17. 17.The metered and per-use tier, read 24 August 2026 — Cloudflare's pay-per-crawl documentation (“closed beta”, last updated 28 July 2026) beside its pay-per-use direction statement of 1 July 2026; Microsoft's Publisher Content Marketplace announcement of 3 February 2026; ProRata's published Gist split; Perplexity's Comet Plus publisher pool; and TollBit's realisations, which reach the public only through journalism. The cohort of six is fixed at first reading and does not change between Scorecards.
18. 18.The buy-side intake pages, read 22–24 August 2026 — micro1's data-partnerships page (tiers at $100k, $500k and $1m-plus, a $5m referral programme paying up to $25,000 per company introduced, and eligibility criteria of 30 or more employees with documentation in English) and Mercor's data page (referral payments up to $100,000). Company claims; both companies buy the supply they describe.
19. 19.Anthropic — Claude Opus 5 system card, 24 July 2026, the source of the four named training-data categories and of the roughly 1.5 million reinforcement-learning episodes across thousands of training environments, read together with the consumer-terms update of 28 August 2025 that has permitted training on assistant sessions since. Company claims; the other four cards in the R03.5 census are read from their own publishers.
20. 20.Spencer Fane — The New Distressed Asset, 7 July 2026, the strongest adjacent evidence on the sell side and a practitioner claim rather than an offering: “a genuine buyer market that did not exist five years ago” for accumulated corporate data pools, written without reference to any named case. Law-firm publishing is marketing, and the row that cites it says so.
21. 21.Andreessen Horowitz — Big Ideas 2026: Part 1, 9 December 2025, vertical-AI section; the source of the claim adjudicated in the verdict registry and of the base rate cited against CALL A.