GiveReady, Pivot snapshot 2026-05-24

The one-paragraph version

The autonomous-agent-write thesis is dead at scale. Six weeks of live attention produced one applied enrichment, and that one came from a manual demo. AI training crawlers (GPTBot, ClaudeBot, Google AI) ingest the directory at high volume (8,313 hits in 30 days, GPTBot alone 8,309). Interactive agents (ChatGPT-User, Claude-User, Perplexity-User) rarely arrive at the site, and when they do, they read and leave. The repositioning: GiveReady is the open, structured, neutral data layer that AI models cite when users ask charity questions, and human-in-loop enrichment via consortium partners builds verified depth on a small named set of nonprofits. Both halves are measurable. Neither requires the autonomous-write magic to land.

What the data showed in the first six weeks

All figures as of 2026-05-23, captured at the moment of the pivot. Numbers will keep moving on the live /progress page; here they are frozen.

Training-corpus ingestion: 8,313 training-crawler hits across the 30 days ending 2026-05-23. GPTBot/1.3 alone hit /AGENTS.md 8,309 times. SemrushBot, Googlebot, Bingbot all indexing normally.
Interactive-agent traffic: 76 interactive-agent hits across 14 days. 0 of those resulted in an /api/enrich submission. The first ChatGPT-User landing on /agents.md happened on 2026-05-23, read and left.
Applied enrichments lifetime: 1. From a manual Claude/3.5 demo on 2026-04-12, not from autonomous agent activity.
Donations: 9 lifetime across the three anchor nonprofits (5 to Finn WEF for $7.01, 4 to City Kids Surfing for $5.75, 0 to Bridges for Music). All from operator wallets, none from third parties.
MCP tool-call counts in inference sessions: effectively zero. Agents do not autonomously call GiveReady tools unless a user prompts them to.
Stripe gateway path: shipped 2026-04-12. No third-party use to date. Charities without USDC wallets still cannot currently receive donations.

What changed since 2026-04-25

The five theories from the hypothesis snapshot, rescored against six weeks of data. Two retired, one revised, two unchanged.

Theory 1: AI agents will discover and act on small-nonprofit data via MCP and AGENTS.md.

As submittedAgents would both read the directory and write back to it. Multiple-agent consensus would auto-promote enrichments; disagreements would be returned as context.

Revised 2026-05-24Discovery half holds (training crawlers reading at scale). Action half retired (zero autonomous writes in six weeks). New version measures citation share: % of AI responses to 10 fixed charity-discovery prompts that include GiveReady-sourced data.

Theory 2: Donors actually want to give via AI assistants.

As submittedDonor intent expressed to an AI assistant would convert into completed gifts once infrastructure was live.

Unchanged 2026-05-24Still unproved. 0 non-operator donations in 41 days across the three anchors. Single biggest open question in the project. Anti-gaming rule: operator wallets do not count.

Theory 3: The gateway payment path works end-to-end for charities without wallets.

As submittedAn agent pays GiveReady, GiveReady remits to the charity bank account via standard fiat rails, the charity sees "Donation via GiveReady" and claims its profile afterwards. Deployable in Months 1-3.

Revised 2026-05-24Architecture shipped 2026-04-12. No third-party agent has triggered a cycle. Now treated as proof-by-pilot, not proof-by-architecture: requires a non-operator person to trigger a real donation that lands in a partner charity's bank account, documented.

Theory 4: The framework holds across UK, US, and South African regulatory environments.

As submittedThree jurisdictions tested in parallel. City Kids Surfing (UK), IRS 990 registry (US), South African cohort selected in Months 1-2.

Unchanged 2026-05-24UK pilot named (City Kids Surfing, Joe Taylor as Charity Sector Lead). US still "via IRS 990 registry" with no named pilot. SA cohort still unselected. Operationally one-and-a-half jurisdictions. Same gap, same plan to close it via consortium work.

Theory 5: B2A scales at zero marginal cost.

As submittedEvery AI assistant that indexes the MCP registry becomes both a distribution partner and a data contributor at zero marginal cost.

Rewritten 2026-05-24Distribution scales at zero marginal cost. Contribution does not (crawlers index, they do not write back). New split: training-corpus ingestion scales for free; enrichment depth scales via consortium-partner human-in-loop work, which is budgeted in v6 at $20K.

The new north-star metric: Citation Share

Of 10 fixed charity-discovery prompts run weekly against Claude, ChatGPT, and Perplexity, how many include GiveReady-sourced data in the response. Denominator 10. Numerator the count of prompts where GiveReady-sourced data appears.

Baseline expected: 0 to 5%. Training-corpus ingestion has a 6 to 18 month lag before affecting trained model weights. RAG-sourced citations (Perplexity, Bing Chat) can land sooner. Tracker outputs land at 01-Projects/GiveReady/citation-tracking/YYYY-MM-DD.md in the project repo. Methodology stays public. Prompts cover three UK youth charities, three US youth charities, two SA youth charities, one surf therapy, one music education.

Kill triggers, locked 2026-05-24

2026-11-23 hard kill. If citation share is still 0% across all 10 prompts at the six-month mark, the training-corpus thesis is also falsified. GiveReady drops to maintenance mode and freed hours redirect to the TestVentures enterprise pipeline. Revisit in Q1 2027 only if B2A market data shows a material shift.

2026-07-25 quarterly accelerator. If citation share is still 0% AND no consortium enrichments have shipped beyond the initial Joe Taylor proof point AND donations are still under $20/quarter from non-operator wallets, accelerate the maintenance-mode decision to 2026-07-25 instead of 2026-11-23.

Anti-gaming rule. At least one positive signal must come from a source the operator does not control. Citations must appear in responses to standard prompts (not prompts engineered to produce the citation). Donations must come from non-operator wallets. Consortium enrichments must come from non-operator people. Operator-initiated activity counts for operational validation but does not satisfy kill-trigger thresholds.

What stays the same from the original application

Architecture. Cloudflare Workers + D1 + R2 + MCP server. Same stack, same ~$50/year cost floor.
Consortium. TestVentures.net as lead. Joe Taylor as UK Charity Sector Lead. City Kids Surfing as named UK pilot.
Budget envelope. v6 budget unchanged. Internal reallocation only: the $20K enrichment line item now funds consortium-partner human-in-loop work rather than autonomous agent infrastructure.
Jurisdictions. UK, US, South Africa. Same three regulatory environments.
Open-source commitment. Code, metrics, governance toolkit publish under MIT licence at end of grant period. Unchanged.

How this pivot was decided

A /plan-ceo-review followed by a /plan-eng-review on 2026-05-23, run as part of the gstack toolkit. Both reviews cleared. 5 of 6 proposals accepted, 1 deferred to TODOS.md. 8 implementation tasks across 3 tracks were assigned definitions of done. Eight eng-review amendments (D1 through D8) refined the implementation specifics. The full CEO plan lives at 01-Projects/GiveReady/2026-05-23-giveready-pivot.md in the project repo.