Fieldstone is a strategic thinking partner — not an answer machine. Its job is to help practitioners, candidates, and decision-makers develop positions that are genuinely theirs, through structured provocation, citation-enforced research, and a quality gate that requires you to state your position before advancing each stage.

The name comes from dry-stone construction. Fieldstones are rough, irregular, imperfect. Stacked with skill and pressure-tested over time, they build structures that last for centuries without mortar. Fieldstone works the same way — your raw thinking, shaped under genuine challenge, into something that holds.

The positions in a Fieldstone session are yours because you earned them against resistance. Not because they were handed to you pre-formed.

In early 2026, a candidate completed a six-task strategic assignment in a Lagos hiring process — a strategy role at a pan-African organisation. The work was produced in genuine collaboration with AI: the candidate brought the strategic instincts, market knowledge, and judgement; the AI provided structure, research, and written execution.

The contribution was not 50/50. The candidate estimated 70% of the strategic direction was theirs. But the AI wrote most of the final language. When the hiring team reviewed the submission, they flagged it as AI-generated and disqualified the candidate — citing "overwhelming reliance on plug-and-play AI-generated responses."

The transcript told a different story. Three moments from it:

None of those moves were AI-generated. They were strategic judgements that shaped what the AI then developed into structured arguments. But there was no way to show this — no audit trail, no record of the pushbacks, no contribution split.

Fieldstone was built to make that record. The contribution tracker, the quality gate, the change log, the voice editor — every feature exists because of what happened in that session. The tool is the audit trail that would have made the difference.

P — Provoked. Every position is challenged before it is accepted. The tool never agrees without testing. Pushbacks are recorded as human contribution events — they are the primary signal that the thinking is yours, not the AI's.

A — Arrival. The goal is a destination, not a process. You end somewhere defensible. The quality gate enforces this — you cannot advance until you can state what you now believe and why. Activity without arrival is just activity.

R — Reasoning. The change log tracks how you got there. Not just the conclusion but the evolution — what you believed at stage one, what changed at stage two, what survived challenge at stage three. The reasoning is the audit trail that makes the position defensible.

Most AI tools make it easier to generate answers. Fieldstone makes it harder to accept them uncritically. The friction is the feature.

Anyone with a high-stakes thinking task where the output has real consequences — a decision that gets made, a submission that gets evaluated, a pitch that gets funded or not.

Particularly effective for African and emerging-market contexts where global frameworks require significant local reinterpretation — the Nigerian market context mode is calibrated for Lagos dynamics, corporate culture, and local precedents.

The space is either academic debate tools, generic SaaS counterargument generators, or bare ChatGPT wrappers. None treat the applied professional context as the primary user — someone with a real deadline and a submission that will be evaluated.

The critical difference: in every other tool, AI is the author or the teacher. In Fieldstone, AI is the challenger. You remain the author throughout. The positions in the output are yours because you defended them — not because the AI wrote them and you edited the language.

Citation enforcement is built into the system prompt. Every factual or evidential claim the AI makes must carry a citation tag in one of three tiers. The tiers render visually inline in the chat so you can see immediately what is sourced, what is approximate, and what is inference.

Each citation is also checked automatically against free sources — the URL is verified live, matched against a curated pool of free-access outlets (official data agencies, open-access scholarly hosts, general and African tech/business news), and cross-checked against real scholarly databases (Semantic Scholar, CrossRef, OpenAlex), SEC EDGAR filings, World Bank documents, or news coverage where applicable. A small badge on each citation shows the result — verified, broken, or unconfirmed — instead of only the AI's own self-reported confidence tier.

For citations to the World Bank or IMF, a second badge goes further than the link: the actual number the AI cited — a figure like "$19.5B" or "22.4%" stated just before the citation tag — is extracted from the sentence and checked against that source's own live data for the same country and year. A live link with the wrong number now shows as a mismatch, not a pass; a claimed year with no data at the source shows as unconfirmed rather than silently compared against the nearest available year. This is narrower than it sounds — it only fires on a handful of common macro indicators (GDP, inflation, remittances, unemployment, population, and similar), and only for Nigeria and a few neighbouring countries — but where it applies, it checks the actual claim, not just that the citation resolves.

Important caveat: citation enforcement and the automated free-source check together reduce but do not eliminate hallucination. The AI can still fabricate sources, and the automated check only covers what a free, no-key source can confirm. World Bank and IMF citations get real value verification against those sources' own live data, as described above — but only for the specific indicators this build curates, not their full catalogues. OECD statistics are checked for link liveness only: OECD's own data API requires precise multi-dimensional query keys rather than a simple lookup, and Nigeria is not an OECD member, so this wasn't built out. Nigeria's own NBS/CBN data is checked for link liveness only as well — neither agency publishes a documented public API, so any value verification there depends on a small, manually-checked reference set rather than a live source, and that set may simply have nothing entered yet for a given figure. Always verify links independently before submitting or publishing. The APPROX and LOGICAL tiers are explicit acknowledgements of uncertainty — treat them as such.

It cannot replace domain expertise. Fieldstone can challenge your reasoning and surface evidence, but it cannot substitute for deep contextual knowledge of your specific market, organisation, or problem. The Nigerian market mode helps — but local knowledge beats calibrated prompts every time.

It cannot guarantee you pass an AI detection test — and that isn't its job. The contribution tracker shows the split; the voice editor is a readability pass that makes AI-drafted prose sound like you, not a way to defeat a detector — the rewrite is still logged as an AI turn, not hidden from the record. If you primarily use Fieldstone to generate text rather than challenge thinking, the output will still read as AI, and it should.

The contribution tracker is approximate. It measures word contribution and pushback events — a useful signal, not a precise measurement. A single sharp insight that redirects the entire conversation matters more than three paragraphs of elaboration. Pasted text is not credited as original thinking (typing it out is) — but this is a heuristic based on paste events, not proof of authorship: it cannot tell your own notes from someone else's, or a genuinely original pushback that happened to be composed elsewhere first and pasted in.

It works best when you bring something real. If you arrive with no instinct at all — just asking the AI to generate one — PAR collapses into sophisticated autocomplete. The method requires you to have a view, however rough, before you begin.

Fieldstone is a static HTML file. Your API key lives in the browser tab and disappears when you close it. Your conversation itself never touches any infrastructure controlled by Fieldstone or Extrafemi Labs.

API requests go directly from your browser to your chosen provider — Anthropic or OpenRouter. Neither provider uses API-tier data to train models by default. See the Anthropic Privacy Policy ↗ and the OpenRouter Privacy Policy ↗.

One disclosed exception: citation verification. When an assistant response contains a citation tag, that response's text is sent to a small function Fieldstone runs, which checks the citation against free public sources (scholarly databases, SEC EDGAR, World Bank, official news outlets). This never includes your messages or your API key — only the AI's response text, and only to check its citations. Your IP address is used for rate-limiting abuse, and results are cached for up to 24 hours to avoid re-checking the same citation repeatedly. Nothing is retained as a transcript.