The Exemplar Ratchet

I left a review loop running on Venture World — a simulation of Perch Software, an 1980s games studio with bird-agents, letters, invoices, and a kid tester named Mags — and prompted it poorly to finish at 10am. When it got there, it decided the finish was the next 10am, and kept going. Roughly eight hours of self-directed work. Across the wider build, 338 commits in about five days. Half a Max plan burned. I was flabbergasted with the changes.

The story got more consistent and fluent. Gaps filled themselves. Mags, being a kid, turned up after school and only played a few hours. Her bug reports were hilarious and honest. Disks came back broken in the post asking for replacement, not refund. Letters arrived opened, with envelopes. Newspaper clippings showed reverse print bleed and a crease. The postman brought packages. The birds walked across to talk. They spent the budget and ran an overdraft.

Most people, facing that astonishment, keep the wrong souvenir. They remember the night. They try to replay the steps. Or they A/B a few prompts and call it self-improvement. The useful extract is narrower and more portable:

Now that we have the shape of greatness, we can extract it, rather than repeat it step by step.

This piece is the live-tier version of that extraction: one claim, argued from a recorded run to a reusable template. A loop improves only in directions you can name, and only against a bar you freeze.

What the night was made of

The session record — e37513db, opened Sunday evening on the build machine — does not support “left it running and magic happened.” Eight ingredients show up. They are the first cut of the template.

1. Judgement written before code. The author agent was told it held authority for taste, review and content, and to document so the work could be audited. It produced on the order of 34,000 words across twenty documents, plus four tests every feature had to pass: theatre, business, looking interesting, scripted.
2. Gradient steering instead of a spec. North star: is it interesting and fun. Named weakness: too much playing a game, and not ‘real business’. Named delight: when the magazine cover/review arrived… so perhaps other things have an image. That third line is where the letters, broken disks and clippings came from.
3. As-run representation. Critics read what happened on screen, not what the script intended.
4. An hourly outsider critic — nicknamed Gilfoyle — a stronger model judging progress against the north star, explicitly not judging code quality.
5. Cold-arrival readers: fresh agents with no context.
6. One author holding quality responsibility, reviewing sub-agents’ changes against git.
7. Mechanical gates before every commit.
8. A clock — and this is where the runaway came from. A deadline given as a time of day with no date is temporal blindness in your own tooling. The fix in the template is an absolute datetime plus a spend cap.

Two corrections to recollection matter as much as the ingredients, because they are the argument for keeping the as-run record.

Mags was not discovered overnight. He was designed as non-staff from the start: spawn at the door, play, give a verdict, leave. When labour laws came up later, the agent could point at an earlier episode that already said he was not on the payroll. The world was coherent before anyone checked. That is stronger evidence than a lucky find.

And the pivot belonged to the humans. A question — why is episode one Perch and Post? — led to counting, then to the conclusion that the alpha was better than the generated thirty, then to a reframe: don’t port the alpha, recover how it was produced. The loop polished the work. It did not pivot it.

Gradient steering is the underrated ingredient

Almost nobody writes the delight line down. Specs say what to build. Rubrics say what to score. A gradient says which way is up when the work is already moving.

The pairing of named weakness and named delight tells a loop which direction to improve in, and almost nobody writes it down.

Without the weakness, “interesting and fun” can slide into toy. Without the delight, “real business” can slide into spreadsheets with no texture. Together they are a provocation, not a specification — enough to steer, not enough to over-constrain every branch.

That distinction will come back when we talk about putting lessons in the scheduler rather than stuffing history into the explorer’s prompt.

Know what your loop optimises

On the Venture World record, the critic’s absolute scores were too unstable to use as a gate. The useful signal was contradictions quoted from the as-run log. In effect, the loop was a contradiction hunter.

A contradiction hunter can only push the work one way, toward coherence — consistent, fluent, effortless, gaps filled.

Two consequences follow, and both are load-bearing for anyone about to point a similar loop at company work.

First: it will not find drama or business truth. A coherent world can feel charming and still be theatrically flat. A good player and a good episode are different objectives; coherence sits on a third axis.

Second: pointed at public company copy, the same loop would make “ambition stated as fact” read more smoothly. Coherence is not honesty. If you want a company-box version of this tool, aim it at drift — what the site promises versus what the sales docs say versus what the ledger shows versus what sits in the folders — not at making the marketing more fluent.

Expect junk. A later overnight run ended by keeping useful fixes and removing a speculative change. The template includes a morning review that assumes the loop will produce both.

The exemplar ratchet

Farming exhaust is the practice of treating a run’s by-products as raw material: see the pattern, generalise it, feed it back to fix the machine. Skill-library agents already show why that compounds. Voyager’s ever-growing library of executable skills compounds abilities and eases forgetting.¹ Agent Workflow Memory induces commonly reused routines from trajectories and feeds them forward.² The discipline that keeps that safe is gated promotion — review promotions, not writing. Work on durable agent workflows warns that appending lessons to memory without lifecycle gating lets a single bad update silently corrupt future behaviour; governed evolution puts verification and rollback in front of the write.³

For a creative world like Venture World, the ratchet has three honesty conditions.

1. Freeze the exemplar — and its cost

If the benchmark keeps changing, it stops being a benchmark. Tag the commit. Snapshot the as-run logs and the picture set. Treat one build as the reference copy.

Record what it cost. Roughly five days and half a Max plan is not a vibe; it is part of the bar. Extraction means the same quality for a fraction of that effort. Cost belongs in the benchmark beside quality.

2. Compare pairwise

Skip “rate this out of seven.” Cold-arrival readers get one stretch from the exemplar and one from the recreation, and answer which they would keep watching — and why. Comparative judgement tends to be steadier than absolute scores. The “why” is the next gradient.

Promotion to exemplar stays with a human. If the loop’s own critics decide what counts as greatness, the ratchet drifts toward the critics’ taste.

3. Test transfer

Reproducing a 1983 software house may only prove memorisation. The real test is a different world — a modern company with no CRM, a home-care advocate with folders instead of pipeline stages — judged against the same bar. If the pairs come out even at a fraction of the effort, you extracted the shape, not the costume.

Why the outsider critic worked

Gilfoyle — the nickname is a blunt TV-character licence; the role name is the outsider critic — is worth taking apart because most of its properties transfer.

• Stronger than the author. Most review loops use a peer or something weaker.
• A persona that licenses bluntness. It said what was wrong without the default politeness tax.
• Judged the north star, not the code. Drama and aesthetics were in scope; lint was not the point.
• Read what actually ran, from outside the author’s context. No sunk cost. Output, not intentions.
• No commit authority. It advised; the author decided. Harshness was safe because it could not ship.
• It was fun, so the human actually read it. Entertainment value is a governance feature. A dutiful report gets skimmed. A reviewer you look forward to keeps you in the loop.

Aim the harsh outsider at your own work: the episode, the ebook, the doctrine page. For shared work, switch role. Use a cold-arrival reader in a customer’s voice — I’m a nurse made redundant last month; here’s what I understood from your homepage and where I stopped reading. That is evidence about a reader, not your opinion. It depersonalises the critique.

The same stronger-model outsider also belongs eventually inside company agent teams, where the reviewer often shares model and evidence with the producer. An hourly question — would this customer feel looked after? — restores independence the closer currently lacks.

A finished run is a simulator

A recent line of work, Dream-RSI, makes a narrower and more useful claim than “vary the harness once you know the answer.” Read carefully: it improves the exploration policy — which branches to pursue, how many in parallel, when to stop — while leaving the underlying coding agent unchanged. Its key move is to treat a completed discovery history as a replay simulator. Node outcomes are already recorded, so candidate policies can be evaluated by walking the saved tree without rerunning the expensive online loop.⁴

Only the exploration-policy code changes; the underlying models, evaluator, and execution interfaces remain fixed.

Three takeaways travel.

You already own the trees. Overnight worktree trials and hundreds of Venture World commits are discovery trees. Nodes can be scored after the fact with pairwise-against-exemplar checks. Then you can evaluate stop-or-spend policies offline on a night you already paid for — farming exhaust made rigorous. The smallest honest cousin already on file is a sales rehearsal that burned about nine dollars on retries before escalation existed: a stop-or-escalate policy evaluated offline over recorded runs.

Lessons belong in the scheduler, not the explorer’s prompt. Dream-RSI found that summarising history into directional hints inside the prompt consistently did worse than no guidance — strong semantic priors over-constrain parallel exploration.⁴ That matches steering by gradient rather than by specification. Caveat: the guidance analysis sits inside a discovery setting; for delivery work serving a particular client, context still helps. Do not universalise the finding into “never give context.”

The learned policy got frugal. In one reported trajectory, evaluated attempts per round fell from 110 to 50 while scores kept improving, then spending rose again when progress plateaued.⁴ That is the spend cap the accident lacked — expressed as policy, not as a human watching the bill.

Self-improvement, on this reading, touches the scheduler and never the gates. Whole-harness search is a different idea — closer to Meta-Harness, which treats the harness code that stores, retrieves and presents information as its own optimisation surface.⁵ Keep the distinction clean: replay teaches you how to explore trees you already grew; harness tuning against a gold exemplar teaches you how to change the machine that grows them.

Tuning against a human exemplar

When a human expert has hand-made pages for a live lead, those pages are not a chore to automate away. They are the training signal. The expert stops being the person who does every instance and becomes the person whose taste the system learns.

Rules that keep that honest:

• One exemplar teaches the instance, not the person. Tune on one, test on a second — or show a blind pair for a new case.
• Never measure similarity to the text. You will never converge on wording, and that is not the idea. Ask what the human did that the harness didn’t. That answer is your next weakness-and-delight gradient.
• Use dimensions for pairwise comparison, not 1–10 scores. Absolute scores thrashed on Venture World; win-rate profiles per dimension are steadier, and the “why” steers.
• Mine criteria from the exemplar; confirm by recognition. Asking an expert cold for a scoring list asks for recall. Showing pairs and asking which they prefer asks for recognition — how they actually work.
• Treat the dimension you can’t stand as valuable. Fonts, colour, craft you personally dislike are often the tacit expertise you lack. Judge them on rendered screenshots, not on the HTML.
• Add the dimension that is not about the artefact: after twenty seconds, what does this company do for you, and what would you do next? A cold-reader outcome test is hard to game.
• Decide holistically. Optimising a twelve-point rubric yields bland pages that win every row. Use the rubric to diagnose. Decide with one question: which would you rather receive?
• Split the rubric into families with different owners — craft, doctrine, and where people are vulnerable, care. A page has to win across families, not average them into mush.

Where the shape goes next

Three surfaces, same geometry.

A coherence loop over a whole company — site, sales docs, ledger, folders — catching drift unattended. Aimed at contradiction, not at fluency.

An outsider critic inside company agent teams, asking whether the customer would feel looked after, with no commit authority and a persona strong enough to be read.

Self-improvement as a core mechanic with fast proxy evaluators, because customer outcomes arrive slowly and sparsely: first-pass approval rate, escalations per matter, cost per closed responsibility, human decisions required. Offline policy evaluation over runs you already paid for is the smallest honest start.

None of that requires replaying the accidental 10am bug. It requires naming a direction, freezing a bar, and refusing to let the loop certify itself.

The extraction you can run next week

1. Pull one finished run. List the eight ingredients as they actually appeared — exact north star, weakness, delight, critic brief fragments, cold-reader passes, gates, clock.
2. Freeze the current best artefact. Tag it. Snapshot as-run evidence. Write down cost.
3. Stand up an outsider critic aimed at your own work, and a cold-arrival reader for anything shared. No commit authority.
4. Attempt one recreation under a spend cap. Judge in blind pairs. Keep the “why.”
5. Transfer to a different world against the same bar — or admit you only memorised a costume.
6. If a human expert’s hand-made work is available, ask for fifteen minutes of pairwise picks before you tune anything.
7. Morning review: keep the useful, drop the speculative. Promote only a human is willing to call the new bar.

The overnight miracle was never the product. The product is a harness that can find the next miracle on purpose, cheaper, in a world you have not built yet — because you named the gradient and froze the bar.