What this project refused, and what each refusal cost.

The constitution constrains how the engineer must write, not what the owner must know: explain tradeoffs in plain language and teach as you go. Every option put to them has to be legible without the jargon, or the ruling that comes back is a rubber stamp on a sentence nobody understood — so the rule is enforced against the writer, and a decision taken on an unread sentence is the failure it exists to prevent.

CLAUDE.md

Every request for a decision goes in one fixed block at the end of a reply, and anything outside it does not count: any ask NOT in this block is deemed not asked.

docs/redteam-protocol.md

Each open question carries a default that applies if nobody answers, and the default is constrained rather than chosen: Defaults must be reversible choices, never one-way doors (pricing, data schema, deletions = no default, PM must answer). Silence can therefore only ever produce the undoable option.

docs/redteam-protocol.md

The guardrail this project runs on is not a preference for simplicity. It is written down as a cost: complexity is a cost, not a value — and either party may object by citing it.

restructuring_decision_memo_2026-07-11.md

Work is reviewed in the smallest increment that can be proved on its own, and the reason is written into the protocol: self-review honesty is inversely proportional to the amount of sunk work under review.

docs/slice-protocol.md

Inference — the engineer's reading, not a recorded ruling

Both protocols are aimed at the same weakness, and it is not incompetence — it is ownership. A reviewer goes soft on work they built, so the rules shrink what is under review and force the ask into a place it cannot be buried.

docs/slice-protocol.md · CLAUDE.md

Two documents once described a quality gate that had been abolished months earlier, as though it were still owed. The rule that came out of it is stated as a matter of truth rather than tidiness: a gate nobody performs any more, written as a thing still to be done, is a false statement in the repository.

src/content/retired-gates.test.ts

The repair was not the two sentences. Fixing the two sentences leaves the class open, so the rule became a test that scans every document on every run, proved in both directions, because a guard that has only ever returned clean is not known to check anything.

src/content/retired-gates.test.ts

The same rule now binds the files this site publishes about itself. They described an instrument of eight clips long after it had grown to sixteen, so the quantities are derived from the shipped item pool instead of being retyped: change the pool without changing the sentence and this fails, naming both numbers.

src/content/published-text.test.ts

The honesty rule is stated as a constraint on output, not an aspiration: no score, percentile, or claim the data can't support.

restructuring_decision_memo_2026-07-11.md

Before an estimator is trusted with real answers it is run on simulated ones generated from a known model, and required to recover the known parameters. The claim that buys is deliberately modest: I validated the estimator by parameter recovery before fielding it.

docs/artifact-pivot-2026-08-07.md

Nothing simulated is allowed to pass as observed, anywhere it might be seen: in-app, in charts, in the write-up, in the repo. The badge is not small print. It is the reason the analytics pages are allowed to exist before a single person has taken a test.

docs/artifact-pivot-2026-08-07.md

Where a measurement is noisy the product must show the uncertainty rather than hide it behind a label: report the band, never the point, because a point estimate from a noisy measurement is a claim the measurement cannot support.

src/engine/delicacy.ts

A fifth instrument, to turn preference into words

Refused under N3, and the arithmetic the proposal produced about itself

Inference — the engineer's reading, not a recorded ruling

The four instruments here each have a right answer — damage you can or cannot hear, a label's pull, a critic's gaps. None of them touches the thing listeners actually report, which is that they cannot say what they like. Another was specified for exactly that: you say what you prefer, then choose blind between two versions of the same passage differing in one respect, and the product is the moment your words and your ears disagree. It was approved, sized, and killed by the first slice that did its arithmetic. A preference has no right answer, so the only measurable thing is whether blind choices agree with each other — and that takes twenty-eight of them per dimension from a decisive listener. Three dimensions is eighty-four pairs; at sixty seconds a pair, eighty-four minutes — longer than all four shipped instruments together. It had been sized against a figure a quarter that size, which engineering stated without deriving. Run against the numbers in its own specification, two of its four findings did not survive, the contradiction it existed to deliver among them. Nothing further is added to this product on easier terms than these: an instrument arrives with the arithmetic for its own sitting length, or it does not arrive.

What it cost. The largest one on this page, and it is unpaid rather than accepted. Two findings came out of listening to people: that past listening predicts less than present and forming taste, and that almost nobody can describe their own taste in words. This product serves the first and does nothing at all for the second, which is the one with somebody in front of it. The refusal does not say that instrument was a bad idea — it says this design could not be built at a length anyone would sit, and no better design has been found. Somebody else may well find one.

docs/rt-answers-2026-09-13.md · src/engine/preference.ts

The Taste Gem — a five-faceted picture of your result

Refused under the anti-clone clause, and the writing pass that made it unnecessary

Inference — the engineer's reading, not a recorded ruling

A visual was held back until the product's sentences had been through a writer, on the rule that if the sentences landed the picture was decoration. Three batches of them have now been written, applied and shipped, and every result screen ends in prose rather than in a unit. The picture would add no fact the sentences do not already carry. What it would add is five facets, most of them dark for most people — because a reader has usually taken one instrument, not four — and a shape with slots to fill in is a completion meter however carefully it is drawn. This product refuses those by name.

What it cost. The one thing the product will never have is an image a person can post without reading a word. Every result here has to be read to be understood, which costs the share loop most of its reach and is the second time that trade has been made deliberately: the ranked verdict went the same way. What it buys is that nothing on a result screen can be understood as a score out of five.

docs/handoff-2026-09-01.md · docs/handoff-2026-09-04b.md

The Taste Index — one number standing for a person's taste

Refused under N3, and the ruling on ranked tiers that it would have repeated

Inference — the engineer's reading, not a recorded ruling

The design that opened this phase ended at a single composite over five sub-scores. The five are a percentage of movement toward a label, a detection band, a threshold in cents, a count of distinguished works and a calibration score — five different units measuring five different things. Adding them requires deciding how much each is worth, and that weighting can only be argued from a population this product does not have: the cohort is zero. A number assembled from an unjustifiable weighting is not a summary of five measurements, it is a sixth claim resting on none of them. There is no Taste Index, and there will not be one.

What it cost. The product gave up the one thing it could have put on a share card and in a headline — a single figure a person could compare, remember and repeat. What ships instead is five readings in their own units, each meaningless outside its own context, on five screens nobody has to visit in order. That is a worse product to market and the only honest one available, and it is the same trade the six ranked tiers lost: a sharper claim given up, rather than kept in the hope nobody checked.

docs/artifact-pivot-2026-08-07.md · docs/handoff-2026-09-01.md · src/engine/delicacy.ts

Six ranked verdict tiers on the Delicacy result

Refused under N3, applying RT-90a — report the band, never the point

They shipped first, and then the measurement meant to justify them killed them. Asked how often the six tiers put a person in the right one at the shipping length: 30.5%. No coarser cut rescued it. A tier name is a point estimate wearing an adjective.

What it cost. The result screen lost the one line a person could repeat to a friend and got an interval instead — wider, duller, and true. Earning a ranked verdict honestly would land on ~42–45 trials = 21 min, which is the session 15 was chosen to avoid. The product kept the shorter session and gave up the sharper claim, rather than keeping both and hoping nobody checked.

src/engine/delicacy.ts · docs/handoff-2026-08-22.md

The paid training arc — the entire business model

Refused under the D4 amendment

The plan was to give the assessment away and charge for the training arc. It was withdrawn in one line — there is no paid tier, and no pricing question — because a paywall on the training loop would have put the honest deliverable, whether your ear actually moved, behind the wall.

What it cost. The project gave up its only means of showing that anyone would pay for this, at a point where monetization remains a goal but as proof of commercial viability, not income. It also created upkeep nobody budgeted for: six weeks after the ruling, three published sentences still promised the tier — on two reading-room pages and in the file the product serves to AI crawlers. Writing a rule down does not enforce it.

CLAUDE.md · restructuring_decision_memo_2026-07-11.md · src/content/voice.test.ts

The $3.99 consumer product, and the funnel built to feed it

Refused under memo C1 — a conclusion of record rather than a rule

Viral consumer distribution for a $3.99 impulse product is dead, concluded on twenty-nine visitors across a month, with the World Cup front door spreading to nobody at all.

What it cost. A quiz, a share-card pipeline, a paywall and a Merchant-of-Record payment adapter all became legacy in a single decision. And here is the part that is easiest to leave off a page like this: the paid product itself was never tested (4 paywall views). The verdict was reached on distribution evidence, and the pricing question it looks like it answers was never actually asked.

restructuring_decision_memo_2026-07-11.md

The human ear-check on every audio clip

Refused under a gate only one person can discharge is debt; artifact pivot §1

Quality control was a person listening to each clip and approving it. It was abolished — The PM never judges a clip again — on the owner's own finding: Ear-passes by a non-musician = unstable labels = no value. The gate was not adding quality. It was adding a delay only one person could clear.

What it cost. The replacement has two layers, and the one the pivot itself calls the real gate — item difficulty and discrimination estimated from response data — has never run, because there are Zero real responses. What gates clips today is the acoustic layer alone: loudness, spectral distance, silence, clipping. It can measure how large a manipulation is. It cannot notice that a clip is bad in a way nobody thought to model.

docs/artifact-pivot-2026-08-07.md · docs/blueprint-vs-reality-2026-08-25.md

Speaking about the person, on one surface only

Relaxed: D1 — the product describes what you did, never what you are

Every reading on this site is a statement about a performance. That was a rule rather than a habit: it is written into the constitution as D1, and it is why a five-tap personality verdict with no measurement behind it was killed rather than improved. On 2026-09-16 the owner relaxed it, against the engineering recommendation on file. The card that turns a measured threshold into words a person can use may speak to the reader about themselves. The amendment's own wording was that D1 is suspended for the prompt card, and for nothing else; a second named surface followed a week later, recorded below. Every instrument readout on this site still says only what you did.

What it bought. The one thing here anybody would keep. The measurement ends in a threshold in cents, the number is evidence, and it had been standing in the position of the deliverable — which is why a technically sound instrument was neither enjoyable to use nor convincing to look at. A sentence that is only about a performance cannot be the thing somebody leaves with.

What it cost. This project can no longer say that every sentence it shows is about performance. That was true, it was one of the plainest things the product could say about itself, and it is now false — the exception is real even though it is one surface wide. The constitution also gains an exception, which is complexity it did not have, and every surface built from here has to ask which side of it it falls on. The rule that survives is narrower and harder to hold: offer, do not assert.

CLAUDE.md · docs/rt-answers-2026-09-16.md

A second surface that speaks about the person: the snack

Relaxed: D1 — the product describes what you did, never what you are

On the morning of 2026-09-23 the five-tap music snack was retired so that two sentences on this site would be true. The same day the owner restored it, on the argument that it was the part of the product carrying the product's own thesis — that taste carries cues about feeling, and that the gap between what a person's taste reveals and what they know about themselves is where insight lives. The constitution records the decision in one line: it extends the 2026-09-16 suspension to one more named surface. A reading built from that thesis later is not covered until it, too, is named.

What it bought. The part of the product people could enjoy, and the half of its thesis the instruments never reached. The instruments test whether a listener can hear; the snack is where the product speaks to what a listener might be going through — as a playful verdict that says at its own door there is no measurement behind it.

What it cost. The pivot concluded the five-tap verdict dead, and the reversal above names it as the reason D1 exists. It is back, beside the instruments. The line between a reading about the person and a measurement of a performance is now held only by naming surfaces one at a time, and the card's own disclosure had to narrow from "on this site" to "in the gym". One line did not move: nothing on any surface asserts anything about trauma, abuse or mental health.

CLAUDE.md

The snack retired again, and the reading made the product

Relaxed: D1 and D3 — the product describes what you did, and the Prestige Test was the flagship

Hours after the snack came back, on 2026-09-23, the owner retired it for good and withdrew its exemption from D1. Its questions asked people to describe their own taste, which is the one thing the interviews behind this project found almost nobody can do, and its verdict was written by a language model. The same ruling moved the flagship: the Prestige Test is no longer the front door. A reading of a listener's recent plays is, and the four instruments become the hearing section behind it. The ruling adds the rule the reading is built under: no surface may assert a feeling. And nothing on any surface asserts anything about trauma, abuse or mental health.

What it bought. A front door that shows the idea the project was started for: a listener's recent taste read into lines that can be checked against the plays, argued with, and carried into a prompt. And no sentence on this site is generated by a model when a visitor arrives any more, so every one of them is a fixed template a test can read.

What it cost. The snack was the one part of the product a person could enjoy without headphones, and it is gone. The reading that replaces it at the door runs on three illustrative listeners, so the first thing a visitor meets is simulated plays, labelled as such, where it used to be a measurement of the visitor. And the reversal above now records a decision that lasted less than a day.

docs/rt-answers-2026-09-23-audit.md

2026-09-02 · broke N3 — nothing the data cannot support

Before the retest arc was allowed to tell anyone their ear had moved, the size of change it can resolve was measured: the whole hazard here is that subtracting two noisy numbers manufactures progress. Simulating the same unchanged person through two sessions at the shipped length puts the floor on the pitch ladder at roughly 3.5 times — the threshold has to more than halve before the difference can be told from ordinary run-to-run wobble. On the prestige test it is eight points of the scale. The delicacy trials cannot support an arc at all: six of their fifteen pairs would have to change hands.

Since then. Most retests are therefore told, in as many words, that nothing changed the instrument could hear. That refusal is the ordinary output of this feature rather than its edge case, and the sentence names the floor in the reader's own units so it reads as a fact about the instrument rather than a verdict on them. The only thing that lowers the floor is returning: pooled across four sittings it falls to about two and a half times, which is the entire reward this product offers for coming back.

src/engine/arc.ts · docs/analytics/e14-arc-resolution.txt

2026-08-07 · broke N2 — the anti-theater guardrail

A ruling had already been made: post the flagship instrument on its own, within one to two weeks, and do not let the second instrument gate it. The second instrument got built instead. The plan written that day says it without softening: Delicacy got built instead. That is the N2 launch-avoidance pattern, on the record. And directly above it, the diagnosis: Nothing is blocked by engineering. Everything is blocked by the launch not having happened.

Since then. As of the revision date at the foot of this page, it still has not been posted. The product has had 29 real visitors, ever. There are Zero real responses, which is why every psychometric figure in the Lab is generated from a known model and badged as simulated — the dataset that was named as the project's proprietary asset does not exist. Building is the part that feels like progress, and it is the part that was never the constraint.

docs/endgame-plan-2026-08-07.md · docs/blueprint-vs-reality-2026-08-25.md

2026-08-07 · broke N2 — the same guardrail, applied to the response rather than the act

Inference — the engineer's reading, not a recorded ruling

What happened next is the part that is harder to read, and this reading is mine rather than a recorded ruling. Within the same week the project adopted a direction that made the avoided thing optional: Resume value cannot be hostage to a launch the owner has no energy to run, and after it, The 2026-09-15 deadline is not a live constraint. That argument is sound on its own terms. It is also, in sequence, a project noticing that it was avoiding something and then removing the requirement to do it.

Since then. I cannot tell from the record which of the two it was, and neither can a reader, so the page says so rather than choosing the flattering reading. The test that would settle it is not an argument: it is whether the instruments are ever put in front of strangers. Until they are, the honest description of this project is that it has built four working instruments and measured them against simulated respondents.

docs/artifact-pivot-2026-08-07.md · docs/endgame-plan-2026-08-07.md

A constitution the engineer is held to

Every session starts from one written constitution, and its first standing rule is about the engineer: every proposal must cite the memo decision (D1–D6) or guardrail (N1–N3) it serves.

CLAUDE.md

at work in commit ef33160

Hooks that enforce the loop

Git refuses a code commit that does not carry the slice's north star, three red-team findings and a confession. It checks presence, not honesty. A second hook, run by Claude Code, is the stop between slices: a git commit arms the latch, and file edits are refused until the owner replies. It stands down while the owner's standing auto-advance is in force.

.githooks/commit-msg · .claude/hooks/slice-latch.py

at work in commit 9523984

Skills that quote their rules

Three skills load a procedure into a session at the moment it is needed. The red-team skill packages a procedure every session here already follows from prose; the session close packages the closing rules; the blueprint check packages the blueprint rule of record.

.claude/skills/red-team-slice/SKILL.md · .claude/skills/close-session/SKILL.md · .claude/skills/blueprint-check/SKILL.md

at work in commit 9523984

A reviewer that cannot edit

The red-team skill hands a slice's staged change to a second agent, which returns up to three findings and cannot change the code: you report; the main session fixes. It has no tool that writes because a hook meant to keep it read-only was tested and did not deny a write, and a guard that did not deny a write cannot be the thing that makes a reviewer read-only.

.claude/agents/red-team-reviewer.md

at work in commit 9523984

A server that quotes the record

A small read-only server answers a session's questions about the blueprint, the rulings and the abandoned hypotheses with the files' own words, because a tool that returns the file's own text cannot paraphrase it.

packages/record-mcp/README.md

at work in commit b8c024f