The process.

The steps are drawn in order because they execute in order, but the run does not end where it stops. Audit can return work to Plan. Verify can return it to Research. A failure does not just get fixed — it amends the specification that permitted it. A near-miss that could have taken the device down produced a root-cause diagnosis, which produced two revisions of the operating rules, which now close that class of failure permanently.

Two rule changes are on the record. A session limit was loosened after fourteen clean runs showed it was costing more in restart overhead than it was saving. And when a genre definition was corrected mid-analysis, the first batch was not patched — it was deleted and re-judged in full, which moved the tag from 3 of 14 songs to 9 of 14. That jump is the rule change, not a drift in judgment, and it is recorded as such.

An executing agent gets the authority its task requires and no more. It does not choose its own scope, does not extend the job it was given, does not fix things it was not asked about, and does not act on inference. Anything outside the stated scope is raised rather than performed.

The same principle covers the machine. The whole system runs on a phone, so correctness is only half the specification; the other half is what the hardware can physically survive. Limits are stated before the run starts rather than discovered during it — a floor below which the work stops, thresholds that end a run rather than push the device toward failure, and a hard prohibition on installing anything. On this hardware an install can silently fall back to compiling from source, which costs hours and fills memory that is never released. That has already happened once. A missing capability is a finding to report, not an obstacle to route around.

The agent is capped so that it fails before the device does. A crashed agent is recoverable. A locked phone is not.

• One counter governs everything. Every pass that changes anything takes the next number, and documents, reports and deployments all carry the number of the pass that produced them — so no two current documents can disagree about which state they describe.
• Version numbers are non-contiguous by design. A gap means a pass happened that changed documentation rather than the site, and that is the system working rather than failing.
• A numbered document is never edited in place. The specification is read before every task, so a silent amendment is invisible until it causes damage.
• Edits are surgical. A diff between consecutive versions shows only the substantive change, and nothing still true is deleted — an unsuperseded rule stays, in its original words.
• The convention came from a failure. The previous rule left documentation-only passes with no way to advance, and produced three separate document sets sharing one version number.

• Role and authority — whether the AI is analyst, planner, auditor, executor, verifier, or documentation reviewer. Authority is never assumed.
• Mode locking — analyze-only, plan-only, execute, and verify modes kept separate, so inspection cannot quietly become modification.
• Input contract — the source files, reference documents, permitted context, and the order materials must be read in. Read scope is part of the contract, not a suggestion.
• Hard constraints — protected originals, exact duration, metadata, lyrics, cover art, folder locations, filenames.
• Tool boundaries — which tools may be used and which operations require explicit approval.
• Measurement definitions — metrics, calculation methods, tolerances, pass/fail gates, and how conflicting evidence is handled.
• Output schema — structured reports, JSON, indexes, manifests, change logs, status labels, exact naming conventions.
• Uncertainty rules — fabricated certainty forbidden; INDETERMINATE, escalation, or rejection required when evidence is incomplete.
• Safety and rollback — preserved originals, clean staging, recoverable changes, trash rules, no destructive shortcuts. In the run itself this is a sequenced step of its own, not only a field in the prompt.
• Handoff discipline — concise context packages so a new model or session can continue without stale or hidden assumptions. The rules governing what transfers are set out below.
• Stop condition — what done means, stated so it can be checked rather than argued about. Never "improve the page."
• Decision rules — concrete conditions with concrete responses, supplied in advance wherever a human would otherwise be asked. "Use your judgment" is not a decision rule. This is what makes unattended work possible rather than merely tolerated.
• Interrupt rules — when the agent must halt and ask, and when it must correct and continue, so a run left alone neither stalls on a triviality nor improvises through a consequential ambiguity.
• Resource limits — what the machine is allowed to be asked to do, stated before the work starts.
• Provenance requirements — every claim attributed to its actual source, with unattributed conclusions treated as defects regardless of whether they measure correctly.
• Read scope — the exact material the agent may open for this task, with reading beyond that scope treated as a defect rather than as diligence.
• Standing prohibitions — no acting on inference, no extending the task, no fixing what was not asked about.
• Supplied figures — where a prompt inserts numbers or names into a document, the values are written into the prompt and the agent computes nothing. That removes both the cost of the work and any opportunity to invent a figure.
• Rollback reference — the exact current state and the exact state to restore if the pass fails partway, named in the prompt rather than assumed.

A long conversation is not a free resource. Every session accumulates superseded instructions, abandoned approaches, and stale state, and a model reasoning across that residue will eventually act on something that is no longer true. Context is therefore managed deliberately, as a threat to project integrity rather than as an inconvenience.

• Sessions are ended on purpose — new conversations are started before context degrades, not after something goes wrong in one. Ending a working session while it is still healthy is a deliberate act, not an interruption.
• Only what must transfer, transfers — a handoff carries the authoritative documents and the current state, not the history of how that state was reached. A fresh session inherits conclusions and their evidence, never the arguments that produced them.
• Reading is two operations, not one — material processed by a program is cheap and released when the program ends. Material read into a model's own context is expensive and permanent, and is budgeted accordingly. Most tasks that look like they require reading everything require reading nothing and running one program instead. Confusing the two is what destabilizes a session.
• Read scope is restricted — the agent is told what it may open for the task at hand. Reading beyond scope is a defect, not initiative; it costs context and imports assumptions the task never needed.
• Stale artifacts are removed — superseded reports are cleared rather than left to be found later by a session with no way to know they are old.
• The authoritative state lives outside the conversation — the specification, the current state, the measured evidence and the operating rules are held as files, not as chat history. They survive context loss, model changes, new sessions, and moving the same work to a different AI system entirely. That is the whole point.

Context bloat is not a performance problem. It is an accuracy problem, and it is handled like one.

• Specialized work. Different systems perform different functions inside one larger project. One researches, another analyzes, another challenges the assumptions, another consolidates the documents, another verifies the result independently. The assignments are chosen for the problem in front of them, not fixed in advance.
• Parallel problem-solving. When a hard problem stays unresolved, the workflow inverts: the same problem goes to several systems independently. Their explanations, evidence, proposed fixes, criticisms and alternative approaches are then compared, and useful findings are carried between them and back into the main work. Disagreement is treated as information rather than as something to hide.

The objective is not to collect the largest number of AI answers. It is to reduce dependence on any single model's assumptions, and to improve the quality of the decision a human then has to make.

FretNova is the clearest example. Some development problems on it went past what the primary workflow could reliably resolve — an architectural question that kept producing plausible but wrong answers, an implementation that failed for reasons none of the obvious explanations covered. At that point the problem was widened rather than repeated: several systems examined the same failure independently, their reasoning was compared and challenged against each other, the parts that survived were consolidated, and the result was carried back into the main project. Some of those problems took sustained work across several systems over days. None of them were solved by a better single prompt.

For consequential work, independent agents may analyze the same source without seeing one another's findings. Their results are consolidated separately, the consolidation is independently verified, and execution occurs only after that review chain is complete. Roles remain deliberately separated so that no stage silently expands its own authority or certifies its own work.

Nothing in that chain lets a stage grade its own work. The analyst does not consolidate. The consolidator does not verify. The verifier does not execute. The executor does not certify. And the documents the chain runs on were themselves written only after the source information had been independently checked and reconciled across the wider stack, in a deliberate order, with the human's design decisions, rules and constraints settled and written in before any agent was given a job to do.

The layers are not interchangeable, and that is the point. Planning, execution and record-keeping are held apart: whatever decides what should happen does not carry it out, and whatever carries it out holds no authority to decide what should happen. Execution is bounded by a written specification rather than by judgment in the moment. A third responsibility sits alongside both: when something structural changes, the specification is updated in the same pass, unprompted, and the new version says what it supersedes.

None of the three can quietly take on another's job. The planner cannot execute, because it has no hands. The executor cannot expand its own scope, because the scope arrives in a document it must read first. And neither can amend the rules, because a proposal is not active until it is approved.

None of this is choosing between AI-generated options. The human defines the problem, sets the objective, supplies or creates the source material, establishes the constraints, designs the workflow, decides which systems are involved, weighs their disagreement, directs the revisions, determines what evidence is sufficient, accepts or rejects the result, and maintains the authoritative record everything else is measured against.

On the music side that matters more, not less. Songs are not limited to text-prompt origins. They may begin as lyrics, guitar playing, arrangements, compositional ideas, or material written years before any of this existed. AI enters afterward, in transformation, development, analysis, refinement, production, or verification, depending on what the work actually needs.

• Blinded or independent analysis prevents the second model from merely repeating the first model's conclusions.
• Deterministic measurements create a shared factual baseline that does not depend on model opinion.
• Pre-execution audits inspect a proposed plan before any source file is changed.
• Post-execution verification checks whether the approved result was actually achieved.
• Acceptance is separated from creation: a generated output may still be rejected, revised, or discarded.
• Final human listening, artistic judgment, and release decisions remain outside automatic approval.
• Completion claims are treated as unverified until the artifact itself is re-read. After a batch job reported embedded artwork it had not written, the standing rule became: verify by extracting the result and hashing it, never by trusting a success log.
• Playback hardware is part of the test protocol. A known-defective speaker is disqualified as a QA device; small-speaker translation is judged only on a designated reference speaker.
• The ear outranks the meter. If a file measures correctly and sounds worse, the measurement loses and the file is revised.
• Provenance is checked separately from correctness. A conclusion that passes every measurement can still fail on where it came from; sourcing labels on the published statistics files exist for that reason.
• Inherited claims are re-verified rather than assumed. A published figure, a prior report, or a setting carried over from an earlier session is treated as a claim until it is traced.
• Research and independent checking happen continuously, not once. Most claims that reach a specification were researched and verified before they entered it, and are checked again after the work is done.

A claim of success requires evidence from an actual check, not from the fact that a command finished without complaining. Every kind of claim has a kind of proof attached to it before the work starts, and the proof has to come from the artifact rather than from the report about the artifact.

That is a harder standard than it sounds, because most checks verify something adjacent to what was actually asked. A syntax check does not prove a page renders. A copy that succeeds does not prove the copy is identical. An archive that builds does not prove it holds the right files. So the instruction is procedural rather than technical: ask what could still be broken if this check passed — then go and check that instead.

• The live site outranks any working file. A file measuring part of a published metric is not evidence against that metric. When a computed number disagrees with a published one, the first assumption is that they measure different things — establish what the published figure counts before treating the gap as an error.
• A reproducible artifact outranks the copy written about it. A build's own generated statistics, verifiable against the thing they describe, are the measurement. A page quoting a different number is the thing to fix.
• Superseded is not wrong. Artifacts from an earlier true state are historical records, not errors. Archives are never swept, and finding an old figure inside one is not a finding.
• Restriction by masking, not exclusion. Unreleased material is withheld from public copy by name, but included in every catalog-wide statistic. A privacy rule is not allowed to corrupt the arithmetic.
• Coverage is reported before any dataset is called complete — how much of the catalog appears in it, and in which role. A partial dataset is usable as a floor and describable as such. It is never a total.

Some judgments cannot be settled by a tool — whether a song belongs to a genre, whether a recurring image is a motif. Those are governed by a written rubric rather than by taste, so that the same song judged months apart gets the same answer.

• A gate is checked before anything else, and can disqualify a candidate outright regardless of how well it scores afterward.
• A fixed question set, where every question must be a clear yes. Ambiguous evidence counts as no, and the ambiguity is recorded by name rather than resolved quietly.
• A frozen calibration set: fixed reference examples every batch is judged against, so a later batch and an earlier one mean the same thing.
• An explicit list of criteria that must not be required, each identified as a generator of false negatives, alongside a named list of the failure modes the test exists to prevent.
• No threshold is invented mid-run. A standard changed halfway through makes everything judged after it incomparable to everything judged before.
• No method and no output is ever tuned toward a published figure.
• Confidence is recorded per judgment, and low confidence is expected. A batch with no low-confidence entries is more likely overconfident than accurate.
• An outlier is a finding, not an error. A song that breaks its album's pattern is reported as such; judgments are never adjusted to make a set look internally consistent.

The catalog analysis is the better demonstration of what this page claims, because nobody watches it happen. The work is defined once, in writing, together with the record of how far it has got. A session starts, determines its own position in that record from the document rather than from a person, does the next piece of work, writes down what it did, and stops. No stage is named by hand. Sessions have a budget and end well before they reach it, and an agent that notices it has contradicted itself or lost track of what it already read is required to stop on that basis alone.

Every session produces a report, including a session that changed nothing. A system that demands a report from a run with nothing to report is a system that expects to be audited.