Procedures

Methods for AI-assisted work, proven on the system that wrote them.

Every procedure here was forged the same way: a friction observed in real use, measured, cured, and only then written down as a standard — distilled from the record, never invented. They run FRACTAL’s own development, and they generalize to any AI-agent-assisted workflow. This section will grow: next up, measuring context rot itself — how project memory degrades and what structure provably protects.

Nobody needs a micro-close.

A wish from a simulated user became a better ritual — by deleting the feature the user asked for.

FRACTAL sessions end with a close: a checklist walked over the living documents, two machine checkers, a handover for the next session, and — when decisions were made — a protocol recording the reasoning. It is the ritual that makes project memory trustworthy. And it has a price, sized for sessions where real decisions land.

During a simulation program, a synthetic user — a content creator running many small sessions — hit exactly that price: for a five-minute session, the full ritual felt absurd. The wish was recorded faithfully: give us a micro-close. A lighter tier, fewer steps, for small sessions.

The obvious cure is a second ritual. The better question, asked by the founder, was: could the close automatically touch only what needs renewal, so that a small session’s close turns out to be a micro-close by itself?

The close was conditional all along

Reading the ritual closely: every row of the closing checklist already carries a refresh trigger; the protocol rule already contains an escape (“or a recorded reason none was issued”); the store checker only fires when the store was touched. The design was adaptive from birth. What made every close cost the same was two manual habits sitting on top: verifying that nothing changed costs as much as changing, and the narrative artifacts default to full size.

The adaptive close, in three rules

File hashes drive the mechanical half. The close tooling knows exactly which files changed; unchanged checklist rows mark themselves.

The decision ledger is the judgment flag. Machines prove what did change, never what should have — but the should-questions are only hard when something decision-shaped happened, and that is mechanically detectable: the ledger moved, or it didn’t. Ledger untouched ⇒ the recorded-reason line is offered automatically. Ledger moved ⇒ full protocol, as ever.

Narration scales with the change-set. A session that changed three files hands over in six lines — and that is a conformant handover.

Under these rules there is no micro-close to build — and nothing to choose. A small session’s close emits the small shape; a decision-heavy session gets the full ritual it deserves. One mechanism, scale-invariant. The feature request was cured by deletion.

The boundary that stays: a close of any size ends its conversation — adaptivity changes the close’s cost, never its meaning. And the judgment on rows that did change stays human; the mechanization program drew that line deliberately.

Of record: decision C-112, riding the diet program’s mechanization slate (C-110). The micro-close wish: finding RF1, refinement simulation, 2026–08–17.

Put the memory on a diet — then prove the pointers hold.

Living documents accrete. The cure is eviction to pointers — adopted only after a blind experiment showed agents actually follow them.

A project memory has two tiers: a frozen record (protocols, dated reports — append-only, complete) and a living tier the agent reads at session start. The living tier accretes: histories re-narrate what the frozen record already holds, ledger rows grow annotations, one decision re-tells itself across six surfaces. Measured on FRACTAL’s own corpus: one register grew 19 KB to 200 KB in five days; one paragraph was growing 460 words a day.

The instinct — write everything compactly from day one — is wrong. Early in a project, rich clustered prose is what lets an agent interpret anything at all. The correct model is a cache: accretion is the default mode at every age, and a diet is the callable eviction pass — cold entries compact down to pointers into the frozen record. Nothing is deleted; the trade is ambient availability for efficiency.

The experiment: do agents follow pointers?

Eviction is provably lossless on disk. The open question was behaviour: does an agent reading a compacted row follow the pointer into the frozen record when depth is needed — or does it stop at the pointer and improvise? So before adopting, the compaction was gated on a blind two-arm test: identical corpora, one with the evictions applied; twelve gold-answer questions across three depth classes; six agents, told nothing of the design, answering from whichever corpus they got; every answer scored for correctness and for whether the frozen evidence was actually opened.

Result

23 of 24 correct. On every task requiring depth, lean-arm agents followed the pointers into the frozen record and matched gold — twice exceeding it. The single failure was the shallowest task: a headline row informative enough to feel like an answer got quoted instead of followed.

The two failure modes the experiment surfaced became the standard’s two gates. The pointer grammar: every compacted row names the concrete file holding its narration, and index tables declare themselves index-only — a row that says what without flagging that the why lives elsewhere invites being treated as the answer. The fact-presence gate: before any row compacts, verify the target actually holds the narration — the adversarial verifier caught one row whose content existed nowhere else; file-existence checks pass exactly this defect. Content born in a living row relocates into the frozen record first, or the row keeps its prose: eviction waits, never forces.

The phase answer falls out of the same gate: a young project has almost no frozen record, so almost nothing may be evicted — rich prose is protected structurally, and the pointer form becomes reachable region by region as the record matures. No mode switch; the gates simply become satisfiable.

Of record: Fractal_Diet_Procedure_v0.1 (decision C-111), distilled from the diet measurement flight (DF1) and the compaction-behaviour experiment (CB1), adversarially verified before ratification — 22 findings, all cured in draft.

The Scan: hunting what the machines can’t see.

Checkers guard what they were built to guard. Drift lives in their declared blind spots — so a commissioned review hunts exactly there.

A governed project runs mechanical checkers — but a checker only guards claims someone taught it. What about contradictions between documents, present-tense claims that quietly became false, mechanisms relied upon but never defined? For that, FRACTAL runs Scans: whole-project adversarial reviews for loose ends not explicitly stated.

Across four Scans the finding count fell 26 → 24 → 8 → 7 while the corpus tripled — the trajectory of drift being converted, catch by catch, into mechanical guarantees.

And the standard holds itself to the discipline it preaches. In FRACTAL, a standard ships in two halves: the prose and the checker — a standard without a checker is a norm, not a gate. When the prose is ready before the machine half, that is said honestly (prose-first) and the checker's build trigger is armed as a hard condition: the Scan standard shipped with the rule that the next Scan cannot close until its conformance checker exists — a tool that mechanically validates the report's grammar, its evidence block, and that every finding actually landed in the project's ledger with a disposition. Until that trigger fires, the obligation sits visibly in the queue; it cannot be quietly forgotten, because forgetting it is exactly the drift a Scan exists to hunt. The standard refuses to become its own blind spot.

Of record: Fractal_Scan_Procedure_v0.1 (decision C-108), distilled from the four executed Scans of 2026–08; the checker gate is its §12.

Multi-agent work under a contract.

Dozens of agents are cheap. Ungoverned, so are their mistakes. The flight protocol makes the swarm accountable to one commissioning contract.

Modern agent tooling will happily launch a fleet — parallel readers, adversarial verifiers, judge panels. The failure mode isn’t capability; it’s governance: programs that launch before anyone agreed what they may touch, scopes that grow silently mid-run, results that arrive as assertions nobody can audit. FRACTAL runs multi-agent programs as ultracode flights under a ratified protocol:

The procedures on this page compose with it: a Scan, a diet, a behaviour experiment can each run standalone — or plug into a flight when the job needs a fleet. The protocol itself was distilled from the programs that preceded it, frictions included: it exists because an early program launched undeclared, and the record of that friction is cited in the standard.

Of record: Fractal_Ultracode_Flight_Protocol_v0.1 (decision C-108), distilled from the flight programs of 2026–08 — including a 67-agent adversarially verified capture program and the experiments described above.