The Three Clocks of a Learning System

Bronze remembers events, the queue manages uncertainty, and gold preserves meaning. A mature system knows more — so it needs to look at less.

Scott Farrell · LeverageAI · Long-form article

Most intelligence systems get more expensive as they learn. More records mean more search hits, more candidate matches, more alerts, more model passes, more things a human must review. Storage is cheap; continuous re-thinking is not. If one store is asked to remember everything that happened, keep every open question warm, and rewrite durable understanding on the same schedule, the system eventually spends most of its cognition re-reading what it already knows.

That failure is not a model problem. It is a clock problem.

A continuously operating personal intelligence system — the kind that scrapes the world, holds open cases, and compounds a private wiki — needs three layers that tick on different times. Bronze grows with what happens. The queue grows with what remains unresolved. Gold grows only when understanding changes. Those are radically different growth rates. Forcing them onto one clock is what makes always-on systems explode.

A scalable intelligence system does not minimise what it stores. It minimises what it must keep thinking about.

This article is the economics of that separation and the artefact that makes it operational: a three-clock growth-and-cadence model specified to the point of implementability, three worked curves from a real decision trace, and an honest line between what those curves show and what has not yet been measured. It sits inside the larger object named in The Semantic Market Model; it does not re-teach that whole-object design.¹ It reuses the Signal-Case Queue’s review-priority formula and Cache the Significance’s dear-versus-regenerable split by name rather than re-deriving them.^2,3

What it is not: a medallion data-lake tutorial, a full signal-case lifecycle, or a summary-caching doctrine. Medallion diagrams answer how clean a table is. Three clocks answer on what timescale each layer may grow, and what scarce resource growth is allowed to consume. If you implement three storage buckets and still reprice every bucket hourly, you have shelves and one clock — and the doctrine has already failed.

One store cannot serve three clocks

Start with temporal permission, not storage fashion.

Bronze is detailed, fast and cheap to append. Its natural size tracks reality’s volume. That is a feature if nothing else is forced to re-think at the same rate. Keep the bronze: deletion is the irreversible operation; re-indexing and re-deriving a map can wait if the territory still exists. The catastrophe is not a large bronze. It is a gold layer or a queue required to track it.

The queue is selective, temporary and operational. It is not a backlog of unread items and not a second archive. It is a market for future cognition. A case stays on the active frontier only while another unit of attention still has expected value.² Uncertainty time ticks when corroboration arrives, first-party evidence lands, expected developments fail to appear, or a case cools after null re-observation. Arrival order is a polite lie about attention. Significance is as-at now.

Gold is compressed, slow and expensive to earn. It ticks when meaning changes — not when a post arrives, not when engagement bumps. When a concept’s role shifts, an attribution corrects, a relationship becomes load-bearing, or a durable lesson consolidates. That is a different information product from “what did this article say?”, a question that is mostly regenerable from bronze if you kept the bronze.³ Gold addresses reality rather than containing volatile detail: topology and routes in the map, figures and session noise read from bronze when needed.⁴

The system owner’s early instinct is the load-bearing move: keep bronze out of gold. Gold holds what is important and what the meaning is. Detail stays in bronze. That is what stops gold blowing up — not cleverer summarisation prompts, but a hard refusal to let event-time volume write into meaning-time storage.

Couple any two clocks and you recreate a firehose at the higher layer. Bronze locked to gold turns every observation into a wiki page: a second firehose that cannot answer what understanding changed. Queue locked to bronze creates a case or a permanent poll for every source row: cognition cost scales with archive growth — the always-on tax. Queue locked to gold keeps settled doctrine hot because the feed still mentions it: the operator is interrupted about ideas already held while open uncertainties starve. Healthy coupling is asymmetric. Bronze may grow without consulting gold. Gold may point into bronze without containing it. The queue may read both without letting either dictate its size. The only growth rate that must stay bounded is the active attention frontier, because that is the rate that spends judgement under time pressure.

Three compressions keep cognition bounded

Clocks without compression ratios still explode. Each stage must be allowed to throw almost everything away as long as the right remainder survives at the right timescale.

Source compression. Many articles, posts and secondary write-ups collapse into one evolving case — the unit of unfinished understanding, not a post ID and not a timeless concept page. Attach is preferred over create when the unfinished episode is the same. The world can emit ninety objects about one incident; the queue should still hold one open question with a growing evidence set. Evidence-set size may track the world. Active case count should track genuinely new unresolved episodes.

Attention compression. Many open cases collapse into the small set worth observing now. Priority factors and diverging cadence ladders decide who earns the next look. You are not committing to scrape every case every hour forever. You are committing to a bounded series of increasingly sparse observations unless new evidence reheats the case.²

Meaning compression. Many observations and case transitions collapse into a durable concept, relationship or learned principle — or into no gold write at all. One hundred articles might produce one evolving case, two newly important relationships, one corrected attribution, and perhaps no durable gold change. Gold should scale with meaningful worldview deltas, not source volume. “No mutation” after a noisy day is a successful outcome when understanding is stable.

WORLD VOLUME
    ↓  source joining and deduplication
CASE FRONTIER
    ↓  heat, uncertainty, cadence and fading
ATTENTION FRONTIER
    ↓  interpretation and consolidation
DURABLE MEANING

Cache the Significance owns the storage gate this pipeline needs: description is regenerable; significance is compiled judgement.³ Aim gold at the dear layer. Require significance claims to keep resolvable pointers. Brilliance with a citation is archive; brilliance without one is marketing. The new claim here is not that split. It is the growth-rate theorem that makes the split economically necessary under continuous operation.

The cadence model — implementable contract

Prose about “fading” is not enough. A competent engineer should be able to code a first controller from this section alone. What follows is the primary artefact: objects, fields, priority, ladders, and six rules with predicates and verb sequences.

Objects and required fields

Bronze record (event time). Immutable after write. Corrections are new rows, not silent edits.

• observation_id, observed_at, source_identity, payload_ptr
• Optional: engagement snapshot, content hash, conversation identity
• Clock tag on write: event

Case (uncertainty time). Stable identity for an unfinished episode of understanding.

• case_id, state (open | significant | residual | resolved | expired)
• Factor fields: importance, uncertainty, expected_new_information, time_sensitivity, retrieval_cost
• heat (derived or stored band), evidence_ptrs[] into bronze
• review_targets[] — each with its own next_review_at and ladder position (origin thread, first-party page, secondary corroboration, and so on may differ)
• wiki_neighbourhood — pointers into gold concepts that make personal significance computable
• consecutive_nulls, residual_until, dirty (boolean)
• decision_log_ptr — append-only: attach, create, reprice, fade, residual, resolve, reactivate, push, gold_write, suppress
• Clock tag on operational updates: uncertainty

Gold mutation (meaning time).

• mutation_id, mutated_at, typed change (concept | relationship | attribution | principle)
• understanding_change — mandatory one-line field; empty string is a hard fail
• support_ptrs resolving to case and/or bronze
• Clock tag: meaning

Clock tags are not bureaucracy. They are how you debug a cost spike: did model spend rise because event volume rose, because the frontier widened, or because gold started churning?

Priority formula

Reuse the Signal-Case Queue market; do not invent a rival score.²

priority =
  importance
  × uncertainty
  × expected_new_information
  × time_sensitivity
  / max(retrieval_cost, ε)

Read each factor as a decision lever:

• Importance — if this resolved against active work and canon, how much would it matter? Alone, it produces anxious forever-hot topics.
• Uncertainty — how unsettled are claim, attribution or implication? Alone, it produces thrashing when the world is silent.
• Expected new information — will another look change anything? This is the fade lever. It falls on nulls and pure mirrors; it rises on dirty events that change evidence class.
• Time sensitivity — does usefulness decay in hours even if the concept will live for years?
• Retrieval cost — price of the next observation and judgement relative to competing cases. Expensive low-yield loses to cheap high-yield.

Policy may band or log-scale factors. What is not optional is multiplicative structure: low expected new information drives priority toward zero even when importance is high. That is how null re-observation earns fade without pretending the topic became unimportant.

Worked examples of the market without invented magnitudes help more than another formula rewrite. A high-importance case with low uncertainty and no expected new information — “we already know, and nothing is moving” — waits, even if the topic is strategically central. A medium-importance case with high uncertainty, rising independent lineage, and a closing action window jumps the line. A cheap sensor check with high expected yield beats an expensive full-page re-ingest that is unlikely to teach. Rank without expected yield is prestige. Rent is economics: ongoing cost justified by ongoing chance of learning.

Promotion is not ingestion. Ingestion writes bronze. Promotion places an unfinished episode on a ladder because serial attention has expected value. A new bronze row can attach to a residual case without promoting it if expected new information remains near zero — pure duplicate mirror, no evidence-class upgrade. Conversely, a single first-party object can promote a new case straight onto a hot ladder because importance, uncertainty and expected new information spike together. Resolution is the clean exit when uncertainty has collapsed; the case leaves the frontier, may deposit gold under Rule 6, and keeps trajectory and bronze. The pitfall is resolution-by-deletion: drop the case, drop the bronze, keep a vague summary. You save a row count and destroy late-bloomer audits, reactivation and honest evaluation.

Schedule ladders

One formula must produce diverging cadences. If the hot cascade and the mild post share an hourly scrape, you are running a tax, not a market. Ladder step lengths are policy constants; the shapes are not optional.

Review targets inside one case may sit on different steps while sharing one case_id. Do not collapse them into a single stupid poll of “the case blob.”

Six rules — predicates and verb sequences

Rule 1 — Enter.

on new bronze:
  if honest_join_exists(case):
      attach(evidence → case)
      log(attach, reason)
  else:
      create(case)
      init factors; choose hot|mild ladder
      set next_review_at
  never write gold from enter alone

Honest join means same unfinished episode, not shared topic tag. Over-joining hides distinct uncertainties; under-joining floods the frontier. Log the reason either way.

Rule 2 — Reprice.

on due review or dirty wake:
  collect deltas (scout-cheap where possible)
  recompute factors and priority
  advance, hold, or promote ladder position
  append decision log
  set dirty = false
  if interrupt_threshold_crossed and personal budget allows:
      push
  else:
      silent-handle on frontier

Reprice updates the frontier. Interrupt is a separate rendering under an interrupt budget. Many significant internal state changes should remain silent-handle.²

Rule 3 — Fade.

if observation is null or low_yield:
  consecutive_nulls += 1
  step ladder down
  reduce expected_new_information
  if consecutive_nulls >= N:   # policy, e.g. 3–5 at current density
      state = residual
      clear engagement-driven polling
      enable event-driven dirty wake
      optional residual_until
  do not delete bronze
  do not delete prior gold
  do not destroy case_id (must stay attachable)

Fade answers “do we still watch?” without answering “did it happen?” or “does the lesson still matter?”. Cliff-delete is not fade. Fade is a visible descent with reasons.

Rule 4 — Resolve.

if uncertainty collapsed
   (answered | window closed | permanently out of scope):
  leave active frontier
  if worldview_delta exists: apply Rule 6
  retain trajectory + decision log
  bronze remains

Resolve is lease end. Fade is rent reduction while the question may still be open. Do not collapse them into one “close” button.

Rule 5 — Reactivate. Residual without reactivation is amnesia with extra steps. Reactivation without fade is a firehose. Encode priors as named predicates, not folklore:

reactivation_priors (any may fire):
  is_first_party_reconstruction_or_disclosure
  is_independent_second_lineage      # second victim / lab / community
  contradicts_case_interpretation
  intersects_active_work            # tests what owner is building/claiming
  operator_flagged

on prior true for residual|mild case:
  dirty = true
  attach(evidence)                  # prefer not to create parallel case
  reprice (Rule 2)
  allow ladder promotion to hot
  push only if interrupt threshold crossed

Non-priors that should not reheat alone: pure engagement thrash, duplicate aggregators, keyword hits with no evidence-class change, celebrity amplification that adds no facts. Evidence class beats popularity as a reheat signal. A first-party technical timeline is a different object from the twentieth press paraphrase. Independent corroboration of a second affected firm is a different object from another comment thread restating the first story.

Rule 6 — Gold gate.

write_gold only if:
  understanding_change is non-empty
  AND support_ptrs resolve
  AND change is not regenerable description of surviving bronze

usually qualifies:
  corrected attribution
  new durable relationship
  material shift in concept readiness / threat class / personal relevance
  lesson that changes next behaviour
  contradiction of a load-bearing prior that must now live in the map

usually fails:
  another restatement of a known claim
  engagement thrash
  regenerable article summary
  "people are talking about X" when X is already on the map

A well-shaped judgement pass evaluates a semantic delta, not the whole archive:

Deterministic collector:
  new engagement deltas, new source objects, participants, prior case state

Queue:
  why this case is due; which wiki neighbourhood matters

Wiki:
  relevant existing worldview

LLM:
  did meaning change, and what — if anything — should update?

When cadence fades, you are not “turning AI off.” You are refusing to buy a judgement the world has not earned. Cheap collection and expensive senior judgement should not share a schedule: residual weekly checks should almost never burn senior tokens without a dirty wake; hot ladders may collect often and judge only when evidence class may have moved.

Two false economies the rules refuse. Delete to save money destroys late-bloomer reactivation and honest evaluation. Keep bronze; withdraw attention. Keep everything hot so you never miss is rational without residual and reactivation; with dirty wake and the priors above, it is optional. Being quiet is great. Being quiet forever is not. Being noisy is no good either. Cadence is the tuning surface between those failures. Heat and fade constants often start as guesswork; they do not have to stay folklore — but do not pretend multi-month optimisation already happened when it has not.

Minimum vertical slice: bronze append API; case table with the fields above; scheduler applying Rules 1–6; two ladders plus residual; gold write that hard-fails on empty understanding_change; decision log. Version policy constants beside the code. That is enough to run the curves below.

Worked cadence curves — what is measured

Honesty about proof: three curves are available from a real decision trace on a continuously operating personal intelligence system. They are narratable trajectories. They are not a multi-week simulation of bronze growth versus model-call counts. They do not substitute for the proposed instruments named at the end. Walk them as proofs of behaviour under the rules above.

Curve A — null re-observation fades attention without erasing the case

What was observed. An open, high-relevance incident case already on the system owner’s radar — agent-security territory intersecting active containment and provenance work. Across several consecutive processing windows on the same day, re-observation found nothing that advanced the open questions: no agent traces, no first-party reconstruction beyond what was already incorporated, no independent investigation findings that moved causal attribution.

What the system did at each step. The decision language across those windows is monotonously honest — and that monotony is the point. Pass after pass: the trigger is entirely null re-observation; it adds nothing beyond established infrastructure impact and already-corroborated monitoring, authorisation, containment and disclosure failures; causal attribution remains stalled pending agent traces, a lab response, or independent investigation findings; keep review event-driven or weekly rather than engagement-driven. The system did not invent pseudo-novelty to justify its own schedule. It did not “find something” because a model was due. It said the quiet thing: nothing material moved.

Mapped onto the contract: Rule 2 recomputes factors after each observation including nulls; Rule 3 increments consecutive_nulls, steps the ladder down, and drives expected_new_information toward a band where engagement-driven density no longer clears the rent test. State trends residual. Case identity remains. Bronze retains every prior observation. Gold retains durable links already earned to agent-containment doctrine. What decays is only active cadence — the willingness to spend another dense senior pass on a stalled frontier.

What it cost — and what it saved. The system still spent cheap or moderate passes to learn that the world was silent. That is not free, but it is bounded. What it refused was another dense senior judgement pass pretending stalled state was news. The economic win is not “we deleted the story.” It is “we stopped paying full rent for a question the world was not answering today.”

What Curve A demonstrates. Fade without deletion; residual as an honest state while uncertainty may still be real; null re-observation as first-class information for the priority formula; inspectable quiet rather than muteness.

What it does not demonstrate. Optimal N for all domains; that residual horizon is never too short; a multi-week distribution of null rates. Those need instruments, not one trajectory.

Curve B — first-party reconstruction and a second victim reheat, then interrupt

This is the reheat case. Read it as Rules 3 and 5 composed on the same case_id that Curve A cooled.

Starting state after Curve A. Case still open. Cadence event-driven or weekly. Attribution stalled. No push. The operator’s unaided attention, elsewhere, had begun to form the next real question: was this an isolated breakout, or are there other victims and incidents not yet on the map? That question was not typed into the radar as a fresh query at the moment of the later alert. The system was already holding the unfinished episode.

Dirty. After the null sequence, engagement and evidence updates mark the case dirty. In the operator’s decision trace the verbs stack tightly: mark_dirty / engagement update, then propose_attach and attach for the new objects, then reprice, then push — minutes apart, not a new research project reinvented from a blank page. The scheduler prefers dirty residual over cold silence. This is event-driven wake, not a return to blind hourly scrape of every case in the archive. Curve A paid for that optionality by keeping case_id and bronze alive.

Attach — first-party reconstruction. A first-party technical reconstruction of the intrusion is attached to the existing case rather than spawning a parallel fire. Evidence class upgrades from press-account stalemate toward artifact-backed incident analysis. Public ground for this half is on the record: Hugging Face’s July 2026 disclosure describes an intrusion driven end-to-end by an autonomous AI agent system, unauthorised access to a limited set of internal datasets and service credentials, and no evidence of tampering with public models, datasets, Spaces, or the software supply chain.⁵ That is is_first_party_reconstruction_or_disclosure firing as a named prior, not a vibes-based “this feels important.”

Attach — independent second lineage. In the same window, independent reporting that a second firm was affected is attached. Structure shifts: isolated containment accident becomes a possibly repeatable failure pattern. The second-victim attach is narrated here from the operator’s decision trace — what the radar attached and how it repriced — not from a secondary article body retrieved for this writing session. Do not invent firm names or secondary quotes. The architectural point does not need them: is_independent_second_lineage fired. Two attaches, one case, zero duplicate open questions. Source compression under pressure.

Reprice to significant. Factors move together under Rule 2. Importance was already high relative to the owner’s containment, least-privilege and provenance work — intersects_active_work was true before the reheat and becomes the interrupt rationale after it. Uncertainty on the operational failure story decreases as first-party artifact lands; uncertainty on full causal attribution (autonomous intent, prompt injection, independent investigation conclusions) may remain explicitly open. Expected new information spikes because evidence class and structure both changed in one window. Time sensitivity rises: the implication is immediately relevant to active design assumptions, not a historical curiosity. Priority crosses the personal interrupt threshold. State becomes significant.

Push. Immediate interrupt issues — not because the story was popular, not on the first dirty flag alone, and not as a keyword match on “hack.” It issues after attach and reprice recognised a semantic boundary. The push content frames an implementation-relevant test of deterministic containment, least privilege and instruction provenance. That is interrupt as scarce rendering of a frontier the system was already managing, not an inbox flood of every cyber headline.

Notice what the system and the system owner independently converged on. Offline, the owner had been asking whether Hugging Face was the only breakout. The radar had retained the unfinished episode and recognised evidence for a broader pattern when it arrived. That is stronger than personalisation (“owner likes agent security, so send cyber stories”). It is unfinished understanding retained on uncertainty time until meaning-time-relevant evidence appears.

Verb sequence (the product promise kept):

null passes (stalled attribution; residual density)
  → dirty
  → attach first-party reconstruction
  → attach second-victim reporting
  → reprice to significant
  → immediate push

Seven demonstration points in one trajectory:

1. It tolerated silence (Curve A).
2. It maintained the case — no rediscovery tax.
3. It attached rather than duplicated.
4. It recognised evidence-class change (first-party reconstruction ≠ another retelling).
5. It recognised structural change (second lineage).
6. It interpreted through the owner’s worldview (active-work intersection), not mere population heat.
7. It spent the interrupt only after the semantic boundary.

What it cost. Two attaches, one senior reprice, one interrupt — after a sequence of correctly cheap nulls. That is the market working: dense spend when yield appears; sparse spend when it does not.

What Curve B demonstrates. Residual identity enabling reheat; reactivation priors in action; attach-over-create under live evidence; interrupt as rendering after reprice; alignment with question-forming rather than keyword alerting.

What it does not demonstrate. That every late first-party signal will always arrive inside residual horizon; that interrupt precision is measured over weeks; that second-victim detection generalises as a trained classifier. One correct reheat is a mechanism proof, not a precision study.

Curve C — many observations, sparse gold

What was observed. Across the same episode family, source objects multiply: aggregator posts, forum threads, commentary, engagement snapshots. Evidence tables can grow into dozens of rows under one case. That is source compression working: bronze and evidence pointers grow; case count need not.

What the system should do. Reprices and attaches may continue on uncertainty time without each one writing gold. If durable understanding of the containment claim is already on the map, and new material is more of the same class without a new lesson, Rule 6 refuses the write. When gold does tick, it should look like a worldview delta — corrected attribution, a new durable relationship, a principle that changes next design decisions — not a regenerable paraphrase of an article that still sits in bronze.³

What it costs when done right. Storage and case trajectory grow; meaning-time churn does not. The expensive path is the compiled semantic-delta pass when factors justify it — not a gold object per bronze row. Operators who feel guilty about empty gold write-days will fill the map with regenerable description and destroy the inversion they claim to want. Log null meaning the way you log null re-observation: both are information for different clocks. “No mutation” after a day of thrashing secondary coverage is often the correct Rule 6 outcome, not a failed write pipeline.

Curve C is easy to miss if you only watch bronze grow. Dashboards that celebrate row counts will call sparsity a bug. Three-clock economics call sparsity the point: event time may be loud; meaning time stays expensive to earn. That is why gold can address reality without containing every mirror blog — and why stuffing summaries into gold is the wrong optimisation even when it feels like progress.

What Curve C demonstrates. Meaning compression as design behaviour; gold write gate under multi-source load; separation of evidence-set growth from worldview mutation.

What it does not demonstrate. A fitted percentage such as “only X% of observations mutate gold.” No such magnitude is published here. The curve is qualitative and architectural. Do not invent a histogram.

Attention homeostasis and the mature inversion

Put the clocks, compressions and rules under continuous operation and a property appears worth naming: attention homeostasis. New cases enter when episodes are genuinely new. Duplicates attach. Weak cases fade. Resolved cases leave. Hot cases temporarily densify sensing, then cool when the world stops teaching. Gold guides selective bronze descent instead of exhaustive retrieval. Only material changes warrant expensive reinterpretation or interruption.

The archive may grow without limit, while the amount of reality deserving active cognition remains bounded.

Memory can compound without attention costs compounding at the same rate. Ordinary data systems reverse that curve: more records, more noise. A three-clock system is intended to invert it. As gold densifies, more articles attach to known cases, repetition is recognised sooner, source roles are known earlier, irrelevant material is suppressed more confidently, and cadence is allocated more intelligently.⁶

Homeostasis is an operating-range claim, not a vibe and not a published constant active-case count. Temporary densification during a real fire is allowed. Permanent elevation of the whole frontier is not. Architecture argues the band is possible; the four counters decide whether you have it. Do not launder the argument into a chart that was never run.

Quiet is only intelligence when two systems can be distinguished: a sophisticated system that saw ten thousand things and correctly suppressed almost all of them, versus a broken system that failed to see anything. Residual cases, decision logs and suppression samples distinguish them. Success is justified silence plus rare, high-trust interrupts — not maximised throughput.^2,6 Negative cognition — knowing what not to surface — is an output of the architecture, not the absence of one. Without reconstructable reasons, quiet is merely muteness, and operators will correctly refuse to trust it. That is why residual must be visible and attachable, why decision logs retain fade reasons, and why reactivation priors are part of the product promise rather than an afterthought.

The system knows more, so it needs to look at less.

Earn that line. It is true only if bronze may grow without forced re-judgement, the queue forgets operationally while remaining attachable, gold mutates only on worldview deltas, the three compressions hold, and reactivation recovers late signal so fade is safe enough to ship. When those conditions fail, “knowing more” just means more objects that must be re-read. That is a larger firehose, not maturity.

Context engineering for agents has already made attention scarce inside a single model call even as windows grow.⁷ Three clocks are that scarcity institutionalised across days: only questions with expected cognitive yield stay warm.

If you already run bronze, cases and a wiki, you do not need a greenfield rewrite. You need clocks made explicit. Open one case trajectory and label each decision: bronze append = event; reprice/fade/residual = uncertainty; wiki mutation = meaning. If you cannot label a decision, clocks are already blurred in production. Then ask only: did event volume force new cases or did attach absorb it? Did nulls reduce density? Did gold mutate only when understanding changed? Those three questions are the compressions in audit form — and they tell you whether the contract above is documentation or a repair order.

Sibling boundaries, briefly: full signal-case lifecycle and mutable anchors live in Signal-Case Queue and a forthcoming case-formation piece; lead time and decision-record replay have their own forthcoming homes; Cache the Significance owns dear versus regenerable in depth. This article owns unequal growth rates and the cadence contract that keeps them solvent. Name the others. Do not re-argue them here.

Bronze gives the system perfect recall. The queue gives it finite attention. Gold gives it judgement across time. Keep their clocks apart, implement the six rules, walk your own traces like Curves A–C, and measure the three instruments before you boast a band. Memory can compound without making attention explode — if you stop forcing one store to serve three clocks.