Logos52
wiki / Dimensions / Deep Processing / Best attempt Encoding

Best-attempt Encoding

concept updated 2026-08-14

Best-attempt Encoding

Part of Deep Processing

You can start practicing retrieval on material you only half understand, and the half-understanding is not something to fix first. The move is to build one organized picture of the topic — the best you can currently make — and then hold it to a single question: can this be tested? A picture that can be questioned enters practice immediately and gets corrected there, by the practice. So the drawing stage is where the real choice sits, between a box narrow enough for one fact and one question type, and a box wide enough to hold an application, a connection, and a why. That choice is best-attempt encoding.

Narrow or wider

Every item can be held two ways, and the real difference is processing depth. Narrow encoding protects one fact against one question type: the item comes back when it is asked the way it was learned. Fast, and brittle — no transfer to adjacent questions, applications, or variations. A slightly wider box costs more time up front applying, integrating, and working across contexts, and it protects the asked question and its neighbours on the same topic.

If the wider encoding cost the same time as the narrow one, every item would get the wider one. There would be nothing to decide. The extra time is the only reason a triage exists. The house feel-test for the two widths is written 1:1 against 1:1.2. That notation is not a hit rate and not a measured coverage number.

The wider box also gives an update somewhere to attach — a new regulation, a new finding, a new example. A narrow item cannot absorb the update. It stays another isolated card.

The two filters

Two cheap questions decide which items earn the wider box. The screen itself, in this system’s own working model, costs about twenty seconds an item. The expense is in the encoding that follows, never in the decision.

Conceptual relevance asks whether the item attaches to something already known to matter. A small change to smoking legislation is an isolated fact on its face, and it attaches to a large theme already known to matter, so it earns the wider box. A change to a speed limit looks attached to nothing, so it gets filed narrow.

Repeated testing relevance asks whether the item keeps surfacing across question types and contexts. The same speed-limit fact stays narrow until retrieval on city planning keeps reaching for road incidents; the fact then turns out to be a usable example across several answers, and the second filter has found the relevance the first filter could not see. The two filters are one story.

When both filters are low, the item is carded and the work moves on. When either is high, the wider encoding is paid for. Flashcards are a tool class, not a product, and a card can still be made afterwards as a reminder — the two are not alternatives. The wider encoding is a set of acts: the fact used inside a written answer, placed into an existing map as a worked example, looked at for what else it connects to.

Testing widens the margin of error on every triage call, in both directions. An item filed as isolated turns out to be connected when retrieval keeps reaching for it; an item encoded wide turns out not to have taken, and the gap shows. The decision does not have to be right, because testing corrects it. The payoff of the wider encoding shows up in the review queue rather than in the encoding session: the card comes up right more often, spaces out, and stops arriving daily. Narrow items keep arriving. Importance-Based Chunking is the same importance judgment at topic scale; the two filters are it applied to a single item.

What the attempt is for

A schema should help recall, not require its own memorization. The test is whether the only way to remember how the material is organized is to memorize the organization itself. Yes is a retention risk: the structure has become a second thing to hold rather than a way of holding the first.

Five moves that widen an item

A generated application forces the concept off the page. One real example is enough.

One connection is fragile; two or three are much harder to lose. The threshold is a house working number, not a measured finding.

The mechanism question — why does this work — is the move that most often pays. Knowing the mechanism means the fact can be rebuilt when memory fades.

A predicted variation asks what would change if one condition shifted.

The chunk boundary is checked last: whether the item belongs inside an existing chunk, or extends it.

These five are the standard for any first encoding. They usually run inside the refinement stage of the Bear Hunter System, the encoding loop that ends by restructuring a map rather than taking notes — and they do not require a completed loop to run. Rushing that restructuring accumulates shallow encoding debt that surfaces later as poor retention.

Calibrating to your own level

A good structure is specific enough to be unique to this topic, intuitive enough that remembering one element surfaces the others, and flexible enough to take an update. Before / during / after is the positive form: remembering one member surfaces the other two, so only a fragment of the structure has to be held. Once a domain is owned, its own trios do this too.

The two failures sit on either side. Too specific: groups named with terminology so particular that the group names themselves have to be learned one by one, and remembering one tells nothing about the next. Too generic: a home-design topic split into interior and exterior, then furniture against art and materials against color — and the next topic takes the same split, and the next. The structure now fits everything, so it locates nothing. The clinical template does this at scale: history, examination, investigation, management, applied to every disease in turn. Mechanism / presentation / treatment is the same shape wearing different clothes.

What reads as intuitive is a function of how much of the domain is already owned. The structure is calibrated to current knowledge rather than to what a specialist would find obvious. The target is one level above the current encoding baseline, not the ideal level. Marginal Gains is why one level above baseline beats the ideal target: small gains held long enough.

Encoding skill moves slowly. This system’s own working model puts the climb from a low baseline to the top at ten months of consistent effort and often years, while most learners are starting mid-scale and need to reach a good-enough level, which takes months. Study time and content volume do not wait for that climb, so aiming every session at the best possible structure buys quality per topic and pays for it in coverage.

A small retention gain compounds and lightens every later review. In the same working model, a learner retaining around 40% of what they encode who improves the method to around 50% has not saved ten percent once. Each item that would have needed roughly ten repetitions now needs fewer, on every item, every week, for as long as the material is being carried — which is why a small encoding gain shows up as hours a month rather than minutes a session. Those figures are the system’s illustration, never a finding.

Where encoding skill is genuinely low, retrieval outranks encoding. The compensation is retrieval volume, not a better first map. Explaining the topic from memory to someone who knows nothing about it forces synthesis whether or not the learner can build structure yet, and failing at it is the gap being found. Encoding becomes the higher-leverage half later, once it determines how much retrieval is needed.

A daily card load past about an hour to an hour and a half that still falls behind is the triage skewed narrow: items are accumulating faster than retention can absorb them. The move is to spend more time integrating the items that keep arriving, not to grind the queue harder. That threshold is the model’s own caution, not a measurement.

Revision is the design

The structure was never going to arrive finished from the source. The prior reading — even if structurally wrong — is what makes the error visible during testing. Testing is what converts a reading into something reliable. Spaced Interleaved Retrieval is the retrieval half of that conversion.

Fixing a gap after retrieval costs a fraction of the original build only when the gap is found early. A structural error found after a large map has been built forces every later branch to move; the same error met in the first half hour would have changed how the map was built. The cost of a revision is a function of how much has been built on top of it, which is what makes the first retrieval pass worth scheduling early.

The cycle, named and handed off: best-attempt encoding, then high-volume retrieval within one week, then gap diagnosis by type, then re-encode the higher-order gaps, targeted retrieval for the lower-order ones, and cards for what stays isolated. How to prepare for ultra high-volume exams owns the schedule, the retrieval ordering, and coverage strategy under exam volume.

Skilled and new learners find the same gaps. What separates them is timing and follow-through: the skilled one expects gaps, goes looking early, and rebuilds the structure quickly; the new one delays the test, puts off the rebuild, and then rushes it. The difference is not accuracy of the first attempt.

Expecting revision is not a concession. It is how the cycle is designed to work. Effort spent rearranging a structure is what buys the better version of the knowledge, and there is no version of the process where that exchange is skipped.

Sources

  • Kapur, M. (2008, 2016). Productive failure: an initial structured attempt that is incomplete, then instruction, outperforms direct instruction on transfer.
  • Tulving, E., & Thomson, D. M. (1973). Encoding specificity: a cue matches the original encoding, so narrow encoding predicts poor transfer.
  • Barnett, S. M., & Ceci, S. J. (2002). Near transfer is common; far transfer is not automatic.
  • Craik, F. I. M., & Lockhart, R. S. (1972). Levels of processing. Craik & Tulving (1975). Elaborative encoding.
  • Pressley, M., et al. (1987). Elaborative interrogation. Chi, M. T. H., et al. (1989). Self-explanation. Dunlosky, J., et al. (2013). Moderate utility for both.
  • Gentner, D., Loewenstein, J., & Thompson, L. (2003). Analogical encoding.
  • Roediger, H. L., & Karpicke, J. D. (2006). The testing effect. Rowland, C. A. (2014). Psychological Bulletin meta-analysis.
  • Kalyuga, S., Ayres, P., Chandler, P., & Sweller, J. (2003). Expertise reversal: structures that help novices can hurt experts, and vice versa.