Augmenting Human Intellect: A Conceptual Framework

The unit of design is larger than the interface

Engelbart examines capability as an interaction among a person, representational language, artifacts, methods, and training. His report develops a framework and a research proposal, not a controlled proof that every proposed tool improves reasoning. The bootstrapping idea asks whether improvements in the activity of improving tools can accelerate subsequent improvement. A feedback loop still needs evidence that its direction is beneficial.

Consider an original example: a small team maintains a knowledgebase about computer architectures. Each technical claim should connect to a source, an interpretation, and any assumptions. A plain document can contain all this text, but it may not preserve the relationships when one statement changes. A linked representation gives claims and sources stable identifiers and records which claims depend on which evidence.

Trace a change through a representation

Create source S1, measurement M1, claim C1, and conclusion C2. M1 records a latency as 12 milliseconds and cites S1. C1 compares M1 with a 20-millisecond threshold. C2 recommends a design because C1 says the threshold is satisfied. Later the team discovers that S1’s value was 12 seconds, not milliseconds.

Changing a number in one paragraph is insufficient. The dependency edges M1->C1->C2 identify affected reasoning. Mark M1 corrected, invalidate the derived comparison, and require C2 to be reviewed. Keep the prior version and the reason for correction so readers can reconstruct the earlier conclusion. The graph does not itself understand whether the inference was valid; it makes a relevant relationship explicit enough for tools and people to inspect.

Stable identifiers solve a different problem from visible titles. Renaming C1 should not break its incoming references. Deleting it should produce a detectable missing dependency or an explicit replacement, rather than silently changing the meaning of every link. These are data-model decisions with direct consequences for the research method.

Change methods and training with the artifact

A team will not automatically use evidence links simply because a toolbar offers them. Define a method: every quantitative claim records units, conditions, and source; every edit affecting those fields triggers a review; reviewers distinguish an observation from an inference. Train contributors on one worked correction, including the case where a source supports only a narrower claim.

Now the language, method, training, and artifact reinforce each other. Without the method, links may become decorative citations. Without training, contributors may confuse a convenient web page with primary evidence. Without an artifact that preserves relationships, disciplined contributors must repeatedly track dependencies by hand. Improving only one part can leave the overall workflow ineffective.

Evaluate and improve the improvement loop

Measure whether a fresh reviewer can trace a conclusion to its assumptions and correctly update it after a source correction. Count missed dependent claims, time spent finding provenance, and errors introduced during revision. Do not equate the number of links or documents with improved understanding. A highly connected graph can still encode unsupported reasoning.

Use the same system to record problems with its own workflow and compare revisions. That is a small, concrete bootstrapping loop. It should retain unsuccessful changes as evidence, rather than treating every iteration as progress. Access control, accessibility, and who can challenge a claim also affect the capability of the group.

The framework’s value is a reminder that tools reshape practices. Its limitation is that a broad account cannot substitute for task-specific testing. An excellent demonstration of linked editing does not establish that an organization will adopt the method or distribute its benefits fairly. Those outcomes require separate observation and design work.

from collections import deque
edges = {'S':['A','B'], 'A':['C'], 'B':['C'], 'C':['D'], 'D':['B']}
seen = {'S'}
queue = deque(['S'])
while queue:
    node = queue.popleft()
    for child in edges.get(node,[]):
        if child not in seen:
            seen.add(child)
            queue.append(child)
assert seen - {'S'} == {'A','B','C','D'}

• Augmenting Human Intellect: A Conceptual Framework: Doug Engelbart Institute