Skip to content
finalstate.aiTacit Knowledge: We Can Know More Than We Can Tell
123Atlas

123THE HEAD THAT WALKS OUT

Tacit Knowledge: We Can Know More Than We Can Tell

Tacit knowing integrates cues into recognition and action without an exhaustive account of how they combine. A record can support learning; a stand-in still needs to demonstrate performance.

The manual remains in the workshop as a technician moves toward the door with a schematic network representing practiced judgment.The stationary manual and moving operator distinguish retained procedures from a person's practiced integration of cues. The network is a metaphor, not a transferable knowledge file.ManualJudgment
Illustrative workshop. The manual remains; the moving cue network represents practiced judgment.

THE CUE BEHIND THE ACT

A cue matters through what it helps you notice

A rattle trace and a load indicator connect to the same machine, then a frame marks the diagnosis receiving attention.This illustrative construction represents attending through cues to an integrated judgment. Neither cue alone is labeled sufficient for diagnosis.RattleLoadDiagnosis

In Polanyi's account of awareness, subsidiary does not mean unconscious. For the workshop technician, the rattle matters together with the load and the machine's history. That dependence on one person can remain a latent exposure after the manual is saved.

  • Focal attention rests on the diagnosis or act
  • Subsidiary cues contribute to that attended whole

RECOGNIZING THE WHOLE

Recognizing a face is easier than specifying every clue

A schematic face remains whole beside separated eye, nose and mouth features.A still interpretation of Polanyi's face-recognition example, not a portrait of Polanyi or an experiment showing that recognition cannot be taught.WholeFeatures
Schematic face, not an author portrait. Polanyi, The Structure of Tacit-Knowing, 1964, pp. 11–15.

In his 1964 Duke lecture, Polanyi discusses face recognition and learning through advice, example, and trial and error (pp. 11–15). He describes attending from particulars to a whole. The argument allows teaching and analysis without promising exhaustive articulation.

MORE THAN WE CAN TELL

We can know more than we can tell.

Michael Polanyi, The Structure of Tacit-Knowing, 1964, p. 12

The word can leaves room for articulation. An interview may reveal a useful cue without becoming a complete account of the skill. The same clause appears in the publisher's description of The Tacit Dimension.

THE REUSE BURDEN

A record made for its author may need work for another reader

A record returns to its author along one path and reaches another reader along a second path only after a contextual insert joins it.The two illustrative routes show different reuse needs, not a measured cost curve or a guarantee that adding context makes reuse successful.AuthorReuserRecord

Markus's theory of knowledge reuse distinguishes four reuse situations (pp. 63–72). Producer, intermediary, and consumer roles can overlap. Time and incentives matter, alongside access, interpretation, and support; author effort alone does not explain success (pp. 82–83).

  • Producers may reuse their own records or exchange them with peers
  • A distant reuser may need context, explanation, and help applying the record

A saved incident is evidence for a handoff

Capture can reduce exposure when someone can use what was retained. The next step is to find out what the account leaves unclear and observe its use in a bounded rehearsal.

PROBE A REAL INCIDENT

Revisit the moment the routine stopped fitting

An incident timeline links a change in the rattle to a stopping decision; a recalled load cue is added to an incident sheet with an unresolved gap retained.The interview recovers a recollection of one incident. A dashed gap remains to distinguish the account from a complete or independently verified event record.RattleStopRecollection

Gary Klein describes the Critical Decision Method as a retrospective interview using probes around an actual nonroutine event. Applied here: reconstruct when the rattle changed, what the technician noticed, and why stopping seemed warranted. Compare the recollection with the incident log; retain disagreements.

  • Choose a nonroutine incident the expert handled
  • Separate recalled cues from corroborating records
  • Keep uncertainties visible

COLUMN COVERAGE IN SIMULATION

820 of 864 descriptions mentioned every source-table column

A fixed bar shows 820 of 864 synthetic runs meeting the all-columns-mentioned criterion; 44 runs do not meet it.The measured count and proportion remain fixed in every state. The denominator is simulation runs, not pieces of tacit expertise. This criterion is separate from the paper's text-quality measures.All columns mentioned820 / 86494.9%Synthetic runs
1820/864 synthetic runs (94.9%) mentioned every source-table column. This measures coverage, not correctness or tacit skill transfer.Zuin et al., arXiv:2507.03811v1, July 4, 2025, §§IV–V.

Zuin and colleagues' arXiv version 1, sections III–V, describes the reconstruction of generated table documentation. Mentioning every column does not establish that each description is correct. The study did not test live employees or the transfer of a practiced skill.

  • Synthetic company networks; GPT-4o mini questioned simulated employees
  • 288 parameter combinations, three repetitions each
  • The 94.9% criterion measured column coverage

WHAT THE ARCHIVE CAN HOLD

An answer can leave the archive. Performance happens outside it.

An incident record moves from a bounded archive toward a stand-in beside a training machine; a source copy remains inside.The retrieved document crosses the archive boundary. The machine and stand-in remain outside, where the applicability of the account and the person's performance must be observed.ArchiveRehearsal

That boundary matters when fixing the wiki's clock. A timely answer can support a rehearsal; it cannot supply the performance evidence itself. Preserve opportunities to practice rather than letting deskilling weaken the person using the archive.

  • Keep the incident, context, and source trail together
  • Check whether a retrieved account fits the present case
  • Observe the stand-in using it

WHILE THE PERSON IS STILL HERE

Let a rehearsal expose the missing cue

Two illustrative training trials show a missed cue recorded after the first attempt, followed by the stand-in noticing the cue and stopping on a later attempt without expert prompting.The first miss remains visible beside its incident note. A return route connects that note to the later trial. The observed stop concerns one controlled case and does not certify general competence.First trialMissed cueRepeatStop
Illustrative rehearsal on a controlled training rig. One later stop without expert prompting is evidence about this case only.

In the illustrated rehearsal, the stand-in first misses the rattle's change under load. After feedback and a record correction, a later trial ends in a stop for inspection. That is evidence about this case, not comprehensive transfer. Keep untested cases open, then continue to When a Company Wiki Becomes a Graveyard, Fix the Clock.

  • Set one exception and a stopping criterion on a controlled training rig
  • Have the stand-in use the archive while the expert observes
  • Record missed cues, revise the account, and repeat without expert prompting
Read the transcript

01 · THE HEAD THAT WALKS OUT

Imagine a workshop where the manual stays and the technician leaves. The manual records the expected steps. The technician carries practiced judgment toward the door. Tomorrow, a new rattle may require more than those steps. In this example, the technician would attend to how the sound changes under load before deciding to stop for inspection. Saving the procedure preserves something useful. It does not establish that another person can notice the same exception.

02 · THE CUE BEHIND THE ACT

A rattle is a cue. The load is another. Together with the machine's history, they bear on the diagnosis. Polanyi distinguished subsidiary awareness of the clues from focal attention to the whole. The cues contribute together; attention rests on the diagnosis. The person may explain some of the clues while leaving part of their integration unstated. Such integration can also be mistaken.

03 · RECOGNIZING THE WHOLE

Polanyi used a familiar face to make this structure easier to see. You recognize the person, yet may struggle to specify every clue you used. In his nineteen sixty-four Duke lecture, he also described learning through advice, example, and trial and error. The teacher can guide attention. The learner still has to develop the ability to recognize the relevant pattern.

04 · MORE THAN WE CAN TELL

We can know more than we can tell. Polanyi's word can matters. It leaves room for an expert to describe a clue, demonstrate a contrast, or correct a learner. In the workshop, asking about the rattle may recover an important observation. The resulting sentence is useful evidence. Whether the stand-in can use it remains a question for performance.

05 · THE REUSE BURDEN

A note that reminds its author what happened may leave another reader puzzled. Lynne Markus's account of reuse includes people reusing their own work, peers exchanging records, novices seeking expertise, and people mining records for new purposes. The note returns to its author. With added context, it can reach another reader. Packaging takes effort, but so do finding, interpreting, and applying the result. An intermediary can help; these roles can overlap.

06 · CAPTURE IS NOT COMPLETE

An archive can preserve evidence without reproducing expertise. Capture can reduce exposure if someone can use what it retains. A useful incident account gives the stand-in something to question and try. A completed interview, by itself, does not show that the handoff worked.

07 · PROBE A REAL INCIDENT

Choose a nonroutine incident the technician actually handled. Gary Klein describes the Critical Decision Method as a retrospective interview using probes around such an event. In this workshop, start with the remembered rattle. Follow the account to the stopping decision. Ask when the load changed and record the recalled cue. Compare that account with the incident log, and leave any disagreement visible. These questions seek a more useful account of one decision; memory is still part of the evidence.

08 · Advertisement · 11x AI Growth Workers

Pipeline is not one task. It is a chain of small handoffs: find the buyer, research the account, respond fast, qualify cleanly, follow up on time. 11x gives sales, marketing, and RevOps teams digital workers for that background motion. The system keeps work moving, while people focus on the conversations that deserve them.

09 · COLUMN COVERAGE IN SIMULATION

A study by Zuin and colleagues tests a different problem. An agent questioned simulated employees to reconstruct descriptions of database tables. Across eight hundred sixty-four synthetic runs, eight hundred twenty descriptions mentioned every source-table column: ninety-four point nine percent. That was the full-knowledge recall criterion. It did not mean every description was correct. The researchers evaluated text quality separately. Both the questioning agent and simulated employees used GPT-four-o mini. No live technician or transfer of practiced skill was tested.

10 · WHAT THE ARCHIVE CAN HOLD

Keep the incident record in the archive with its context and source trail. For a new question, retrieve the relevant account. Move a copy across the archive boundary to the stand-in. The person and training machine remain outside. A timely answer can support a rehearsal, but it cannot perform that rehearsal. The stand-in still has to check that the account fits the present conditions, notice a mismatch, and exercise the agreed stopping rule.

11 · WHILE THE PERSON IS STILL HERE

Choose one exception on a controlled training rig. Agree in advance that a change in the rattle under load calls for a stop and inspection. In this illustrative first trial, the stand-in misses the cue. The expert records the gap after observing the attempt. Correct the account and provide feedback. On a later trial, the stand-in notices the change and stops without expert prompting. Keep both attempts in the record. The later result supports a handoff for this case; it leaves other exceptions untested.

01 / 11 · THE HEAD THAT WALKS OUT0:00 / 5:27