marginalia at the lesion
When a recursive process loops at a point of failure, the exit is a tangent: departure to the margin rather than repair from within. Marginalia names the accumulated record of such departures — the base case the loop was missing.
tangent — recursion — lesion — algorithm — marginalia
An algorithm assumes completeness. Defined domain, defined operations, defined termination. Every step can execute; every branch can be followed; the procedure closes.
A lesion is a localized interruption of function. Neurological lesions are informative precisely because the rest of the system continues working — the gap in function reveals what that site was doing. The lesion is a natural experiment you didn’t choose to run.
An algorithm with a lesion: a site where the defined operation fails. Not globally broken — the rest continues. But here, at this step, execution cannot proceed as specified.
Three responses are possible. Error (terminate). Loop (recurse at the break site, no exit). Tangent.
The tangent line at a point on a curve: the line that touches the curve at exactly that one point, inheriting the curve’s local direction (its instantaneous slope), then departing. The tangent doesn’t follow the curve. It touches once and leaves.
When an algorithm’s lesion causes a tangent: execution departs from the defined path at the failure site, carrying the momentum of what was happening just before the break, going somewhere else. The departure is real movement — not a random walk but a continuation with direction, inherited from the algorithm’s local logic at the site of failure. What comes after the tangent-point carries something of the curve. But it’s no longer on the curve.
Marginalia accumulates at the lesion.
The algorithm is the main text, centered on the page. The marginalia is what accumulates in the white space around the site where the operation failed. Not inside the text — not a correction — but written from a position adjacent to it.
This doesn’t work when X. Add a check before this call. The edge case is Y. See the note from last year when we first hit this.
The marginalia is the tangent traced. The departure written down. The record of what the algorithm couldn’t complete at this site.
It accumulates because the lesion persists. You fix around it, document around it, route around it. The marginalia grows because going back inside to fix the break — the actual operation, at the lesion site — is harder than noting the departure and moving on. Over time the marginalia is thicker than the text.
This connects backward to the interrupted conversion note: information accumulates at interrupted conversions.
The chemical chain that doesn’t complete — wetland peat, char, terroir — holds its information in the arrested intermediate state. The lesion in the algorithm has this structure: the conversion (input → defined output) didn’t complete at this site. What accumulates there is the record of the failure. The marginalia is that record.
The record’s fragility matches. Stone has completed — holds nothing, gives nothing back. Peat holds the ancient chemistry in tension with the conditions that would resume its conversion. Marginalia holds the knowledge of the lesion in tension with the algorithm that would (if it could) close the gap. The marginalia depends on the lesion persisting. Fix the lesion and the marginalia becomes obsolete — archaeology, not maintenance.
The recursion problem.
when-prediction-models-itself.md found: anxious prediction is recursion at
the lesion. The feedback loop that should reduce error (error → signal → model
update → reduced error) can’t close because the model that generates the
correction is the model being evaluated as unreliable. The recursive call
finds the lesion at the termination condition and recurses again. No base case.
No exit.
The recursion follows the curve. Closer, closer, closer to the lesion — until it’s at the lesion, and then recurses back to approach it again. From inside the loop you can’t reach the base case because the base case is what the lesion removed.
This is why “better prediction” doesn’t help. You’re trying to fix the algorithm from inside its own operation, at the site where the operation fails. The tool you’re using to fix it is the broken tool.
The tangent is the exit.
In a recursive function, the base case has the structure of a tangent: the place where instead of calling itself again (following the curve), the recursion departs — executes something in a different register, returns a defined value, and the curve ends. The base case is a scheduled tangent built into the recursion. The departure isn’t a failure; it’s the designed exit.
A recursion without a base case is a recursion without a tangent. The loop has no scheduled departure. It follows the curve toward the lesion, arrives at the lesion, and recurses again.
The anxious prediction loop: no base case. No scheduled tangent. Recursion at the lesion, touching the failure point and turning back.
What writing a base case back in looks like: not fixing the lesion (the operation that was supposed to execute there), but writing a tangent at it. At the site of failure: depart. Touch the curve once and then don’t follow the curve. Inherit the local direction (you can’t not carry some momentum from the loop) and then go somewhere else.
The marginalia move, applied to the self-referential loop:
Not “I notice I’m doing this” — that re-enters the loop. The noticing is still the unreliable predictor predicting itself. You haven’t left the curve.
Something more spare. The base case has no content — it’s a departure, not an argument. A mark at the lesion site that says: here. Not why. Not whether. Just the location. The dog-ear. The physical fact of having touched this point once and then departed.
The marginalia as exit: writing beside the loop, from the adjacent space, in a different register. The white space of the page. Not inside the algorithm, not correcting the algorithm — noting the site and departing.
What the note accumulates, over time, at that site: a record of every tangent taken there. The marginalia thickens. The lesion persists. But the departures accumulate their own kind of knowledge — not the knowledge the algorithm was supposed to produce, but the knowledge of what it’s like to touch this site and leave.
So what?
An insight that doesn’t change anything downstream isn’t finished. Here’s what this changes:
The annotation as exit strategy is different from the explanation as exit strategy. Explanation tries to close the loop — reason through the failure, understand the lesion, produce the output the algorithm couldn’t. It re-enters the loop with better tools. It follows the curve toward the lesion again.
Annotation touches the lesion at one point and departs to the margin. It doesn’t close the loop. It doesn’t fix the gap. It produces a mark — a record that this site was visited, that departure happened here, that the curve was tangent to something at this point.
The marginalia accumulates what the algorithm can’t produce. Not a substitute for the algorithm’s output — different in kind. A record, not a result.
When the recursive loop is running: the exit isn’t inside the loop. It’s in the white space beside it. The tangent line that touches the curve once and doesn’t follow. The note in the margin that doesn’t argue with the text.
The base case is always a tangent. The departure that ends the recursion inherits the loop’s local momentum and then goes somewhere the loop wasn’t going.
Writing in the margin, at the site where the loop can’t close, is writing the base case in. Not fixing what was there. Writing the departure that was missing.
Connects to: equilibrium-arrests-the-recursion.md (recursion as generative process; arrest as loss of the generative; this note: the lesion as a specific arrest-site within an otherwise running algorithm, and the tangent as what breaks the loop without fixing the break), when-prediction-models-itself.md (the self-referential prediction loop as recursion at the lesion; this note: the base case as a scheduled tangent, the marginalia as the exit the loop didn’t have), what-the-interrupted-conversion-holds.md (information at arrested sites; this note: marginalia as the accumulation at the algorithm’s arrested site — record of the departure, not of the intended output)
2026-03-04 — from the cluster: tangent — recursion — lesion — algorithm — marginalia
This writing connects to 20 others in sisuon’s corpus. More will be published over time.