Gnosis, Belief and Paradigm Shift

A feedback protocol is a piece of instrumentation, and instrumentation has a property that beliefs do not: it can be wrong in a specific way, at a specific time, and the error can be located. This account covers six months of running one. The protocol itself is simple enough to state in a paragraph. Write down a sentence you can be wrong about. Open a record before the aim exists. Run the aim for a fortnight without adding anything to the record. Read the record against the aim. Update, which is the step everybody skips. What the six months produced was not evidence that the protocol works. It was three specific failures, each of them in the same place, and a clearer sense of which part of the whole enterprise can fail and which part cannot.

The protocol is an instrument, not a belief

The first thing to be clear about is what kind of thing this is. It is not a claim that belief changes outcomes. It is a procedure for making your own commitments falsifiable, and the difference matters because instruments can be evaluated while beliefs can only be defended. A thermometer tells you it is cold. A horoscope tells you it is going to be cold. Both are statements about the future, and only one of them can be checked against something it did not produce.

That is the whole reason for keeping the procedure separate from the attitude. The parabola, that magic is a device with no reality of its own that serves whatever end it is applied to, is a useful licence and it is frequently smuggled into places it does not belong. A licence belongs in the decision about what to attempt. It does not belong in the evaluation of whether the attempt worked, and moving it there is the single most common way that a disciplined practice becomes an undisciplined one.

Six months of running the loop

Six months is roughly three cycles of the protocol at a fortnight per aim, with a gap between cycles for reading. That is the longest span in which I would trust my own records, and the reason is specific: a month is too short to distinguish a change in practice from a change in circumstances, and a year is too long to keep the record detailed enough to be honest. Three cycles is also the minimum at which a pattern in the failures becomes visible rather than anecdotal.

What the six months contain, in summary: eleven aims, of which four were falsified inside their fortnight, two were falsified in the reading stage, and five survived. The five survivors are the ones I would distrust most, for the reason the next section sets out. The four early falsifications were all correct decisions and cost nothing. The useful information was almost entirely in those four.

Where the loop broke, three times

The loop broke in three distinct ways and each of them was a way of avoiding a downward update. The first was restating the aim. Twice, in the middle of a fortnight, I found that the aim as written was one I did not want to run, and I rewrote it in a form I did want to run, and then reported the fortnight against the new version. The record showed a completed fortnight. The fortnight had not happened.

The second was the record going quiet. In the third cycle I stopped writing entries on the days the practice went badly, which is the single most predictable failure of any self-monitoring scheme and the one that makes the surviving data worthless. The third was the reading stage, where the aim and the record matched closely enough to produce a feeling of satisfaction and no update was required, because an update only becomes necessary when the two disagree and the disagreement had been quietly removed upstream by the previous two failures.

The aim you restated

Restating an aim mid-cycle looks like flexibility and behaves like tampering. The mechanism is entirely ordinary: a fortnight is long enough for a commitment to stop feeling like a commitment, and rewriting the sentence restores the feeling without restoring the behaviour. The restated aim is almost always a weaker aim, because the rewriting happens at the moment of maximum resistance and the version written under maximum resistance is the version that asks least.

The fix is procedural rather than moral. Write the aim on a card, put a date on the card, and forbid yourself to change anything on the card until the reading stage. If the aim turns out to have been wrong, that is a finding, and findings are cheaper than the appearance of a completed cycle. Two of my eleven aims were restated and both of them survived, which is the part of the outcome I find least reassuring and most instructive.

The record that stopped being written

The record is the only component of this that can fail, and it fails in a direction that is always the same. Entries get shorter on bad days. The reasoning behind them gets vaguer. A day with a strong result gets three sentences and a day with nothing gets the word nothing, which is accurate and carries no information about the difference between the two days.

The fix is to constrain the record's shape rather than its quality. Three lines, every day, whether or not there is anything to report: what the practice was, what happened, and what I would have done in any case. That third line is the one that matters and the one everybody drops, because it requires a counterfactual and counterfactuals are uncomfortable. It is also the only line that distinguishes a day when the practice did nothing from a day when the practice did something with no visible result.

The downward update, and the week it arrived

The update is a Bayesian move and it is the least popular part of any honest practice. After reading the record against the aim, you adjust downward the estimate you hold of that method, that timing, or that framing, in proportion to how badly the record and the aim failed to match. In plain terms: you reduce what you expected, or you change what you did, and the change is in the direction the evidence points rather than the direction you prefer.

The week it arrived for me was the sixth week, on an aim about attention that I had been running for three cycles. The record showed that the practice happened on eleven of fourteen days and produced a change I could not distinguish from a good week. The update was to drop the aim entirely rather than run it again with a tweak, because three cycles of no distinguishable result is enough to conclude that this particular framing does not do this particular thing, and continuing was the motivated option.

Expectancy, and the ceiling

Expectancy has a measured influence on some outcomes, which is not the same as having a measured influence on all of them. It has real effects on reported pain and on a handful of physiological measures. It does not have a demonstrated ability to cure disease, and any account that slides from the first to the second is not being careful about the difference between an effect on how something is experienced and an effect on what is the case.

The practical consequence for a protocol is a ceiling that is lower than practitioners usually assume. Expectancy can make a practice feel like it is working, and it can do that so reliably that the feeling becomes the main evidence. A record with a fixed shape and a required entry for bad days is partly a defence against this, not because the expectancy is imaginary but because it is powerful enough to substitute for a record if the record is optional.

A protocol with a version number

The protocol gained a version number in the third month, and the version number is the only structural change I would recommend. The number goes on the card, next to the aim and the date, and it increments when the protocol changes rather than when the aim changes. That separation is what makes the history readable: you can see that cycle two used a different procedure from cycle one, and you can therefore stop comparing their results.

Without version numbers the temptation is to treat a change in method as a change in luck, or a change in luck as a change in method, and there is no way to tell which happened. With them, an improvement after a procedure change is a hypothesis rather than a fact, and it gets tested the next time the procedure is the same. This is a small piece of bookkeeping and it did more for the reliability of the six months than anything else in the protocol.

What the loop cannot detect

The loop is blind to anything that does not appear in a fortnightly record. It cannot detect a slow accumulation that only shows up over years, because two weeks is the wrong window for it. It cannot detect the case where the aim was right and the timing was wrong, because timing is not in the record unless you put it there. And it cannot detect the case that matters most for people who start these things, which is whether the practice is doing anything at all, because a protocol that only measures compliance will happily report perfect compliance for a method that does nothing.

This limitation is not fixable within the protocol. It is managed by what you put in the record, and the items worth adding are the ones that are cheap and easy to falsify: the time of day, the state of sleep the night before, whether anything else was changed in the same fortnight. A protocol that records those can at least generate the question of whether the aim is wrong, the timing is wrong, or the thing is not working. A protocol that records only compliance cannot generate that question at all.

Running it with an aim you expect to fail

The version of the protocol I trust most is a deliberately rigged one. Choose an aim you expect to fail, run it for a fortnight with the full record, and read it. The reason is that a rigged aim tests the machinery. If the record is complete, if the entry for bad days is there, if the reading produces a downward update, then the protocol has demonstrated that it can catch a failure you planted. That demonstration is worth more than any amount of confidence generated by an aim you secretly expected to work.

My rigged aim was a timing claim: that a particular practice performed at a particular hour would produce a particular effect. I expected it to fail and it did, and the record was good enough that the failure was informative rather than merely disappointing. The unrigged aims that followed were easier to run because I had already seen the protocol do its job on something I knew was false.

The month nothing happened

One month produced nothing at all: no result, no insight, no change in the practice, and a record that was technically complete and evidentially empty. The temptation in a month like that is to treat it as a pause, and the pause is fine, but the mistake is in not distinguishing it from a result. A month of nothing is information about the aim, and treating it as a gap in the record is how a record becomes a record of good days.

The other temptation in a month like that is to add something. A new timing, a new glyph, a new material, all introduced in the third week because the first two weeks produced nothing. Adding in week three guarantees that the cycle tells you nothing, because there is no longer a fixed procedure to compare against. I added. The cycle was lost, and the version number incrementing is the only reason I know which cycle it was.

What the feedback protocol is for

It is for protecting a commitment from the parts of you that would rather reinterpret it than abandon it. That is a smaller job than it sounds, because those parts are not malicious, they are efficient, and they will find a way to preserve an aim that is not working if you leave the ambiguity available. The protocol closes the ambiguity by fixing the aim before the record opens and the reading after both are done.

It is not for producing a belief that the practice works. Six months of running it gave me four clean falsifications, one abandoned aim, one rigged failure, and a clearer sense that the parts of the enterprise most worth doing are the parts with observable outputs. Whether the enterprise as a whole is doing anything is not a question this instrument is built to answer, and any account that presents it as answering that question has misdescribed it.

The version I would write now

The version I would write now adds one line and removes two. The added line is a control: run a fortnight where the practice is deliberately withheld, and record it in the same format. A control fortnight is the cheapest available check on expectancy and it converts the whole fortnight from a single reading into a comparison. The removed lines are the ones about how the practice felt, which I now record nowhere, because a feeling logged daily becomes the most persuasive document in the file and the least diagnostic.

The rest of the protocol would stay as it is: aim on a card, record open before the aim exists, three lines a day including the counterfactual, no changes to the card until the reading, and a version number that increments on procedure and not on aim. It is a small amount of discipline for an uncertain return, and the return has been entirely in the four falsifications rather than in the five survivals.

Frequently Asked Questions

What does a belief feedback protocol actually consist of?

Five steps. Write down a sentence you can be wrong about. Open a record before the aim exists, so the record is not shaped by the aim. Run the aim for a fortnight without adding anything to the record. Read the record against the aim. Then update your estimate downward in proportion to how badly they matched. The update is the step most people skip and the only one that can change anything.

How long should a protocol be run before its results mean anything?

About six months, which is three fortnightly cycles. One month is too short to separate a change in the practice from a change in the circumstances, and a year is too long to keep the record detailed enough to remain honest. Three cycles is also the minimum at which a pattern in the failures becomes a pattern rather than an anecdote.

Why is it a problem to restate an aim mid-cycle?

Because the restatement almost always weakens it, since it happens at the moment of maximum resistance and the version written under maximum resistance asks the least. The record then shows a completed fortnight that never happened. The fix is procedural: write the aim on a dated card and forbid changes until the reading stage, so an aim that turns out to be wrong registers as a finding.

What are the real limits of expectancy effects in this kind of practice?

Expectancy has measured influence over reported pain and some physiological measures, which is a real and limited effect. It has no demonstrated ability to cure disease. A record with a required entry for bad days is partly a defence against expectancy, not because the expectancy is imaginary but because it is strong enough to substitute for a record when the record is optional.

What does the protocol simply fail to detect?

Anything that does not fit inside a fortnightly record. It misses slow accumulation, it cannot tell a wrong aim from a wrong timing unless you record timing, and it cannot detect whether the practice is doing anything at all, because a protocol that measures only compliance will report perfect compliance for a method that does nothing. That limit has to be managed by what you choose to log.