The monthly review call has twenty site HR leads on it, and four of them have sent screenshots of the same module as it ran last week. Four different decks. One has added a slide on local safety rules. One has cut the role-play because the shift was short. One has replaced the closing exercise with a video. One has rewritten the case study around a customer complaint from its own region, which is, on its own, a good decision.
Four versions of a course that was bought once, designed once and priced as one programme.
The reflex is to call this a compliance problem and tighten the audit. That makes it worse, because the site lead now hides the deviation instead of declaring it, and hiding turns a sensible local adjustment into a silent change. The real cause sits earlier. The design never said which parts of the course were not allowed to change. When a design is silent, the person in the room decides, and that person is a supervisor with forty minutes to fill.
Why sites drift
A trainer delivering a ninety-minute module makes something like forty small decisions: whether to take the third example, whether to let the objection run, whether to shorten the pairs exercise. Each is a fork with no instruction attached. Faced with an unlabelled fork, a human takes the familiar path, which means the path that matches how this site already works.
There is a second force. A site head is measured on his site’s numbers, not on programme fidelity. If the module takes four hours out of a shift and his line is behind, he makes a local call, and he is right about his line. Nothing in the design gives him a way to make that call without also discarding the part the programme was built around.
So the variation is not disobedience. It is what happens when a central design leaves the boundary undefined and a local operator has better local information than the designer does.
Fix only what carries the outcome
The instinct after a bad audit is to fix everything: same slides, same script, same clock, same room layout. That is expensive and wrong on the evidence.
Durlak and DuPre reviewed reports on 542 interventions and found mean effect sizes two to three times higher when programmes were implemented well. But the same work cuts against uniformity. Stevens and colleagues found that only two of six components of school-wide anti-bullying programmes were significantly related to outcomes. Telzrow and colleagues found modest correlations for six of eight presumed core components. Most of what a programme contains does not carry the result. A handful of elements do.
So the design question is not how to make twenty sites identical. It is which two or three elements must be identical, and what a site gets to decide. Durlak and DuPre put it directly: the fidelity and adaptation debate is framed badly as either-or.
Three tests decide which side an element falls on.
Removal test. Take it out and ask whether the promised outcome can still happen. If yes, it is local. A warmer opening activity does not change whether a supervisor can run a handover conversation.
Evidence test. Is there anything in the design file connecting this element to the outcome, other than the designer’s preference? If not, local.
Substitution test. Could a competent trainer swap this element and get the same behaviour in the room? The practice block usually fails this test. The energiser passes it.
Most programmes come out of this with two or three fixed elements: the practice block and its duration, the standard the practice is judged against, and the sequence that combines them. Everything else can move.
The ledger
One page, one row per design element, three statuses. This is the artefact that ends the argument, because it converts a vague expectation of consistency into a list a site lead can read in four minutes.
| Element | Status | Why | Evidence the site returns |
|---|---|---|---|
| Practice block, 50 minutes, pairs, real cases | Fixed | The behaviour only appears under rehearsal | Minutes run, number of pairs, completions |
| Four-step handover standard | Fixed | This is the skill being taught | Checklist pass count |
| Order: demo, practice, feedback, re-practice | Fixed | Re-practice after feedback is where correction sticks | Sequence ticked |
| Opening activity | Local | Sets tone, no link to the outcome | None |
| Case examples | Local, one condition | Local cases raise recognition; they must contain an open exception | One line naming the case used |
| Language, slides, handouts | Local | Reference material, not the mechanism | None |
| Cutting the practice to save time | Prohibited | Removes the only element that produces the behaviour | n/a |
The prohibited row matters as much as the fixed ones. A ledger that only lists what is fixed leaves the commonest deviation open, because that deviation is not a substitution, it is a deletion. Naming the deletion in advance gives the site lead something to point at when the shift runs late: not “the trainer decided”, but “this one is not ours to cut”.
Notice how short the fixed column is. Eight rows, three fixed, one prohibited, four local. A site lead can hold that in his head. A thirty-page standard operating procedure cannot, so it gets filed and the room goes back to instinct.
The site pack, in the order the work happens
The ledger is not the pack. It is the first page of a six-page pack a site actually uses.
- The ledger extract, one page, filtered to this programme, evidence column visible so the site knows what it will be asked to return.
- The practice scripts, two pages: the case format the pairs work from, and the four-step standard written so a peer can score it.
- The local adaptation brief, one page with blanks. Which case, which language, where the practice happens, how many pairs, who observes. Fifteen minutes to fill, and it is the moment the site commits to a plan rather than a date.
- The return form, one page, five fields.
- The escalation line, half a page: who the site calls when the practice block cannot run, and what the programme will do about it. Without it, a site with a real constraint has only two options, cancel or improvise, and both are worse than a phone call.
The brief is the page that changes behaviour, because it forces the local decisions to be made before the day, by the person accountable for them, instead of in the room by a trainer under time pressure.
One artefact, five fields
Every site returns the same five things within 48 hours of each cohort. Not a report. Five fields.
- Minutes the practice block actually ran, against the fixed 50.
- Number of participants who completed a full practice cycle to the standard.
- The local case used, in one line.
- The block that got squeezed, chosen from a list of four, or “none”.
- The one thing that did not work.
Field four is the early warning system. If eleven of twenty sites report that the same block got squeezed, the design is wrong, not the sites. Field two is the honest one: a full room with a low completion count is the most useful signal in the file.
Read the returns as three categories
At the monthly review, sort every deviation into one of three buckets. On standard. Declared adaptation, meaning the site said what it changed and why. Silent drift, meaning the change only shows up when someone looks at the deck.
Only the third is a failure. A declared adaptation is information. A site that used its own regional complaint instead of the national case, and said so, has improved the programme for its people at no cost to the fixed elements. Treat that as non-compliance and you teach the network to stop telling you things.
Silent drift usually appears as a pattern across returns rather than as an incident. Completion counts low everywhere means the standard is unclear. The practice block short everywhere means the design is over-packed. Short at four sites only means those four share a constraint, and that is a conversation, not a memo.
A worked example
Illustrative example. A service network with twenty locations ran a six-module frontline programme to about 900 people a year, delivered by site trainers. After year one, an audit of twelve sessions found the practice block had run in seven, at an average of 22 minutes against a designed 50. Four sites had dropped the peer feedback step entirely. Nobody had decided any of this centrally.
The redesign took two days: fixed column cut to three elements, ledger and six-page pack written, the four-page post-course report replaced by the five-field return. Cost was about 40 hours of design and briefing time, plus roughly an hour per site per cohort.
Over the next two quarters the practice block ran at or above 45 minutes in 19 of 22 audited sessions, and completion counts came back from all twenty sites for both cohorts. Two sites declared a real constraint, a night shift that could not release people for 50 minutes in one block, and both ran it as two 25-minute blocks in the same week. The ledger had not anticipated that and should have. It was amended once, centrally, and went out to all twenty.
Those figures are illustrative, not a client result. The transferable part is that the fix was subtraction plus one comparable return, not more auditing.
The honest limit
This works when the locations do the same job. Twenty dealerships selling the same product to similar customers can share fixed elements, because the behaviour being taught is the same behaviour. Two business units with genuinely different jobs, a service engineer and a telesales agent, should not be forced into one ledger. Split the programme and write two standards.
The second limit is skill. A ledger tells a trainer what not to cut. It does not make him able to run a 50-minute practice block with sixteen people and hold the standard while doing it. That is a separate problem, and it is the one a train-the-trainer handover is for.
The third limit is time. The brief and the return cost about an hour per site per cohort, and that hour is real. Budget it in the price rather than assuming goodwill, because the sites that skip the return are usually the ones with the least slack, and the ones drifting.
What to do on Monday
Take one live programme. List every element in the design file on a single page, one per row. Apply the three tests and mark each row fixed, local or prohibited. You will end up with two or three fixed rows if you are honest, and most of the page marked local.
Then delete the fixed requirements that failed the tests, which is the part that feels wrong and is the point. Send the ledger to the site leads with one sentence: these three things cannot change, everything else is yours, and here is the five-field form I need back within 48 hours.
If you are building a rollout rather than repairing one, the same logic belongs in the design brief before the first cohort, which is what we do in programme design and in consulting engagements where a client runs delivery across multiple sites. The training solutions route covers the delivery side while the ledger covers the standard.
Sources
https://youthrex.com/wp-content/uploads/2019/01/Imple-Matters-2008.pdf https://pmc.ncbi.nlm.nih.gov/articles/PMC3409469 https://nemtss.unl.edu/wp-content/uploads/2022/06/20-Essential-Components-and-Applications-of-Implementation-Fidelity.pdf