This week I found a small notice on Find a Tender that I almost skipped. The title is just "Web Filtering and Monitoring". The buyer is The David Ross Education Trust, a multi-academy trust in England (a group of schools run by one organisation).
My first thought was that this is boring. Web filtering is a blocklist, right? You have a list of bad websites, the school network says no, done. I have seen this as a student myself. Half of the internet was blocked and we all knew which proxy sites still worked.
But I read the notice and then the government guidance it points to, and I changed my mind a bit. It is more like an access control problem with a safeguarding team at the end of it.
You can read the notice here.
What the notice says
First the facts, then my guesses.
It is a preliminary market engagement notice, so it is not a tender yet. The trust wants to talk to suppliers before it writes the real procurement. It was published on 26 October 2026 and edited the next day. The deadline to register interest was 6 November 2026, so that part is already closed. The estimated date for the real tender is 11 January 2027. The contract would run from 1 July 2027 to 1 July 2030, with a possible extension to 1 July 2032. I did not see a value in the notice.
The things they are interested in include:
- staff and student devices
- filtering and monitoring on-site and off-site
- working across a multi-academy trust
- "Centralised safeguarding, alerting and reporting functions"
- "Role-based administration and access controls"
- following the Department for Education filtering and monitoring expectations
They also ask suppliers if one company should do both filtering and monitoring, or if it is better to split them into separate contracts. I found that question interesting, and I come back to it at the end.
What the DfE guidance adds
The notice mentions DfE expectations, so I opened the filtering and monitoring standards on GOV.UK. The page says it was updated on 16 September 2026.
A few lines surprised me as a developer:
- Filtering should block harmful content "without unreasonably impacting teaching and learning". So blocking too much is also a failure.
- The filtering system should not have "a blanket filtering profile for all users". So one list for everybody is not enough.
- Filtering should also apply to school devices taken off-site, and to personal devices under a bring your own device (BYOD) scheme.
- Some blocklists, from the IWF and the CTIRU, "cannot be disabled, overridden, or altered by any user". Not even an admin.
- Monitoring should give "immediate reports when an incident is classed as high-risk" and also weekly reports.
- The designated safeguarding lead (DSL) is the person responsible for safeguarding issues found through monitoring.
So now I have users with different profiles, different devices, two kinds of rules (locked ones and normal ones), and two speeds of reporting. This is the point where I wanted to write code.
My tiny model
Everything in this example is invented by me. The profile names, categories and rules are not from the notice. I just wanted to see the shape of it.
from dataclasses import dataclass
# lists that nobody is allowed to switch off
LOCKED = {"child_abuse", "terrorism"}
# who may see which category (my own invented rules)
PROFILES = {
"year7": {"social_media": False, "gaming": False, "adult": False},
"year12": {"social_media": True, "gaming": False, "adult": False},
"staff": {"social_media": True, "gaming": True, "adult": False},
}
@dataclass
class Request:
user: str
profile: str
device: str # "school" or "byod"
on_site: bool
category: str
def decide(r):
if r.category in LOCKED:
return "block", "high"
allowed = PROFILES[r.profile].get(r.category, True)
if allowed:
return "allow", None
return "block", "low"
def route(r, risk):
if risk == "high":
return "alert DSL now"
if risk == "low":
return "weekly report"
return "-"
requests = [
Request("amir", "year7", "school", True, "education"),
Request("amir", "year7", "school", False, "gaming"),
Request("lena", "year12", "byod", True, "social_media"),
Request("mr_k", "staff", "school", True, "social_media"),
Request("lena", "year12", "byod", True, "terrorism"),
]
for r in requests:
action, risk = decide(r)
where = "on-site" if r.on_site else "off-site"
print(f"{r.user:5} {r.profile:6} {r.device:6} {where:8} {r.category:13} -> {action:5} | {route(r, risk)}")
When I run it I get:
amir year7 school on-site education -> allow | -
amir year7 school off-site gaming -> block | weekly report
lena year12 byod on-site social_media -> allow | -
mr_k staff school on-site social_media -> allow | -
lena year12 byod on-site terrorism -> block | alert DSL now
The important line for me is the first check in decide. The locked list is checked before the profile, so no profile can open it. If I put that check after the profile lookup, a badly configured staff profile could allow something that should never be allowed. In my first version I did exactly that, and only noticed when I wrote the comment above it.
Where my model is too simple
In quite a lot of places.
My model sends every normal block to the weekly report. But a Year 7 student trying a gaming site at lunch is probably not a safeguarding concern. If the DSL gets 4,000 lines like that every week, the one real problem gets lost. I think a big part of the real product is deciding what is noise.
I store device and on_site but my rules don't use them. In real life I guess they matter a lot. How do you even filter a school laptop at home? My guess is some agent software on the device, or traffic going through the school's system, but I don't know which is more common.
Monitoring is more than URLs. The guidance talks about alerts and observations, and some products watch what people type. That is a big privacy question for students and I don't feel ready to model it.
And the trust has many schools. Who can change the profiles? One central team, or each school? "Role-based administration" probably means that a teacher in one school should not see alerts from another school. That would be one more check in my code.
The one supplier or two question
Now I understand why the trust asks this. Filtering and monitoring share the same data: who the user is and which profile they have. If two suppliers do them, they both need the same user list and they need to agree on who is who. That sounds like a sync problem to me. But one supplier for everything means you depend on one company for five years or more.
I don't know the right answer. I am curious what people think.
If you have worked on school IT or on any content filtering system, did you split filtering and monitoring between tools, or keep them in one? And how did you stop the alerts from turning into noise?
Top comments (6)
The part that caught my attention is that filtering and monitoring are actually two very different responsibilities, even if they share the same user profiles and data.
Filtering makes a decision before access happens. Monitoring has to interpret what happened afterward, and that second part is much harder. A blocked gaming site and a genuine safeguarding concern shouldn't end up competing for the same attention in a DSL's inbox.
I'd be tempted to separate the decision engine from the alerting engine, even if both use the same identity and policy layer. That way, filtering can remain consistent while monitoring applies a more nuanced risk model.
But there's another question here: where should the system draw the line between safeguarding and surveillance? If it starts monitoring searches, typed content, and activity on personal devices outside school, how do you make sure students are protected without making them feel constantly watched?
I suspect getting that balance right is harder than writing the filtering rules themselves.
Thanks, Oleksandr. I like how you split it into before access and after. Filtering only has to say yes or no. Monitoring has to guess what something means, and it can be wrong in both directions.
Your idea also fits the trust’s question about one supplier or two. Two engines on one identity and policy layer sounds like a middle way. In theory they could even come from different companies, as long as they read the same user list. But then that shared layer is the risky part. If the profiles drift apart, a student could be “year7” in one system and “year12” in the other.
On surveillance, I agree it’s harder than the rules, and it’s the reason I didn’t model it. The guidance says the DSL is responsible for safeguarding issues found through monitoring, so maybe a good first rule is to collect only what the DSL would actually need to act on. Still, typed text on a personal device at home feels very different from a blocked URL on a school laptop. Where would you draw the line?
I think the line should be drawn around actionable safeguarding, not around how much data the system is technically capable of collecting. A blocked URL on a school-managed device is one thing; continuously analyzing someone's private conversations or typing on their personal device at home is a completely different level of intrusion.
If monitoring is necessary, it should be limited to the minimum information needed to identify a genuine safeguarding concern, with clear boundaries around what gets collected, who can access it, and how long it's retained. Otherwise, you risk creating a surveillance system that happens to be operated by a school.
I also think context matters enormously. A keyword or isolated message shouldn't automatically trigger an escalation without some human assessment. How do you prevent the system from turning ambiguous behavior into a safeguarding incident simply because it has no understanding of the context?
I like “actionable safeguarding” as the line. It gives you a test for every field: would the DSL do anything with this? If not, why collect it?
About context: my model can’t handle it at all. route() only looks at the category. It doesn’t know the time, the lesson, or what the student searched before. A history class researching extremism and a student searching the same words alone at 2am would get the same result.
The guidance talks about alerts and observations, and I think that difference helps here. A keyword could create an observation that waits for a person, not an incident. Only a human, probably the DSL, turns it into an incident. That’s close to what you said: the system collects and sorts, but it doesn’t decide what something means.
The hard part is volume. If every keyword becomes an observation, the DSL is back to 4,000 lines a week. So maybe the system’s real job is grouping: same student, same topic, several days in a row. Do you think a pattern over time is a fair signal, or does it just make the surveillance deeper?
I think patterns over time can be a fairer signal than isolated keywords, but only if the system is careful about what it actually concludes. Repeated searches could justify prioritizing something for human review, but they shouldn't automatically turn into an incident or imply intent. Context can still be missing, and repetition alone doesn't tell you why someone is researching a subject.
I'd also put limits on how much history gets retained and how broadly it's accessible. Otherwise, grouping observations solves the alert-volume problem but quietly creates a much more detailed behavioral profile of a student. The distinction between reducing noise and expanding surveillance is probably where the design needs the strongest safeguards.
Good point: a pattern can move something up the review list, but it can't tell you why. I hadn't thought about the grouping itself becoming a profile, and I think you're right. The same data that makes the alerts quieter also makes the profile bigger.
So the limits you mentioned earlier (what gets collected, who can see it, how long it stays) matter even more once you start grouping. Thanks for taking the time, this was a really useful thread.