Conclusions and next steps

Eight teams built eight independent submissions at House London #0 (see solutions). Underneath all eight, and underneath the data briefs and policy briefs teams worked from (see problems), is a small number of decisions, made by identifiable actors, at identifiable moments, under identifiable incentives. This page is organized around those six decisions, who decides what, when, where, why, and how, rather than around the eight submissions or around topics, because that's the frame that makes "House London found something the hackathon didn't" legible without having read any of the underlying repos.

Claims below are labelled descriptive (directly measured, this session or by the source projects) or inferential (concluded by combining measurements, or genuinely open). Several of these decisions have real, acknowledged evidence gaps; papering over them would defeat the point of asking who/what/when/where/why/how in the first place.

What's original to House London

Worth featuring prominently: these are findings the project itself produced by connecting or stress-testing the eight submissions, not things any single submission claimed on its own.

The two statutory deadlines are the same mechanism at two different scales. The 8-week council-decision deadline (The Spike's finding) and the 3-year construction-lapse deadline (found this session, untested by any of the eight) produce the identical signature: sharp bunching exactly at the legal cliff, and the bunched cohort performing worse (more refusals; fewer completions) than the non-bunched cohort. Nobody who built the eight projects connected these.
London Land Ledger's and Stalled London's stalled-homes figures are both probably overestimates, for different reasons, confirmed by testing named sites against live Planning London Datahub data, not by critiquing their methodology in the abstract. Both projects' own caveat sections already admitted the risk; House London is what actually checked it and found real, large, under-construction sites misclassified.
Havering's ~0% refusal rate has a diagnosed, specific, fixable root cause: a substring-matching bug in how PlanIt maps Havering's native portal status text, not just a "this looks broken" flag. Every hackathon submission that touched Havering treated it as an exclusion; none diagnosed why.
A clean, quotable answer to "which scale of development actually delivers London's housing": small applications are a stable ~15% of delivered units (both all-time and since 2015), large (100+) schemes 59–72% and rising, computed directly from the Planning London Datahub, something none of the eight submissions stated as a single number.

The six decisions

Read decision-by-decision rather than topic-by-topic, the same evidence base underneath all eight submissions and both briefs tracks reorganizes around who decides what, when, where, why, and how, for the six decision points that chain together into "the housing crisis."

A. What to propose

Who: The applicant: anyone from a homeowner to a major developer.
What: The specific scheme: extension, conversion, new dwelling(s), major regeneration.
When: Locked in at submission (amendments happen, but the core proposal rarely changes character).
Where: The mix varies sharply by geography. Among Planning London Datahub records with a real unit count, Inner London is 82.7% small (1–9 units) schemes vs. Outer London's 86.3%, with Inner London skewing toward the very largest (1.3% at 250+ vs. Outer's 0.8%): density translates into scheme type, not simply scheme size.
Why: Site economics, existing built form, what's financeable: not directly tested this session.
How: This single choice is the dominant predictor of what happens next: The Spike's text-only model predicts approval almost as well alone (ROC-AUC 0.70, replicated even stronger at 0.78 against fresher data) as with borough and application type added (0.71 / 0.79); what you propose matters more than where. Concretely: conversion into flats and change of use are refused far more than the London average (63.8–65.7% vs. ~80–85% base), trees/hedge-works and basement work far less (86–92%).

Status: descriptive throughout, well-evidenced.

B. Whether to approve, and when

Who: A planning officer under delegated authority (86%+ of minor applications) or, rarely for this class, a committee.
What: Approve / approve with conditions / refuse.
When: Bound by a statutory 8-week (minor) or 13-week (major) decision clock.
Where: Enormous borough variation on every axis measured: raw approval rates from 63.8% (Greenwich) to 97.2% (Old Oak Park Royal); deadline-bunching from 1.2% (Camden) to 39.8% (Bexley) in the most recent data.
Why: Councils are measured against national timeliness KPIs (MHCLG's P152/P154, "% decided within 8 weeks"), a metric that rewards deciding by the deadline regardless of whether the case is actually ready, which is The Spike's central thesis.
How: Decisions bunch sharply at the exact statutory deadline, 21.6% of decisions historically, 13.9–15.2% in 2023–2025 data, both far above the ~7–10× chance baseline, and decisions made in that rush are refused meaningfully more often (82.1% vs. 71.9% historically; 86.7% vs. 75.5% replicated fresh).

Status: descriptive (the bunching, the borough variation, the refusal gap) and now causally tested rather than merely observed. Two concrete alternative explanations, committee-referral scheduling and case complexity via consultation intensity, were tested directly against fresher 2023–2025 data and both failed to explain the gap away; it persisted, and in one case widened, under stratification.

C. Whether/when to start building

Who: The developer/landowner holding the permission.
What: Mobilise finance and contractors and legally "commence", or don't.
When: Bound by a 3-year statutory lapse clock (permission lapses if not "implemented" within 3 years of decision, unless conditioned otherwise).
Where: The bunching-at-deadline behaviour itself appears at every scheme size tested (excess ratio 3.4×–7.2×, no clean size gradient); but the resulting stock of stalled homes concentrates overwhelmingly in large schemes: 250+ unit schemes are under 2% of stalled sites but 70.6% of stalled units.
Why: The lapse rule creates the same "use it or lose it" incentive structure as the 8-week decision clock in B, a second, independent instance of the identical pattern, at a different threshold.
How: Commencements spike 6.9× at exactly day 1,095 after decision (189 on the day itself vs. a local daily mean of 27.3), and near-cliff commencements are measurably less likely to be genuine: only 83.1% go on to complete vs. 96–98% for commencements well before the deadline, and those that do complete take a median 548 days vs. 252–364 for organic starts.

Status: descriptive and directly tested live against 94,718 Planning London Datahub records this session. One honest caveat carried over: the completion-rate gap can't fully rule out a confound (near-cliff commencers being already-troubled schemes that both delay and under-deliver for unrelated reasons), but that confound doesn't explain the spike's sharpness, which is hard to produce except via a hard calendar deadline.

D. Whether to actually complete

The biggest evidence gap in the whole session.

Who: The developer, subject to conditions genuinely outside this data's reach.
What: Seeing the project through, vs. leaving it stalled indefinitely.
When: Only 57.5% of homes approved 2019/20–2023/24 have completed (112,104 never even started, per Homes vs Hotels); independently, 100,477 homes approved 2015–2021 show no recorded start, averaging 7.7 years (Stalled London).
Where: This is where plan and delivery diverge most sharply, and geographically it's the same borough on both counts: Kensington and Chelsea holds unstarted permitted capacity equal to 131% of its entire ten-year target, while simultaneously carrying the highest demand pressure measured anywhere in London (8.0× the London-wide average).
Why: Genuinely open. Financing conditions, phased land assembly, deliberate banking, and simple multi-year build-out time for very large schemes are all consistent with what's measured; nothing this session isolates why a specific stalled scheme is stalled.
How: Partly via the token-commencement mechanism in C; partly via genuinely long build times that scale steeply with scheme size.

Status: the volume figures are descriptive and triangulated three independent ways. But the two headline stock figures both turned out to be measurably too high, for different reasons on each side. Of London Land Ledger's top-15 stalled sites, the only two with a recoverable, genuine application reference both turned out to be under active construction when cross-checked (Barnet, 2,088 dwellings; Hackney, 2,696, Woodberry Down Estate). Stalled London's headline Tower Hamlets flag (Ailsa Wharf) tracks the lapse of a Section 73 minor amendment application, not the primary permission it explicitly amends; the site is independently reported under construction elsewhere. Both errors point the same direction: the true scale of decision D is smaller than either project's headline number, though by how much isn't quantified.

E. How much housing London commits to, and where

Who: The Mayor of London / GLA (London Plan), working within each borough's Local Plan.
What: A 10-year borough-by-borough housing target (487,660 net homes, 2019/20–2028/29 combined) and a small-sites policy split target (~38–40% of supply).
When: London Plan 2021.
Where: Targets are allocated per borough; London Land Ledger's own data source (SHLAA-style capacity assessment) implies a capacity-led allocation, not a demand-led one.
Why: Not directly tested this session; inferential from the mismatch below, not from the GLA's own stated methodology.
How: The allocation doesn't track demand pressure: 31 of 33 boroughs contain no neighbourhood at all where supply exceeds demand (Where Should London Build?), and 14 boroughs combine above-average demand pressure with delivery below their own target, yet actual delivery looks nothing like the small-sites policy split either: small applications are a stable ~15% of units delivered against a policy target of 38–40% that was never, per the available evidence, actually being hit.

Status: the demand-mismatch and small-sites-shortfall figures are descriptive. The claim that the target-setting methodology itself is capacity-led rather than demand-led is inferential: nobody in this session read the GLA's SHLAA methodology directly to confirm it.

F. Who bears the cost while the gap persists

Who: Central government (DWP) sets the subsidy formula; individual boroughs absorb the shortfall.
What: Housing Benefit subsidy for temporary accommodation capped at 90% of January 2011 Local Housing Allowance rates.
When: Since April 2011, "intended to last until March 2013" per the original DWP consultation, now 14+ years unchanged.
Where: Falls hardest on boroughs with the least council-owned housing stock: boroughs with 20%+ council/HA stock lose £6,688/household/year unfunded vs. £12,062 for boroughs under 20% (TA Watch).
Why: A specific, identifiable statutory mechanism, not council mismanagement: TA Watch's falsification check (nightly-paid-accommodation share vs. unfunded cost, r=−0.05) rules out "some councils just buy pricier placements" as the explanation; only stock ownership predicts the gap (r=−0.62).
How: A flat percentage cap on a frozen 2011 baseline, with one explicit statutory exception: households placed in council-owned stock outside the Housing Revenue Account are reimbursed at current rates.

Status: descriptive, well-sourced by TA Watch's own methodology. The closing, cross-cutting finding worth carrying into this framework: decision F's geography doesn't track decision D's. TA pressure and failure-to-start approved homes are statistically unrelated across boroughs (rank correlation −0.21, p=0.25, Homes vs Hotels): the fiscal pain of the whole chain above doesn't land where the theoretically fixable stalled capacity actually sits.

By audience: what to do with this

Policymakers / MHCLG

Replace "% decided within 8 weeks" with an appeal-survival or quality-adjusted metric (The Spike's own recommendation, and this session validated it holds on fresher data). New, not in any submission: the 3-year lapse rule shows identical gaming; consider requiring evidence of substantial works, not a token act, to satisfy it. Also new: national/GLA registers need enforced commencement and completion reporting; the brownfield register's "Started" flag covering 49 of 5,186 sites is not a monitoring system.

Councillors / campaigners

Homes vs Hotels' per-borough lookup plus pre-written councillor email is the reusable model. TA Watch's finding that the subsidy gap is a statutory mechanism (frozen 2011 rates), not a council-competence failure, redirects the ask toward central government, not local blame. The 14-borough demand-pressure/under-delivery overlap (Where Should London Build?) is a concrete accountability list.

Planning authorities / officers

The Havering fix is a template: every borough should audit its own PlanIt/Planning London Datahub status-mapping for the same class of substring-matching error. Committee referral does not explain deadline bunching in minor applications (tested and ruled out this session); process fixes belong in the delegated-decision workflow itself, not committee scheduling.

Developers / housebuilders

The text-prediction finding (what you propose predicts approval far better than where or who decides) is directly actionable: conversion into flats and change-of-use are structurally harder to get approved everywhere, not just in specific boroughs. Large schemes are simultaneously the most productive and most stall-prone segment; anyone financing or regulating major schemes should treat a right-at-the-3-year-mark "commencement" as needing scrutiny, not as settled progress.

Data scientists / journalists

Site-ID joins between planning datasets fail more often than they work (use coordinates instead), per-record staleness varies wildly within a single dataset, and a populated "Started" or commencement field is not proof of real progress. The validated PlanIt–Planning-London-Datahub join (88% match rate, explainable misses) is itself a reusable finding for anyone else in this space.

How safe each claim is to state publicly

Safe to lead with, minimal hedging needed

Directly measured, replicated, or drawn from official statistics:

Small applications ≈ 15% of delivered London units; large (100+) schemes 59–72% and rising.
The 8-week and 3-year deadline-bunching findings (both independently tested this session).
TA Watch's £5,198 theoretical / £5,374 observed shortfall-per-household match.
Homes vs Hotels' approved-vs-completed gap (320,203 vs. 184,169, 2019/20–2023/24).

Real, but needs a caveat sentence if quoted

London Land Ledger's 171,800 and Stalled London's 100,477 stalled-homes figures: both are directionally right (large-scale stalling is real) but this session found both numerically overstated. State the finding as "at least X, likely somewhat lower" rather than repeating either headline number unqualified.

Genuinely open, present as a question not a finding

Why stalled schemes actually stall (financing? phasing? deliberate banking?), and whether the GLA's target-allocation methodology is capacity-led rather than demand-led: both flagged as inferential or untested. Good material for the next-steps half of a conversation; wrong material to state as a conclusion.

What this framework adds

Read decision-by-decision rather than topic-by-topic, the same evidence base makes one thing sharper: decisions B and C are structurally identical. A statutory deadline, 8 weeks, 3 years, creates an incentive to act at the deadline regardless of readiness, produces measurable bunching, and the bunched cohort performs worse than the non-bunched cohort on the outcome that matters, refusal, non-completion. Nobody who built these eight projects independently, nor this session before running the decision-C test, had connected those two as the same mechanism at two different scales of the pipeline. That's the value of asking who/what/when/where/why/how across the whole chain rather than auditing each project on its own terms.

The honest gaps are decision D's "why" (open) and decision E's "why" (inferential, untested directly): both are where a next pass of work would do the most to complete the picture.