Planning odds — appeals, land supply, constraints

v5, 24 Oct 2026: housing need, land supply and the HDT for all 94 councils

Nick approved (24 Oct): load all of it for every covered council. Before, only the 5 Oxfordshire councils had a land supply row; the other 89 fell through to "shortage route unknown" in words but ×0.3 "route shut" in the factor (a bug: routeOpen was null and the code read null as shut). Now all 94 have a row.

Three figures per council (housing-supply.csv, one source URL per figure):

Why the need figure is worked, not copied: MHCLG published a standard-method table only with the Dec 2024 NPPF. Annex D says use the latest stock and ratios, so councils recompute it every year. Check: the same sum over all England LPAs gives 367,693 (293 LPAs; 3 with changed codes missing), against 371,117 reported nationally for the May 2026 data (Pegasus, Turley). Babergh 790 and Mid Suffolk 757 match the councils' own published figures exactly. Councils' own figures can differ by a few % where they used last year's inputs (Guildford 1,168 vs our 1,151).

The rule, current NPPF (published 17 Aug 2026, updated 29 Sep 2026):

• S5(1)(j): outside settlements, housing that meets "an evidenced unmet need (including … where the local planning authority cannot demonstrate a five year supply of deliverable housing sites or scores below 75% in the most recent Housing Delivery Test)" and is physically well-related to a settlement.
• Annex D para 12: HDT under 95% → action plan; under 85% → 20% buffer; under 75% → unmet need is deemed to exist for S5(1)(j). Para 13: results apply from the day after publication.
• Annex D para 8–9: supply is measured against the plan requirement, or the standard method once the plan is over 5 years old; buffer 5%, 20% after under-delivery, or (decisions only) 20% where a plan adopted in the last 5 years under a pre-Dec-2024 NPPF has a requirement of 80% or less of the standard method (9(c)).
• Annex A para 3: with 5 years' supply and HDT over 75%, a standard method above the plan requirement is not by itself unmet need for 5 years after adoption. Not encoded: it never opens the route, only stops one argument for it.

Encoded (data/rules/land_supply_build.py): route open = under 5 years (or the council says "unable to demonstrate") or HDT under 75%. Shut = 5 years or more and HDT 75% or more. Unknown = no supply figure and HDT 75%+ (Fenland, Barnet, Kensington & Chelsea, Waltham Forest). Ashford scores exactly 75%: not "below", so its route is open on supply alone (3.62 years). Result: 70 open (short on housing), 20 shut, 4 unknown.

Unknown is not shut. Outside a built-up area the factor is ×0.6 open, ×0.3 shut and now ×0.45 unknown (the midpoint; an assumption, labelled), and the Hard cap applies only when the route is known to be shut.

The card line (planning odds section, visible without a tap; also the v1–v5 plot page's Planning row): "West Suffolk: needs 1,185 homes a year · 6.0 years' supply (Apr 2025) · built 89% of target (HDT 2025)". Each figure links to its source (need → this section; supply → the council's statement; HDT → the gov.uk publication). The date is the statement's base date. "· short on housing" is added when the route is open.

Brief filter "Council short on housing" (short_on_housing in the criteria): narrows the council list passed to suss_opportunities to the 70 short councils (intersected with any councils chosen). No migration. Unknown councils are left out. Not yet in screening agents or watches.

Read the figures with care:

• Many statements predate a rule change. The London Plan passed 5 years old in Mar 2026, so London boroughs without a newer plan should now measure against the standard method; several published figures (Croydon 5.1, Haringey 5.18, Sutton 8.8) use the old target and would likely fall. Not recalculated: the card shows the council's figure and date.
• Plans adopted under the old NPPF with a requirement under 80% of the standard method need a 20% buffer for decisions from 1 Jul 2026 (9(c)). Babergh and Mid Suffolk republished for this (5.34 → 4.68; 7.70 → 6.73). West Suffolk (6.0), Merton (5.5) and Tunbridge Wells (5.60) had not, as published; noted in their rows.
• Where an appeal tested the figure it is in the appeal column (e.g. Mid Sussex: council claims 5.69 on adoption, but agreed a maximum 3.75 years at the Ansty appeal, Jun 2026; Wandsworth 6.1 → 5.5 agreed at the Battersea Bridge inquiry).
• Figures read only through search excerpts, because the council site refuses automated requests, are marked "(excerpt)" in supply_confidence (22 councils). Worth a human check before relying on them.

Refresh: HDT every year (Aug); need when Table 125 (May) or ONS ratios (Mar) update (data/rules/housing_numbers.py); supply as councils publish (most in Nov–Jan): edit land_supply_notes.py, run land_supply_build.py, then tools/build-planning-context.mjs.

v4, 18 Oct 2026: every covered council has its own rates

Before: London (33) and East (10) used one England rate (74% of minor housing schemes granted), so odds didn't separate a Croydon plot from a Camden one. Now all 48 covered councils use their own figures.

Sources (both OGL v3, free, no account):

• MHCLG live tables on planning application statistics, "PS2 data – open data table (202606)": decisions, granted, refused, per LPA per quarter, minor dwellings (1–9) and major dwellings (10+). All 311 England LPAs loaded, last 8 quarters (2024 Q3 – 2026 Q2). Script data/summary/planning-context/scripts/ps2_all.py → council-decision-rates.json (192 kB). Oxfordshire figures match the 17 Oct pull exactly.
• Planning Inspectorate appeals casework database, Casework_Database_-_2609.xlsx: housing appeals (Minor/Major dwellings) for all 48 councils, Sep 2021 – Aug 2026, matched on ONS local authority code. 5,391 appeals, 5,305 geocoded (postcode centre, OS Code-Point Open). Script scripts/build_covered.py → appeals-covered-residential.csv. The 399 Oxfordshire rows are identical to the old file.

Method (changes from v3 in bold):

1. Base rate p1 = the council's PS2 rate for the scheme size (minor 1–9 / major 10+), shrunk to England by 20 imaginary decisions: (granted + 20 × England) ÷ (decided + 20). Same rule as the appeal term. Why: City of London decided 4 minor schemes in 2 years, Kensington 0 major; a raw 50% or 0/0 would swing the card. For councils with 100+ decisions it moves the rate by under 2 points (Oxford 66.4% → 67.5%).
2. Constraint factors, own-plot appeals, "2+ dismissed within 500 m" and caps: unchanged, now live outside Oxfordshire because the appeals are loaded there.
3. Appeal term: unchanged formula, now with each council's own appeals.
4. Fallback: a council PS2 doesn't list gets England's rate, labelled "England rate: PS2 has no figures for X". None of the 48 covered councils needs it.
5. Label: the card's source reads "MHCLG PS2, Croydon, minor dwellings, 2024 Q3–2026 Q2 + Planning Inspectorate appeals, Croydon"; the odds row shows the council's published rate, and the blend when it differs.
6. Own-plot rule tightened (the opportunities rule, ingest 45 fix 6): within 10 m no longer counts when the address puts the site on another road ("134 Crowther Road" on a Holland Road plot). Addresses that describe land from a road ("Land south of 19 Fox Lane", "rear of…") or name no road keep the 10 m rule. Test: cards at 57 sampled appeal points, 3 attributions dropped (Croydon, Hillingdon, Oxford Juxon St), all 9 round-6 Oxfordshire sites unchanged.
7. Backland factor (card-odds.ts) is now measured in 7 councils with 10+ "rear of" small appeals: Barnet ×0.5, Bromley ×0.9, Croydon ×1, Ealing ×1, Havering ×0.82, Richmond ×1, Sutton ×0.92. Others keep ×0.75.

Re-score needed (for the ranking agent, not run here): geo.appeals and geo.candidates still hold the old odds. 45-candidates.mjs loads geo.appeals from planning-context.json, so a full run now loads 5,239 appeals (was 384, Oxfordshire only) and London/East plots can get appeal_dismissed; then 46-score-candidates.mjs re-scores odds_first / odds_band with council rates and the measured backland factor. East and Oxfordshire rows are stale until then (East used England's 74%).

Exposed for the council page and agent: approvalRate(slugOrName) in planning-context.ts, and approval_rate on councilInfo().

Sanity table: covered councils, PS2, 2024 Q3 – 2026 Q2

Range: Havering 43% to Hammersmith & Fulham 96%. 18 of the 33 London boroughs grant under 60%.

Read with care:

• The inner boroughs with 90%+ (Camden 900 decisions, Islington 974, Kensington 1,175, Wandsworth 1,358, H&F, Southwark) decide 2–5× more "minor dwellings" applications than outer boroughs with only a few dozen appeals. PS2 counts every application creating 1–9 homes, so this is probably flat conversions and small changes of use, not new-build infill. Their rates likely overstate the odds for a garden or garage site. Not checked against council registers (no free application-level data).
• Westminster's 48 minor decisions in 2 years is implausibly low for its size: likely coded as "other" in its PS2 returns. Its rate rests on few cases.
• Development corporations: OPDC (Old Oak, parts of Ealing/Brent/H&F) decides its own area; LLDC did until Dec 2024. A plot there gets the borough's rate.

Before / after on sample plots (card code, 18 Oct; homes = the land value's count)

Permitted plots keep their 80% floor whatever the council rate: the floor dominates. The change bites on unpermitted plots in councils far from England's 74%: outer London (Croydon, Havering, Harrow, Kingston, Hounslow, Newham, Barking) and the high-rate inner boroughs.

v3, 17 Oct 2026 (round 5): one home count, permissions, no contradictory numbers (Critique 2 #2, #3, #9)

1. Odds are for the land value's home count, not the low end of "homes fit". The server runs the same starting-point rule as the land value (bestStart: houses or flats, affordable cliff), and the card says "For 8 homes, the land value's count". Odds for each size band (1–9, 10–49, 50–299, 300+) come with the card, so changing the homes switches the odds at once. In a National Landscape at 6–14 homes the card shows both sides: "9 homes: Fair · 10+: Hard". Chipping Norton police station: v2 said Hard, 3 in 10 at 11 homes (major rate × AONB 10+); v3 says Good, 8 in 10 at 8 homes.
2. A permission on the brownfield register for this plot (matching size, permitted, home count) floors the first-time chance at 80% for a scheme no bigger than the one permitted, shown as its own factor ("Already permitted for 8 homes (2021) × 1.6"), and lifts the Look-closer cap. Assumption: renewals and like-for-like schemes usually pass; to calibrate when application data is licensed.
3. No number that disagrees with the band. Pass → "Hard: the card says Pass ()", no number. Outside a settlement with the shortage route shut → Hard, no number. Look closer → Fair at most, and at most 7 in 10, with the flag named. Not a small site (over 1 ha, or open space) → no number, "draw the part you mean".

v2, 18 Oct 2026: what changed and why (Critique 1 #2)

v1 said "Good, about 9 in 10" on all 6 critique plots, including backland in a conservation area with 12 listed buildings nearby, and a 21-acre AONB-edge parcel for 345–517 homes. Two causes: the headline counted a won appeal as a yes, and constraints only touched the appeal term (~2 points). v2:

1. Headline = first-time chance. first = p1 × factors (council PS2 rate × the plot's constraint factors, floor 3%). Shown "about N in 10 first time". The appeal route is shown on its own line, "with an appeal: about N in 10 — adds 6–12 months and the cost of appealing". Never folded in.

2. First-time factors (multiply; judgement, anchored on appeal splits and national policy; to calibrate when nearby decisions are licensed):

   On the map	×	Anchor
   Conservation area	0.80	PINS England CA appeals ×0.68; Oxon back-test ×0.74 (below)
   1–2 listed buildings within 50 m	0.92	PINS "setting of listed building" ×0.92
   3+ listed buildings within 50 m	0.85	Oxon back-test ×0.81
   National Landscape, under 10 homes	0.85	NPPF great weight; Oxon back-test ×0.78
   National Landscape, 10+ homes	0.50	NPPF: major development needs exceptional circumstances
   Part green belt (under 50%)	0.60	inappropriate unless an exception (infill, grey belt)
   Part flood zone 3 / flood zone 2	0.60 / 0.80	sequential test; PINS flooding ×0.49
   Outside a built-up area, shortage route open	0.60	NPPF 2026 S5(1)(j), case by case
   Outside, route shut (and not brownfield)	0.30	plus the Hard cap below
   Edge of a built-up area (≤25 m, or under half the plot inside)	0.85	may be read as outside
   2+ housing appeals dismissed within 500 m, more than allowed	0.90	each was refused first

3. Band on the first-time chance: Good ≥ 75% · Fair 55–74% · Uncertain 35–54% · Hard < 35%.

4. Caps: card Pass → Hard. Outside + route shut + not brownfield → Hard. Look closer → at most Fair. 50+ homes → no number, at most Uncertain ("the local plan decides; we can't see allocations"). 300+ homes (low or high estimate) → Hard, no number ("a strategic site: only an allocation gets it through").

5. Appeal term unchanged from v1 (council appeals shrunk to England, × 0.68 CA or × 0.49 flooding).

Critique plots, v1 → v2 (live, 17 Oct 2026):

Back-test, light (18 Oct 2026)

Our own map flags tested at each Oxfordshire housing appeal's postcode point (379 decided appeals, Sep 2021–Aug 2026; ingest-style script in the session, SQL against geo.constraints):

Reading: the heritage and landscape flags on our map line up with lower win rates, in the same direction and roughly the size the v2 factors use. Outside built-up areas appeals won more — the pre-Aug-2026 tilted balance with Cherwell, South and West short of land. The outside factor (×0.6) is therefore about first-time officer decisions under the new S5 test, not appeal evidence; it is the least supported factor and the first to recalibrate when 2027 appeals arrive. This tests inspectors, not first-time decisions: we still have no free application-level data (PlanIt licence pending).

Draft, 17 Oct 2026. Implements parity.md "5 ways to be better" #2. Data and sources: README.md. Pilot: the 5 Oxfordshire councils.

First, a correction to #2

The "tilted balance" is gone. A rewritten NPPF took effect on 17 Aug 2026 (gov.uk). Old para 11d is replaced by policies S3–S5:

• Inside a settlement (S4): approve "unless the benefits … would be substantially outweighed by any adverse effects". This applies whatever the land supply.
• Outside a settlement (S5): only listed kinds of development. Housing qualifies under S5(1)(j) when there is unmet need, which includes the council being under 5 years' supply or under 75% on the Housing Delivery Test, and the site is "physically well-related to an existing settlement".

So the parity.md line "Cherwell can't show 5 years' land supply, so the tilted balance applies" should become: "Outside the village, Cherwell is short of 5 years' supply, so national policy allows housing next to a settlement (NPPF S5(1)(j))." Inside a town, land supply no longer changes the test.

All 23,003 appeals in our data were decided before this change (to 31 Aug 2026). The rates below describe the old regime. We should expect them to drift.

v1 (17 Oct 2026, superseded where v2 differs)

What the card said (one block, four lines)

1. First time: "Oxford grants 66% of small housing schemes (85 of 128, Jul 2024–Jun 2026)."
2. If refused: "About 1 in 6 appeals like this win (Oxford 5 of 20 for 1–9 homes; less in conservation areas)."
3. Overall: "About 7 in 10 — Fair."
4. Policy line: inside/outside settlement, land supply, HDT, in plain words. Plus the nearest 3 appeals with distance, outcome and a link to the decision letter.

The rule

Every step is one sum the user can tap to see.

Step 0 — size band. From "homes fit" (card-rules.md), low estimate: 1 home · 2–9 · 10–49 · 50+. Bands 1 and 2–9 together are "minor" (1–9), 10+ "major".

Step 1 — first-time grant rate, p1.

• Default: council grant rate for minor or major dwellings, last 8 quarters, PS2 (council-decision-rates.json, OGL).
• When nearby decisions are licensed (PlanIt or council feeds): blend in similar schemes (same band, new homes, within 1 km, last 5 years): p1 = (local granted + 10 × council rate) ÷ (local decided + 10). In words: the council-wide rate counts as 10 imaginary nearby decisions, so 2 or 3 local results can't swing it wildly. With 30 local results the local record dominates.

Step 2 — appeal win rate, pA. PINS casework database (OGL). pA = (council allowed + 20 × England rate) ÷ (council appeals + 20) for the same size band. England rates: 1–9 homes 21.3% (3,158 of 14,854), 10–49 36.0% (416 of 1,155), 50+ 62.0% (552 of 890).

Step 3 — constraint factors on pA. Only where the PINS flag has enough cases and shows a difference. England, 1–9 homes, flagged vs not flagged:

Most rows leave these flags blank (CA filled on 2,001 of 23,003), so the factors come from the minority PINS did tag. We have no free data on how constraints affect first-time decisions; p1 is not adjusted. That waits for nearby decisions.

Step 4 — route to yes. Assumes the developer appeals if refused: P = p1 + (1 − p1) × pA.

Step 5 — band and caps.

• P ≥ 80% Good · 60–79% Fair · 40–59% Uncertain · under 40% Hard.
• Card verdict "Pass" (card-rules.md) → Hard, whatever P says.
• Outside a settlement and the council's unmet-need route is shut (today only Vale: 5.51 years, HDT 192%) and the plot is not on the brownfield register → Hard, with: "Outside a settlement, national policy only allows set kinds of housing. Vale has 5.51 years' supply, so the shortage route is shut."
• Outside a settlement and the route is open (Cherwell, Oxford, South Oxon, West Oxon) → no cap; say so in the policy line.
• "Outside a settlement" proxy: plot centre not in the planning.data.gov.uk built-up-area layer (already pulled). This is ONS built-up area, not the plan's settlement boundary — label it "approx." (inference; the legal test is the NPPF glossary and the local plan). Built (17 Oct 2026): source.settlementFor(geometry) → SQL suss_settlement_at(). inside = a point on the plot's surface is in an ONS Built Up Area 2022; edge = not inside but the plot is within 25 m of one (treat as "physically well-related" for S5(1)(j), still "approx."); outside = neither. Also returns the nearest BUA's name and distance, share of the plot inside, and the brownfield register entry within 25 m (S5 allows brownfield outside settlements — so "outside + brownfield" must not get the Hard cap). Source and licence: ONS via planning.data.gov.uk, OGL v3. Checked for council GIS first: none of the 5 Oxfordshire councils publishes settlement boundaries as open data (see DATA.md). If one does later, load it into geo.settlements with source = 'local-plan' and prefer it.

Never show P to more than 2 figures. Write "about 7 in 10", not 71.7%.

Worked example — the Headington plot

Plot 28775681, 1,055 m², Oxford City, Headington Hill conservation area (example.ts). The HMO Article 4 is ignored (card-rules.md).

Policy line: "Inside Oxford, national policy says approve unless harm substantially outweighs the benefits (NPPF 2026 S4). Oxford built 65% of its target (HDT 2025), and extra homes on existing plots get substantial weight (L2)."

Nearest appeals (all 4 within 1 km, from headington-nearby-appeal-letters.csv):

2 allowed of 4. None of the 4 letters mentions land supply; each turned on character, the conservation area or living conditions. That fits the new S4 test: in a city, the odds turn on design and neighbours, not on housing numbers (inference from 4 letters).

Without the CA factor the answer would be 0.664 + 0.336 × 0.231 = 74%. Same band.

Does it predict anything? Held-out check (run 17 Oct 2026)

Trained Step 2 on appeals decided before 1 Sep 2025, tested on the 12 months after.

Lower Brier is better. The rule beats a flat guess, by a little. It under-predicts: appeals allowed rose from 23.8% (2024) to 27.4% (2025), and the rule averages 5 years. Fix to try: weight the last 2 years double. Report this hit rate on the card, as #2 promised.

First-time odds (Step 1) can't be tested without application-level data (PlanIt).

What we can ship now vs later

Refresh

• PINS: quarterly (file name = YYMM). HDT: yearly (Aug 2026 last). PS2: quarterly.
• 5YHLS: councils publish around December; South & Vale published 12 Dec 2025. Vale is at 5.51 — the next statement could flip it and open the shortage route.
• Cherwell: if the Local Plan Review 2042 is adopted, NPPF 2026 Annex A para 3 can switch the shortage route off for 5 years. Watch the examination.

Three things that would most change this

1. How inspectors apply S4/S5 after 17 Aug 2026. Every rate here is pre-change. The first 6 months of new-regime appeals (PINS Mar 2027 file) may move allowed rates a lot, inside settlements especially.
2. The PlanIt licence. Without nearby decisions, Step 1 is one number per council (Oxford 66%) and can't tell a good Headington plot from a bad one.
3. Oxford's real decision-taking land supply figure. Not found. HDT 65% already opens the shortage route, so the card holds; but if a figure appears it should replace the "not published" line.

Round 6 (17 Oct 2026): the plot's own appeals

Critique 3 #2: on 14 of 15 test cards an appeal for the same address sat 2–24 m away and nothing used it. Now an appeal on the plot (address point on it or within 10 m; an address naming a road must name the plot's road) is its own factor, listed first, and excluded from the "within 500 m" count:

Judgement, not fitted: no free data says how often a re-application after a dismissed appeal succeeds. The direction is not in doubt: an inspector has already agreed with the council on that land.

Critique 3 sites, round 5 → round 6 (tap at the site point):

Refused sites warned: 1 of 3 (partly) → 3 of 3.

Rules v12 (19 Oct 2026): London first-time factors fitted on decisions

London only (card.rules.region === 'london', CAL.londonOdds). Fitted by ingest/65-fit-odds.mjs on every London new-homes decision 2020–26 in geo.bt_truth (GLA planning history, 1–30 homes, each scored as of the day the application went in): 2,811 decisions in the 15 boroughs of the fit half (name hash), judged on 4,725 in the other 18. Model: chance = PS2 base rate × London × Π factor (log-binomial), so each term stays a named ×multiplier on the card. Kept: 30+ flagged decisions in each half, same direction in both. Numbers in calibration.ts LONDON_ODDS.

Held-out boroughs (4,725 decisions): odds AUC 0.535 → 0.548, Brier 0.2429 → 0.2396 (always guessing the average 0.2445). Bands said → happened: before Hard 31→41 · Uncertain 46→54 · Fair 63→62 · Good 86→72; after Uncertain 51→47 · Fair 65→60 · Good 78→73 (no plot falls under 35% any more). The held-out boroughs grant less than the fit half (57% vs 65%), so they read about 5 points high on average: borough-level differences PS2 doesn't carry. Outside London the judgement factors stay; there are no free first-time decisions to fit them on.