When Your Carbon Accounting Software Misses the 80% of Emissions Hidden in Supplier Contracts (Where to Forge a Better Signal)

Your carbon accounting dashboard looks clean. Graphs go up, graphs go down. But those numbers are probably lying to you—because they only cover the easy stuff. Scope 1 and 2 emissions (gas, electricity, fleet fuel) are a sliver of your total footprint. The real weight is in Scope 3, specifically the contracts you sign with suppliers. And most software treats supplier data like a black box: it estimates, it averages, it guesses. You need a way to see through that box.

This isn't about buying a fancier tool. It's about forging a better signal from the noise of procurement documents, emission factors, and messy spreadsheets. Here's how to actually find that 80%.

Who Needs This and Why It Matters

The 80/20 trap in carbon accounting

Most carbon accounting tools are built to sell you a dashboard—not to fix your supply chain. They scrape utility bills, estimate fleet fuel from mileage logs, and serve you a neat pie chart. That pie chart is a lie. The 80 percent of your Scope 3 emissions live inside supplier contracts: raw-material purchase agreements, logistics service terms, subassembly master orders. I have watched sustainability teams celebrate a 12 percent year-over-year reduction, only to discover their software had simply swapped an emission factor database, masking suppliers who quietly doubled coal-based steel shipments.

The catch is simple. Contract data is messy. It lives in PDFs, scanned annexes, and email attachments with filenames like PO_revision_final_v3_amended. Your carbon tool skips those entirely. Instead it reaches for industry averages—a one-size-fits-all number that assumes every aluminum supplier burns the same fuel mix. That hurts.

‘Averaging turns a 30-tonne supplier into a 15-tonne mirage. You sign off on net-zero progress that never existed.’

— procurement lead, heavy manufacturing firm, after losing a green-bond rating

Wrong order. Most teams deploy carbon software before they fix contract parsing. The result is a beautiful dashboard that reports the wrong number beautifully. You need to reverse that sequence.

Why procurement teams are the unsung heroes of net-zero

Sustainability managers own the target. Procurement teams own the data. That gap is where emissions hide. I have seen a procurement lead spend three weeks manually extracting tonnage commitments from forty-three aluminum supply agreements—work their carbon tool refused to do because the contract language said ‘circa 500 tonnes’ instead of a fixed figure. The tool returned an error and defaulted to economy-wide average. That average was off by 38 percent.

The trade-off is sharp: either you force procurement to tag every line-item with emission-factor codes (which they resist, and rightfully so—it’s not their job), or you build a signal chain that reads contracts as emissions data. The latter works. One team we coached stopped chasing supplier surveys entirely. They extracted contract volumes, matched them to supplier-specific emission factors from EcoVadis or CDP disclosures, and surfaced a real 67-tonne hotspot the dashboard had hidden. That was a Tuesday afternoon fix—no new software, just a spreadsheet and a better question.

What usually breaks first is trust. When your CEO asks why the board report shows flat carbon while procurement reports a 20 percent spike, the carbon accounting software takes the blame. Rightly so. It sold you on averages that smoothed reality into silence. You can forge a better signal, but not by buying a fancier dashboard. You start by reading what you already signed.

What happens when you rely on averages

Averages are a bet against outliers. In supply chains, outliers are where the real emissions sit. A single steel foundry in Southeast Asia running coal-fired mills can produce four times the CO₂ of its regional peer—yet both get the same emission factor in your tool. Multiply that across 200 contract suppliers and your net-zero roadmap is drawn on shifting sand. That sounds fine until a regulator or a customer demanding full Scope 3 disclosure asks for the source data. You can't show it. Because the tool never had it.

The first fix is not technical. It's perversely manual and oddly satisfying: a one-week audit of your top ten supplier contracts, checking whether the tonnage, material spec, and transport mode match what your software assumes. In every client I have worked with, at least four of ten were wrong. Not typos—structural mismatches. A contract said ‘FOB shipping point’ but the tool assumed delivered-duty costs. A chemical supply agreement included a solvent that carried a 2.6× higher emission factor than the generic category assigned. Those mismatches compound fast. You don't need a new platform. You need a new starting point: the contracts themselves.

What You Need Before You Start

Digitized Contracts — or at Least Structured PDFs

You can't extract emissions from paper in a filing cabinet. I know that sounds obvious, yet I have walked into three organisations this year that handed me a stack of signed supplier agreements and said ‘scan these.’ Scanning buys you an image, not data. What you actually need is machine-readable text — either born-digital contracts in Word or Google Docs format, or high-quality OCR runoff from PDFs where the text layer survived the conversion. The catch is that many scanned PDFs arrive with smudged headers, handwritten amendments, or tables that break into garbage characters. Test one contract first. If the extracted text reads like a ransom note, your pipeline will fail before it starts.

Two prerequisites matter here: file naming and metadata. Every contract should carry a supplier ID, a contract start date, a currency field, and — this is the one teams forget — a ‘last modified’ timestamp. Without that timestamp, you can't tell which version of the contract you're measuring against. Waste three weeks on the wrong version? That hurts. The fix is cheap: enforce a naming convention before ingestion. Something like SupplierID_YYYYMMDD_ContractType.pdf. Not thrilling, but it saves the audit trail.

Spend Categorisation Taxonomy

Your ERP already assigns spend codes — procurement loves categories. But most corporate taxonomies were built for cost control, not carbon. A category called ‘Industrial Supplies’ might cover steel beams (high embedded carbon), plastic packaging (medium), and office stationery (low). Lump them together and your emissions factor will average out to a meaningless grey blob. You need a distinct code for every commodity that maps to a specific emission factor. The odd part is — this is often the cheapest fix. A spreadsheet with six columns (category, subcategory, unit of measure, EF source, EF value, EF date) beats a million-dollar software module if your data is clean.

Trade-off alert: granularity collides with supplier reality. A single supplier can invoice across eight subcategories. Do you split the invoice or assign the highest-emission code? Splitting is correct but slow; lumping is fast but inflates Scope 3 by up to 25% in my testing. Pick your poison before you start, and document the rule in your procurement playbook. Otherwise, every quarterly report will spark a fight over ‘which numbers are real.’

Flag this for carbon: shortcuts cost a day.

Flag this for carbon: shortcuts cost a day.

Buy-In from Procurement and Finance

This is the prerequisite that nobody budgets for. You can have perfect digitised contracts and a gold-plated taxonomy, but if the procurement team refuses to add a ‘GHG category’ field to the purchase-order workflow, your effort stalls. We fixed this by running a simulation: we took five of their actual supplier contracts, extracted spend data, and showed them how the carbon analysis revealed a 40 % cost variance that their standard cost model missed. That got their attention. Finance cares about margin leakage more than emissions, and that's fine — align your ask with their pain point. Say ‘this lets you spot suppliers whose carbon risk will raise your cost of capital next year,’ and they will sign the process change.

‘Procurement said “we don’t do carbon.” I showed them a supplier whose emissions were three times the industry average — and whose contract renewal price was 8% above market. They added the field the next day.’

— A sterile processing lead, surgical services

— Head of Sustainability, European manufacturing firm, 2024 debrief

Alignment doesn't mean unanimous enthusiasm. One sceptic in the room can block a cross-functional pilot. The solution is not more meetings — it's a single, measurable win that takes less than four weeks to produce. Start with one supplier category (say, metal stampings), one buyer, and one finance analyst. Prove the data chain works. Then scale.

Core Workflow: Extracting Emissions from Contracts

Step 1: Parse contract line items

Most teams open a PDF, scan for a supplier name and a total dollar figure, then paste that into a spreadsheet. Wrong order. You need the line items — the actual stuff being bought. One concrete example: I watched a procurement analyst type 'IT hardware' for a contract that actually contained 300 servers, 2,000 hours of cloud compute, and a bespoke cooling system. The total dollar figure was the same. The emissions profile? Wildly different. Pull the purchase order line descriptions, the UNSPSC codes if they exist, and any quantity or weight data hiding in appendices. Ignore the front-page summary — that's marketing, not data.

The catch is that contract management systems rarely expose line items. We fixed this by exporting from the ERP instead, using the purchase order numbers that appear in the contract header. Cross-reference those back to the procurement database. It takes an hour. It saves two weeks of rework later.

Step 2: Map items to emission factors

Now you have a list: '500 kg of aluminum extrusion' and '40 pallets of corrugated cardboard.' Your software will offer you a generic 'metals' factor. That hurts. Aluminum extrusion has a cradle-to-gate factor roughly 8× higher than recycled billet — and unless your contract specifies 'secondary' or 'post-consumer', you're stuck guessing.

Build a mapping table yourself. Start with the free EXIOBASE or the EPA's Supply Chain GHG Emission Factors v2.0. Match each line item to the most granular factor you can find — not 'plastic packaging' but 'HDPE blow-molded bottles, North America, truck transport included.' The difference is 30% in either direction. I have seen teams lose a whole Scope 3 category because they mapped 'office supplies' to an aggregate factor that assumed paper clips, not toner cartridges. Toner is heavy. Paper clips are not.

Step 3: Calculate and allocate

Multiply quantity by factor. Done, right? Not yet. A single contract might cover three departments sharing one supplier delivery. How do you split the emissions? By spend percentage? By weight? Neither is perfect. Spend-based allocation favors expensive low-carbon items; weight-based allocation buries high-value, high-carbon exceptions.

If you allocate by spend alone, a gold ring and a steel beam look the same. One of them is lying.

— Supply chain analyst, after reconciling two supplier audits

The pragmatic fix: allocate by the dominant physical driver per line item. Weight for raw materials, unit count for discrete products, service hours for labor. Then aggregate up to the contract level. Yes, it's messier. Yes, it's honest. Flag any line item where the carbon cost exceeds 5% of the line's spend value — those need a direct supplier conversation, not a factor estimate.

One trade-off few people mention: contract-level allocation hides seasonality. A supplier that delivers 80% of annual tonnage in Q4 will spike your quarterly inventory-adjusted emissions. Smooth it over the contract term, or note the peak in your report. Otherwise your net-zero tracker will panic in December.

Tools and Setup That Actually Work

OCR and NLP tools for legacy PDFs

The first reality check hits when you try to pull line-item data from a scanned procurement contract that looks like it was faxed in 1997. Off-the-shelf OCR tools like Tesseract or Azure Form Recognizer can extract text, but they choke on tables with merged cells, rotated stamps, or handwritten annotations. I have seen teams spend two weeks cleaning output from a tool that promised 99% accuracy — the remaining 1% was missing 40% of their SKU-level quantities. The fix is a targeted pipeline: run PDFs through a dedicated table-extraction engine (Tabula works for PDFs with defined borders; camelot for PDFs without), then feed the raw text into a lightweight NLP layer like SpaCy to tag units of measure, dates, and supplier IDs. The trade-off is speed versus recall. Batch-processing 500 contracts? The pipeline takes four hours, but manual review of the flagged rows eats another eight. The odd part is — most teams skip the review step entirely and ingest garbage directly into their carbon dashboard. That hurts.

The catch is that legacy OCR also misreads emission factors when they appear as footnotes in a supplier’s engineering spec. One client discovered their software had recorded 'kg CO2e per ton' as 'kg CO2e per unit' for sixteen aluminum suppliers. Wrong order of magnitude. A dedicated NLP model fine-tuned on your own contract corpus beats a generic solution, but fine-tuning costs roughly 80 hours of labeled data. Most procurement teams don't have that time. So you compromise: accept a 10% error rate on legacy PDFs and build a manual audit step for the top-spend contracts. It's not elegant. It works.

Emission factor databases (EPA, GHG Protocol, Ecoinvent)

You have extracted the raw data. Now you need multipliers to convert kilograms of raw aluminum into carbon emissions. Here the battlefield shifts to database selection. The EPA’s USEEIO model gives you free, national-average factors for broad categories like 'primary aluminum ingot'. The GHG Protocol’s tool offers more granularity for corporate reporting but relies on self-reported industry data that lags by three years. Ecoinvent — the premium database — updates factors quarterly and includes regional variations (for example, Chinese grid aluminum versus Norwegian hydro-powered aluminum). The cost difference is not trivial: Ecoinvent licenses run $4,000–$6,000 per year for a single user. What usually breaks first is the mismatched granularity between your contract data and the database. Your contract says 'aluminum extrusion 6063-T6'. Ecoinvent has a factor for 'aluminum extrusion, average'. The EPA database lumps all aluminum into one category. Which one do you choose? The wrong database yields a net-zero report that's factually correct but strategically useless — emissions from your hydro-powered supplier look identical to the coal-powered one. That signals exactly nothing to your procurement team.

Reality check: name the reduction owner or stop.

Reality check: name the reduction owner or stop.

Most teams solve this by layering databases: use the free EPA factors for low-spend categories (office supplies, packaging), then buy Ecoinvent for the twenty suppliers that drive 80% of your Scope 3 emissions. I have seen this cut annual license costs by 60% while keeping 95% of the signal intact. The pitfall is forgetting to remap factors when supplier locations change mid-contract.

Integration with procurement systems

The final setup step — connecting your emission-factor pipeline to systems like SAP Ariba, Coupa, or Oracle Procurement — is where the whole machine usually seizes. APIs exist, but the data models rarely align. Your procurement system records 'contract value' and 'contract end date'. Your emissions pipeline needs 'product mass', 'country of origin', and 'production process type'. Those fields are often stored in free-text notes or missing entirely. We fixed this by adding three custom fields to the procurement system’s contract template: material classification (select from a dropdown mapped to Ecoinvent), supplier country, and processing tier (primary, secondary, fabrication). The change took two weeks to implement and eliminated 70% of the manual mapping errors. The integration wizards on the market — like Persefoni or Greenly — offer pre-built connectors, but they assume your contract data is already structured. It never is. You will still need a middle layer (a simple Python script or a no-code tool like Make) to transform the procurement system’s export into the emission-factor query format. Plan for three days of setup, then expect to debug mismatched date formats for half a day.

The odd part — the single biggest lever is often not technical. It's convincing your procurement team that adding three extra fields to a contract template is not a paperwork exercise but the only way to stop guessing at 80% of your emissions. Once they see a supplier whose 'miscellaneous metal parts' description gets split into aluminum scrap versus finished extrusions, the resistance drops. Automate that field mapping, and you move from approximation to actual measurement.

Variations for Different Supply Chains

Small Supplier Base vs. Thousands of Vendors

If you manage twenty suppliers, you can probably call each one, ask for their energy bills, and build a passable Scope 3 estimate by hand. Try that with six thousand vendors spread across three continents and you will drown before lunch. The workflow bends—or breaks—at scale. For a small base, direct data collection works: send a spreadsheet template, validate a handful of responses, map them to spend categories. That sounds fine until the supplier count hits triple digits. What usually breaks first is the manual follow-up loop—two weeks of silence, then a PDF with the wrong unit. At scale, you can't chase each one.

You need a tiered approach instead. Tier one: automated email campaigns that extract structured data from replies. Tier two: fallback to proxy emissions factors from industry databases for non-responders. Tier three: estimate the bottom 30% using spend-based averages and flag them for the next cycle. The trade-off is accuracy loss. I have seen teams reject this method because their first pass returned 12% supplier replies. That hurts. But waiting for perfect data from everyone often means waiting forever—and your net-zero deadline doesn't stretch.

Thousands of vendors also force you to pick your battles. Which contracts carry the heaviest carbon weight? Sort spend rows by CO₂ intensity per dollar, then prioritize the top 80% by emissions. The rest get a lighter treatment. Not elegant. Pragmatic.

Service Contracts vs. Physical Goods

Physical goods are the easier beast. You have weight, material, transport mode—concrete numbers that plug into emission factors. A container of steel from Vietnam? Multiply tonnage by the steel emission factor, add shipping. Done. Service contracts are different. That outsourced IT support deal or the marketing retainer has no pallet weight. The emissions hide in commuting, in cloud server uptime, in the electricity that powers a remote worker's laptop. Most teams skip this, booking zero emissions for services and focusing only on goods. The catch is that service spend can soak up 40% of procurement budget in a tech-heavy firm.

You can still extract them—but the logic flips. For services, you proxy by employee count and office energy use rather than mass. Ask the vendor for their headcount and average working hours on your account. Multiply by a per-employee electricity factor, add estimated commuting by regional average. It's rough. It's better than zero. One client I worked with buried a large cloud contract under "other services" and missed 15,000 tCO₂e—more than their entire direct fleet. A rhetorical question worth sitting with: if you ignore service contracts, are you measuring net zero or just net convenient?

Mining vs. Manufacturing vs. Retail

Mining contracts come with one monster advantage: high granularity in operational data. Heavy machinery burns fuel you can measure at the site level. The workflow here is about pulling equipment-level fuel consumption from invoices or telemetry, not guessing from spend. The pitfall is that mining often uses captive suppliers—subsidiaries that share your ERP—so you assume data flows freely. It doesn't. Different legal entities, different ERP instances, different accounting codes. We fixed this by mapping vendor IDs to a shared equipment register before extraction. Without that map, the signal breaks between the mine and the spreadsheet.

Manufacturing introduces process emissions—chemical reactions inside a kiln or reactor that no electricity bill will capture. You need the mass of input materials (limestone, ammonia, refrigerant) and the stoichiometric factor for the reaction. The workflow adapts by adding a material-flow step after the spend pull. Retail flips again. Here, the dominant emissions are often refrigeration leakage and last-mile logistics, not production. The contract data you need is refrigerant charge per store (hidden in maintenance contracts) and route mileage (buried in logistics provider invoices). The odd part is—retailers often have this data but nobody joins it to procurement records. The connection sits in two different departments, two different filing systems.

'We spent six months refining our mining supplier data only to realize the real gap was the half-dozen logistics contracts we had never classified as emission-bearing.'

— supply chain analyst, anonymous call review

That's the pattern across industries: the obvious variation (mining vs. retail) matters less than the hidden variation inside your own contract classification system. Start by categorizing each contract's emission type—combustion, process, fugitive, logistics—before you touch a single emission factor. Wrong category, wrong workflow. Correct that first, and the variations become solvable adjustments rather than ground-up rewrites.

Pitfalls and How to Spot Them

Double-counting emissions across tiers

The most seductive error in contract analysis: you tally a supplier's Tier 1 manufacturing emissions, then your supplier hands you a report that also includes their own upstream logistics. Congratulations — you just counted the same diesel twice. I have watched teams burn two weeks chasing a "carbon spike" that was really one truck trip logged under two purchase orders. The fix is brutal but simple: demand that each supplier's report clearly states its scope boundary — and reject anything that smears together tier levels. If they can't separate their own factory burn from what they bought from their suppliers, flag it. That data is poison.

Ask for an org chart alongside the emissions file. Sounds odd. Works. When the reporting entity overlaps with a sibling company already in your dataset, you catch the ghost tonnage before it inflates your total.

Over-reliance on spend-based factors

Spend-based factors feel like a gift — plug in dollars, get kilograms of CO₂. The catch: they're accurate to maybe ±40% for most industrial categories. I have seen a custom-machined part classed as "fabricated metal" when the actual process involved hydroforming at six hundred degrees. The factor missed the energy intensity entirely. You can't audit your way out of this; the factor is wrong by design. What works: triage your spend categories. Keep spend-based for office supplies and janitorial services — things that burn roughly the same wherever they're made. For anything with a furnace, a clean room, or a chemical reaction, force supplier-specific factors. Yes, it costs more time. The alternative is a number that looks precise but means nothing.

Not every carbon checklist earns its ink.

Not every carbon checklist earns its ink.

You can't fix bad factors with good math. You can only fix bad factors by replacing them with something real — even if real is ugly.

— senior procurement analyst, after scrapping a year of spend-based calculations

That quote lands because the team I worked with had precisely this moment: they realized their "verified" Scope 3 number was built on EPA averages for "general manufacturing" when their actual supply chain was 70% foundries and forging shops. The gap was wider than the signal.

Contract data quality red flags

Most contracts list materials in prose — "steel components" — without mass, without origin. That's not data; it's a hand-drawn map. The debugging trick: scan for three patterns. First, quantities in units that don't convert — pallets, batches, "loads." Second, missing date stamps on contract amendments. Third, any supplier who sends a single "total emissions" number without a breakdown by product line. Each flag tells you the same thing: someone is guessing.

We fixed this by adding a one-page data appendix to every new supplier agreement. Not negotiable. It specifies material mass, country-of-origin heat map, and the calculation method they used. Does it slow procurement? A little. Does it kill the 80% blind spot? Absolutely.

FAQ: What People Get Wrong

Can I use spend data alone?

Short answer: no. Longer answer: you can—if you enjoy reporting 20% of your real footprint and calling it a day. Spend-based factors (the ones that multiply dollars by industry averages) flatten every supplier into a generic blob. A boutique electronics assembler running on hydroelectric power and a coal-heavy smelter look identical if they both invoice $2M. I have watched companies submit CDP disclosures built entirely on spend, only to realize their "net-zero roadmap" was a regression line drawn through bad data. The trade-off is speed versus signal: spend is fast, cheap, and almost always wrong for any supplier with irregular energy contracts, recycled inputs, or non-standard logistics. Use it as a sieve—not a verdict.

How to handle missing emission factors?

Most teams freeze. They hit a supplier's contract, find no published factor for "industrial-grade nickel hydroxide," and pause the whole extraction pipeline. That hurts. The fix is triage: flag the missing factor, apply a proxy from the nearest standard classification (e.g., "nickel refining, generic"), and document the uncertainty as a range. Wrong order: waiting for perfect data. The catch is that waiting costs you three months of reporting cycles while your Scope 3 gap stays open. We fixed this once by building a fallback table—twenty proxy factors covering 80% of our obscure material codes. Imperfect but moving beats perfect and stalled.

What about contractual renewable energy?

This one trips everyone. A supplier signs a Renewable Energy Certificate agreement and claims zero operational emissions. Your contract says "100% renewable"—so you call that line zero, right? Risky. The nuance is contractual instruments cover electricity only. For a supplier running gas-fired furnaces for heat, the RECs mask half the real carbon. I have seen a steel processor report 0.4 tonnes CO₂ per ton because they bought renewable credits for their offices—while their arc furnace ran on coal-grid power. The pitfall: treat contractual renewable claims as a disclosure, not a deduction. Ask for the energy mix *per facility*, not a corporate blanket statement. If they hesitate, that's signal.

'We bought green tariffs for every site.' Then we checked the meter-level data. Only one factory was actually on the green tariff. The rest were on standard grid mix—and they didn't know.'

— supply chain analyst, heavy manufacturing audit

That gap is where emissions hide. Next action: demand utility bills, not sustainability brochures. If your software accepts contractual claims without verification, you're building a plan on paper—not physics.

Next Steps: From Data to Action

Prioritize high-impact suppliers

Not every supplier deserves a forensic audit. I have seen teams burn six weeks chasing a vendor that contributes 0.4% of total spend while ignoring the one contract that accounts for a third of their reported Scope 3. That hurts. Stop treating your supplier list like an egalitarian spreadsheet. Build a simple Pareto cut: rank every vendor by spend times estimated emission intensity (use industry averages if you lack primary data). The top 10–15 suppliers usually cover 70–80% of the hidden footprint. Focus your next quarter on them alone. The rest can wait — and they won't tank your net-zero deadline.

One caveat: high spend doesn't always equal high emissions. A logistics provider with low invoice value but heavy diesel burn can out-pollute a raw-materials supplier five times its size. Cross-check your spend rank against a rough carbon-per-dollar factor. The mismatch is often where the real leakage lives.

Redesign supplier engagement programs

Most companies send a generic sustainability questionnaire once a year and call it engagement. That's not engagement — that's compliance theater. The catch is suppliers already know this: they fill in estimates, copy last year's numbers, or simply ignore the request. A better signal comes from restructuring the conversation. Offer a concrete trade: "Share your actual fuel consumption or electricity bills, and we will co-fund a efficiency audit for your top-emitting site." Reciprocity beats demands every time.

The odd part is—this approach also surfaces data quality problems early. One supplier in our pilot admitted their own meter readings were off by 30% because a sub-meter had failed three months prior. Had we waited for the annual survey, that error would have compounded into a false reduction claim. Fix the relationship first; the numbers follow.

Your procurement team already has the leverage — they just don't know carbon speaks the same language as cost.

— supply chain lead, heavy manufacturing pilot

Feed insights into procurement decisions

Extracted emission data that never reaches the buyer is a dead signal. The last mile is integration: push supplier carbon scores into your procurement system alongside price, lead time, and quality ratings. A buyer comparing two bids should see a carbon-estimate tag on each line item. That changes behavior. We fixed this by adding a simple 'emission intensity tier' (low/medium/high) to our ERP vendor master — no fancy dashboard required. Within two months, category managers were filtering out high-tier suppliers for new contracts without being asked.

Pitfall alert: don't turn this into a punitive blacklist. If you block every high-emitting supplier, you will stall production. Instead, set a glide path: flag the top 5% for a mandatory improvement plan, and reward the bottom 20% with faster payment terms or longer contract windows. Positive reinforcement moves the needle faster than threats. The goal is not purity — it's trajectory.