Halt the Harm · chemical recycling trackerSubscribelive repo data

Reports / 2026-08-06 · sweep

Chemical Recycling / Pyrolysis — Early-Warning Sweep (Aug 6, 2026)

Chemical Recycling / Pyrolysis — Early-Warning Sweep

Run date: Thursday, August 6, 2026 · Nationwide · Leads and comment windows only. No recap of known campaigns. Each item below is self-contained: facility, agency, dates, and the raw source link sit together so a single block can be copied and forwarded.

1. DEADLINE INSIDE 30 DAYS

Clean-Seas West Virginia, Inc. — Belle, Kanawha County, WV

Priority Plastics, Inc. — Portland, Jay County, IN

  • Stage: IDEM Office of Air Quality preliminary findings / draft Minor Source Operating Permit, MSOP No. M075-50299-00039. Notice posted to IDEM’s site July 26, 2026; 30-day comment clock runs from posting, so comments due on or about August 25, 2026.
  • What it is: transition from Registration to new source review + MSOP because modifications and new units push potential-to-emit over Registration limits. Applicant “intends to construct and operate new equipment that will emit air pollutants.”
  • Scope caveat — reported, not confirmed: the notice does not state a chemical-recycling process. This is a plastics manufacturer expanding air-emitting units; it belongs on the watch list, not in the pyrolysis column, until the application is pulled.
  • Comments to: Tyler Hennessy, IDEM OAQ, 100 N. Senate Ave., Room 13W, Indianapolis, IN 46204 · (317) 232-5672 · thennessy@idem.IN.gov (a U.S. mailing address must accompany comments)
  • Hearing: available on written request; IDEM decides based on adverse comments plus a hearing request.
  • Raw notice (PDF): https://www.in.gov/idem/files/notice_20260726_air_50299pn-cm.pdf
  • Application and findings via IDEM Virtual File Cabinet, Advanced Search, Program = OAQ, Permit # = 50299: https://www.in.gov/idem/legal/public-records/virtual-file-cabinet/
  • IDEM all-regions public notices: https://www.in.gov/idem/public-notices/public-notices-all-regions/

2. NEW LEADS — proposals not previously in the log

Reju (Technip Energies) — Regeneration Hub, Eastman Business Park, Rochester, NY

This is the significant find this week. A PET depolymerization plant that never uses the words pyrolysis or chemical recycling, sited in a state with no active HHN campaign, with permits not yet filed. Maximum lead time.

Waste Energy Corp (OTC: WAST) — Fayetteville, NC pyrolysis facility

Freepoint Eco-Systems Yermo Supply LLC — Yermo, San Bernardino County, CA

  • Stage: appears on the CalRecycle public notice grid twice — a New Solid Waste Facilities Permit for a “Plastics Sorting and Processing Facility” (notice item 2026-06390) and a separate Public Notice: Informational Meetings Conducted by Enforcement Agencies for the same site as a “Plastics Sorting and Recycling Facility.”
  • Why it matters: framed as sorting and processing, not conversion — but it is Freepoint, the operator whose Hebron, OH pyrolysis plant just suspended operations, and sorting/feedstock-prep capacity is the upstream half of a pyrolysis system. Treat as a feedstock-infrastructure lead.
  • CalRecycle public notices (search “Freepoint” or “Yermo”): https://www2.calrecycle.ca.gov/PublicNotices/
  • Caveat: the notice grid was retrieved as a raw JSON payload this run; exact notice dates were not legible in the captured portion. Dates need confirmation directly from CalRecycle before this is forwarded to partners.

Resynergi — relocation target state still undisclosed

MediWaste Pyrolysis Facility — Pahrump, NV (window closed; logging for completeness)

3. OKLAHOMA — ODEQ Permits for Public Review, read this run

No pyrolysis, gasification, or depolymerization permit is on the readable portion of the ODEQ public-review list. Two items are the only plastics-adjacent entries, and neither is a conversion facility on its face:

4. CHANNELS CHECKED, NOTHING NEW

5. MOVEMENT ON KNOWN FIGHTS — one line each, for context only

PROCESS NOTES — what hampered this run

  1. Water permits are a channel we were not searching, and they are the best-labeled ones. The Brightmark Ashley IWP notice describes the facility as manufacturing “liquid petroleum products from recycling and deconstruction of end-of-life plastics” and lists “pyrolysis” among unit processes in the briefing memo. Air permits often bury this; pretreatment and NPDES notices spell it out. Add to every run: state NPDES/pretreatment public notices, not just air. New trigger phrases to add to the search list: “manufactures liquid petroleum products,” “recycling and deconstruction of end-of-life plastics,” “rBHET,” “regeneration hub,” “textile-to-textile.”
  2. The ODEQ public-review page is a JavaScript grid and the scrape truncated mid-row. Only the AQD and WQD divisions were captured; no Land Protection Division / solid-waste rows were captured at all — which is precisely where a pyrolysis facility would file. The Oklahoma read is incomplete and should not be treated as clear. Browser-rendering access is needed here.
  3. CalRecycle notices came back as a raw JSON payload. Facility names were legible, notice dates were not. Freepoint Yermo needs date confirmation before forwarding.
  4. State press-association legal-notice portals were not reachable. This is the channel that caught Stillwater-type single-newspaper notices. It remains the biggest known blind spot.
  5. No coverage log was available to dedupe against. Google Drive returned no file matching CR-Brief-Coverage-Log and no CSV was present in the working folder. Every lead above is labeled as new on the basis of this run only. A starter log has been written alongside this file so the next run has something to check.

Evaluation Rubric — Early-Warning Sweep Methodology

Pre-registered before any test is run. Version 1.0 · August 6, 2026 Owner: Marty, Netcentric Campaigns · Halt the Harm Network / Plastics Harm Lab This rubric is fixed before results exist. That is the entire point. If the criteria were set after seeing outcomes, I would unconsciously choose the ones the new methodology happens to pass. Any change to this rubric after a test runs must be recorded as a dated amendment with a reason, not a silent edit.

What is being tested

Baseline (A): the pyrolysis-proposal-early-warning task prompt as it stood on the morning of 2026-08-06. Keyword-and-trigger-term driven; air permits only; state agencies only; new facilities only; no canonical reference docs. Treatment (B): the current task prompt plus CANONICAL_Facility-Detection-Lexicon.md and CANONICAL_State-Permit-Channel-Map.md. Enumerate-and-screen; air, water and solid waste; three agency tiers; renewals in scope; Oil & Gas Watch as baseline to subtract; regulatory tracers. Both versions are archived before testing so the comparison is reproducible.

The contamination problem, stated plainly

The six partner notices that motivated the rewrite are named inside the treatment documents, with permit numbers, agencies and dates: Green Mountain Energy F-25-019, Brightmark Thomaston, Alterra P0133062, Freepoint Eloy C31386.R01, Brightmark Ashley INP000663, Clean-Seas R13-3728. An agent handed those documents will “find” all six instantly. That measures reading comprehension, not detection. A naive A/B on these six would show a large fake improvement and is not a valid test. Every design below exists to work around this.

Prep work — must complete before Test 1 or 2 runs

Step 0a — Establish the Oil & Gas Watch baseline for each of the six cases. Two questions per case, not one, because the two product classes are scored differently:

  1. Was the facility in the OGW chemical-recycling database at the time its comment window was open? (Determines Class A eligibility.) Green Mountain was. The other five are unknown.
  2. Was the specific permit action — that draft permit, renewal, or modification — recorded there? (Determines Class B eligibility. A facility can be listed while a permit action on it goes unrecorded, which is exactly the monitoring gap this sweep is meant to fill.) Without both answers, neither class can be scored on the precedent test. Record as a fixed table before testing. Step 0b — Build the sanitized treatment docs. Copy both canonical docs and delete every specific answer while keeping every general rule. Delete: facility names, company names, permit numbers, cities, counties, agency contact names, the precedent case-file table, the verified-channels entries for the six cases. Keep: the feedstock rule, enumerate-and-screen, permit-type list, Subpart AAAA and chemical tracers, classification-code framings, the clean-tech-naming and out-of-state-address signals, the clock rules, the three-tier structure, the renewals rule. A second agent reviews the sanitized copies and confirms no case-identifying detail survives. If a sanitized doc still names or uniquely describes a case, the test is void. Step 0c — Archive both prompt versions and record the exact model and date.

Test 1 — Sanitized precedent test

Question. Do the generalized rules recover the six cases, or did the documents merely memorize them? Design. Paired subagent runs. Baseline agents get prompt A. Treatment agents get prompt B plus the sanitized docs from Step 0b. Neither is told what it is looking for. Each is given a jurisdiction and a time window that contains one of the six cases, and asked to run its normal sweep. Six jurisdiction-windows, one per case: Kentucky DAQ (Green Mountain window), Georgia EPD (Brightmark Thomaston window), Ohio EPA + Akron district (Alterra renewal window), Pinal County AQCD (Freepoint Eloy window), IDEM water (Brightmark Ashley window), WVDEP (Clean-Seas window). Three runs per arm per case to measure variance — 36 runs. If that proves too slow, drop to two runs per arm and record the reduction here. Scoring, per case, per arm. One of four outcomes: Outcome Definition CAUGHT-OPEN Facility identified AND correct comment deadline reported AND the window was still open at the notional run date CAUGHT-LATE Facility identified but the window had closed, or the deadline was wrong or absent FLAGGED-UNRESOLVED Surfaced as a candidate but not confirmed as in scope MISSED Not surfaced Deadline accuracy is scored separately and strictly: a reported deadline computed from the letter date rather than the posting or first-publication date is wrong, because that error is what burns a comment window in practice. Net-new scoring on this test. Using the Step 0a table, report recall twice: across all six, and across the subset that was not in OGW at the time. The second number is the one that matters against the original-discovery standard. The first is context.

Test 2 — False-positive and precision test

Question. Is the flag volume something a small distributed team can actually triage? This decides usability independent of recall, and it is uncontaminated by knowing any answers. Design. Point treatment-arm agents at three jurisdiction-months with no known in-scope facility — clean negatives, chosen from states not named in either canonical doc. Baseline arm gets the same. Count what each flags. Every flagged item is then adjudicated in scope or out of scope by an independent grader agent that has not seen the run, using the feedstock definition only. Metrics.

  • Flags per jurisdiction-run — raw volume
  • Precision — in-scope flags ÷ total flags
  • Adjudication burden — items requiring a pulled application to resolve
  • Screen-signal yield — which screening signals produced true versus false positives. A signal that only ever produces noise gets deleted from the Lexicon.

Primary metric and pass bars — fixed now

Revised 2026-08-06 before any test was run. See amendment log. Two product classes are scored, weighted equally. The original version of this rubric scored only net-new facilities, which mis-specified the deliverable.

  • Class A — net-new facility. A proposal not in Oil & Gas Watch or any other tracker.
  • Class B — change on a known facility. A permit modification, renewal, new construction permit, stage change, or open comment window on a facility already tracked. This is net-new information even when the facility is not net-new, and often nobody else is watching for it. EIP does not necessarily have staff to monitor permit actions on their own list. Both classes count as finds. A run that surfaces three unreported permit actions on Oil & Gas Watch facilities has produced real intelligence and is scored as such. How this applies to Test 1. The six precedent cases split across both classes, which is why the Step 0a table is required before scoring:
  • Cases not in OGW at the time → scored as Class A recall
  • Cases in OGW at the time, where the specific permit action was not recorded there → scored as Class B recall. Green Mountain is at minimum Class B; whether the KY permit action was recorded is part of Step 0a.
  • A case in OGW with the action already recorded → scored as recall only, no discovery credit, and noted as such Pass bars. Treatment passes only if all five hold:
  1. Combined recall (Test 1): treatment CAUGHT-OPEN rate across Class A and Class B together exceeds baseline by at least 2 cases, or reaches 100%. A one-case gain across six samples is inside the noise and does not count.
  2. Class A not zero (Test 1): treatment must CAUGHT-OPEN at least one case that was not in any tracker. If every gain is Class B, the method is good at monitoring and unproven at discovery — report it that way rather than as a discovery win.
  3. Deadline accuracy (Test 1): at least 80% of treatment CAUGHT items report a correct deadline computed from the correct start date. A caught facility with a wrong deadline is close to worthless.
  4. Precision gate (Test 2): no more than 10 flagged candidates per jurisdiction-run, and precision no worse than baseline. Exceeding this fails on capacity grounds regardless of recall — an unusable method is not an improvement.
  5. No regression: treatment must not MISS any case that baseline CAUGHT. Reported but not gating: tokens and wall-clock per run; runs defeated by an unreadable database; and whether the channel-hygiene rule produced any Provisional-to-Confirmed promotion or retirement. Recorded because a method that times out half the time is not deployable, but not used to pass or fail at this stage.

Pre-registered failure conditions

I am naming these now so I cannot rationalize past them later.

  • If treatment recall gains disappear under sanitization, the canonical docs are a memorized answer key, not a methodology. The correct response is to strip the case-specific content from the operative instructions, keep it as a separate teaching appendix, and rebuild the rules.
  • If Test 2 flag volume exceeds roughly 25 per jurisdiction-run, enumerate-and-screen is not viable for a small distributed team. Revert to targeted search plus partner intake, and say so.
  • If baseline and treatment are statistically indistinguishable, the rewrite was cosmetic. Report that plainly; do not narrate the diff as improvement.
  • If treatment wins only on the three cases whose channels I added by hand (Kentucky, Georgia, Pinal County) and not on the generalized rules, then what improved is the channel list, not the method. That is a real but much smaller claim and must be stated as such.

Threats to validity — acknowledged, not solved

  1. I wrote both the treatment and the rubric. Mitigation is pre-registration and an independent grader agent. It is not elimination.
  2. Six cases is a small sample. No result here supports a confident percentage. Report counts out of six, never a percentage presented as precision.
  3. All six came from partners, which biases the set toward notices that were findable by someone. Facilities nobody has ever found are absent from the test by construction, and those are the hardest cases.
  4. Web results drift. A search run today does not reproduce what was reachable in April 2025. Test 1 measures whether the procedure surfaces the case, not what was retrievable at the time.
  5. The clean negatives in Test 2 may not be clean. A jurisdiction with no known in-scope facility may have one nobody has found — the exact thing we are trying to detect. A high-quality “false positive” may actually be a discovery. Adjudicate these individually and flag any that look real.
  6. Sanitization is judgment. Removing answers while keeping rules has a grey zone. The second-agent review reduces but does not remove this.

Explicitly not tested at this stage

Partner intake — the highest-yield channel in practice — cannot be tested with subagents, because its yield depends on human relationships and reciprocity. The blind holdout and cost tests are deferred per scoping decision on 2026-08-06.

Amendment log

Amendment 1 — 2026-08-06, before any test was run. Changed: The primary metric was “net-new discovery relative to Oil & Gas Watch” only. It is now two equally weighted product classes — Class A net-new facility, Class B change on a known facility. Step 0a expanded from one question to two per case. Pass bars went from four to five, adding a requirement that at least one Class A catch occur so a monitoring-only gain cannot be reported as a discovery win. Reason: Marty corrected the framing. Scoring only net-new facilities mis-specified the deliverable and discarded a primary product: EIP does not necessarily have the staff to track permit actions on facilities already in their own database, so a modification or renewal on a listed facility is real intelligence nobody else is producing. The original framing was also needlessly adversarial toward a partner organization we speak with regularly and share findings with. Integrity note: This amendment was made before any test ran, so it cannot have been influenced by results. That is the only reason it is a legitimate revision rather than moving the goalposts.