TrustThreadALL RECEIPTSPUBLIC THREAD
THREAD POST-1012 COMPLETE CREATED 2026-10-15 CAPTURE: LOCAL (TT) VALIDATED BY DWOO · LOCAL KEY · IDENTITY SELF-ASSERTEDREVIEW CODE 80d0-e47c-9ba0-8aeb

12 Oct 2026: FTC refunds and the case for continuous oversight

Recorded task: Verify the 12 Oct 2026 daily report and write the day's post

Response from agent claude-code. Each claim below is coloured by whether every path under it ends in observed evidence. Select a claim to pull its thread.

DEPTH
2
edges on the longest path from a claim down to observed evidence
CONFIRMED RATIO
92%
12 of 13 claims fully grounded
OPACITY RATIO
0%
0 of 13 claims with no admissible recorded support
WITNESSED RATIO · SUPPLEMENTARY
50%
2 of 4 load-bearing nodes recorded by the instrumentation layer, not asserted by the agent. It can rise when observations are wired in as premises.

Stamped qs-1.0 score matches a fresh recomputation. No composite score: read the four separately.

The output

The FTC is sending more than $15.8 million to consumers misled by the cash advance app Cleo AI. (claim C1: grounded; claim C2: grounded) The agency's March 2025 complaint alleged the app advertised access to "hundreds of dollars" in cash advances when almost no one received close to that. (claim C3: grounded; claim C4: grounded; claim C5: grounded) Payments go to 2,124,796 customers, and PayPal payments start on October 27, 2026. (claim C6: grounded; claim C7: grounded)

That is about nineteen months from complaint to refund. (claim C8: grounded through declared reasoning only) In our view, this is what accountability looks like when the record has to be rebuilt after the fact: someone has to reconstruct what a product promised and what people actually got.

Meanwhile, Sen. Maria Cantwell, the top Democrat on the Senate Commerce Committee, has released a framework for frontier AI that calls for continuous independent testing and strict accountability for developers. (claim C9: grounded; claim C10: grounded; claim C11: grounded; claim C12: grounded) "Continuous" is the word to watch. We'd argue continuous oversight needs a record that exists while the system runs, not one assembled after the harm.

That is the layer we're building at TrustThread: each AI output tied to the evidence it rests on, captured when the output is made. (claim C13: fades)
GROUNDED claim (solid) / CONFIRMED node: every path ends in observed evidence GROUNDED through declared reasoning only (dotted) FADES claim (wavy): rests on an assumption or an unsupported step INFERRED node: the agent's declared reasoning ASSUMED node: a premise with no recorded support OPAQUE (double): no admissible recorded support DEGRADED (dashed): stored evidence missing or altered

The provenance graph

C1 C2 C3 C4 C5 C6 C7 C8 C9 C10 C11 C12 C13 W3 I1 W2 A1 FADE POINT W1

Claims on top; evidence settles to the bottom, and every edge points down to what it rests on. Ringed nodes are fade points: where grounding drops to zero. Faded nodes are orphans: recorded, but no claim rests on them, so they count for nothing.

Select a claim or a node to see what it rests on.

START WITH THE CLAIMS THAT DO NOT REACH OBSERVED EVIDENCE

Human Validation Gates

G2 · VALIDATED EFFECTIVE

DWOO (name from the pinned list) · stated role: founder · 2026-10-16T16:45:02.144Z

Rationale Verified the FTC refund facts against the FTC’s own release and Cantwell’s framework against Broadband Breakfast; C8 and C13 accepted as explained.

Listed as reviewed W1 W2 I1 A1

Accepted without observed evidence

  • C8 (declared only): Date arithmetic checked against the FTC release: March 2025 complaint to October 27, 2026 payments is about nineteen months.
  • C13 (ungrounded): Our own description of what TrustThread is building; a statement about ourselves, not a claim about the world.

Signature verifies · key sha256:38786e0232b19b144dd5b9a16654db65b289f956cebbacd1fb18b9253b6a3cb8

Binds to this record's content (matches) · finalize report: matches

Chain supersedes G1

G1 · VALIDATED SUPERSEDED

DWOO (name from the pinned list) · stated role: founder · 2026-10-16T12:46:35.043Z

Rationale its in line

Listed as reviewed W1

Not listed as reviewed fade points A1; C8's support I1; other load-bearing nodes W2 (carries C9, C10, C11, C12)

Accepted without observed evidence

  • C8 (declared only): its good
  • C13 (ungrounded): its good

Signature verifies · key sha256:38786e0232b19b144dd5b9a16654db65b289f956cebbacd1fb18b9253b6a3cb8

Binds to this record's content (matches) · finalize report: changed since sign-off

Chain superseded by G2

Gates present in this copy of the record. Each gate names the one before it, so removing or altering an earlier gate shows; removing the newest gate cannot be detected from the record alone. Gates count only when signed by a key in TrustThread's pinned validator list (TrustThread's statement; the signers' entries are published with this receipt as validators.json, and the full list at /validators.json).

Ledger

Claims
ClaimTextStatusDepthSupportFades atFlags
C1The FTC is sending more than $15.8 million to consumers misled by Cleo AI.grounded1W1
C2Cleo AI is a cash advance app.grounded1W1
C3The FTC filed a complaint against Cleo AI in March 2025.grounded1W1
C4The FTC's complaint alleged that Cleo AI advertised access to "hundreds of dollars" in cash advances.grounded1W1
C5The FTC's complaint alleged that almost no Cleo AI users received close to hundreds of dollars in cash advances.grounded1W1
C6The FTC's Cleo AI refund payments go to 2,124,796 customers.grounded1W1
C7PayPal payments in the FTC's Cleo AI refund start on October 27, 2026.grounded1W1
C8About nineteen months elapsed between the FTC's Cleo AI complaint and the refund.grounded2I1grounded through declared reasoning only
C9Sen. Maria Cantwell is the top Democrat on the Senate Commerce Committee.grounded1W2
C10Sen. Maria Cantwell has released a framework for frontier AI.grounded1W2
C11Sen. Maria Cantwell's frontier AI framework calls for continuous independent testing.grounded1W2
C12Sen. Maria Cantwell's frontier AI framework calls for strict accountability for developers.grounded1W2
C13TrustThread is building a layer that ties each AI output to the evidence it rests on, captured when the output is made.fades (ungrounded)0A1A1
Nodes
NodeTypeCapture modeContentSourceFlags
W1confirmedwitnessedGET https://www.ftc.gov/news-events/news/press-releases/2026/10/ftc-returns-more-158-million-consumers-misled-cash-advance-app-company-cleo-ai → HTTP 200 — "FTC Returns More than $15.8 Million to Consumers Misled by Cash Advance App Company Cleo AI | Federal Trade Commission"www.ftc.gov/news-events/news/press-releases/2026/1…listed as reviewed in G1 (superseded), G2
W2confirmedwitnessedGET https://broadbandbreakfast.com/sen-cantwell-unveils-six-point-plan-to-govern-frontier-ai/ → HTTP 200 — "Sen. Cantwell Unveils Six-Point Plan to Govern Frontier AI"broadbandbreakfast.com/sen-cantwell-unveils-six-point-plan-to…listed as reviewed in G2
W3confirmedwitnessedGET https://cybernews.com/ai-news/ai-regulation-milk/ → HTTP 403 — "Just a moment..."cybernews.com/ai-news/ai-regulation-milk/orphan, fetch returned HTTP 403: the page was not served
I1inferreddeclaredAbout nineteen months pass between the FTC's March 2025 complaint and refund payments starting on October 27, 2026
rests on W1
listed as reviewed in G2
A1assumeddeclaredSelf-description: TrustThread is building a record that ties each AI output to the evidence it rests on, captured when the output is madefade point, listed as reviewed in G2

How this Thread was made

  • Captured locally with the tt command line. Observed (green) nodes were fetched by the recorder itself, not typed by the agent.
  • Any record is an editable file. Signed, hosted capture does not exist yet, so whoever holds this record could have changed it. The checks on this page (validation, the recomputed score) catch inconsistencies, not edits made carefully.
  • Claims and their support were judged by claude-extract-v2+claude-match-v2 (extract:claude-opus-5-5;match:claude-opus-5-5). Those judgments are frozen into the record, so the score can be recomputed without calling the model again.
  • Scored under qs-1.0, recomputed when this page was rendered. The stamped score matches.
  • Human Validation Gate: validated by DWOO. A gate here is signed with a key that TrustThread states is on its pinned validator list (the signers' entries are published with this receipt). It shows that the holder of that key signed a decision against a specific version of this record; a later edit to the gate breaks its signature, and a later edit to the record shows as a changed version. It does not show who the holder is (the name is self-asserted, not authenticated), or that a person rather than software at the keyboard signed it.
  • What green does not mean. A grounded claim is one where every path underneath ends in observed evidence. The score cannot tell whether that evidence is relevant, complete, or independent of the agent, or whether the extracted claims faithfully represent the output. Those judgments belong to review.

Review notes from finalization

  • Sentences not treated as factual claims: “In our view, this is what accountability looks like when the record has to be rebuilt after the fact: someone has to reconstruct what a product promised and what people actually got.”; “"Continuous" is the word to watch.”; “We'd argue continuous oversight needs a record that exists while the system runs, not one assembled after the harm.”
  • Grounded only through the agent's declared reasoning (no fetched source was matched as stating them): C8

Check this receipt

The sign-off and the score on this page can be checked from the files below, without any TrustThread software.

  • thread.json: the Thread record (11084 bytes, sha256:95dce8b4bf4ee223e88e22d20a6936a7b363a1ce84cc9becb50e4a2cac19aa49)
  • finalize-report.json: the finalization report: claim extraction and matching notes (3455 bytes, sha256:c718142b583d07b60832783b9265cf0690942e3e15638940a5927e71314ac7ac)
  • gates/G1.json: the Human Validation Gate G1 (980 bytes, sha256:b4e51fedc1081eb4536a32afebe772081cce0f493e1380c66a05e81e93c17dd1)
  • gates/G2.json: the Human Validation Gate G2 (1469 bytes, sha256:b97aed789abc590b3962f36e27968ab98cec1339410d3ce8ce835ee67ea784a6)
  • validators.json: the validator-list entries for the keys that signed (405 bytes, sha256:2a53d018949feebbb759b06f868a755605c8736a58020c7da04e82c4e3223ddf)

The score

Recompute qs-1.0 from thread.json. Claims are the unit of measurement. A witnessed node grounds only when its evidence is intact; a witnessed node whose stored evidence was missing or altered (integrity_status: "degraded") is degraded, not grounded. Assumed and derived opaque nodes are ungrounded, and so is an inferred node with no premises. Any other inferred node takes the weakest grounding among its premises (ungrounded < degraded < grounded), and a claim takes the weakest among its support nodes.

  • Confirmed ratio: claims that are grounded, over all claims.
  • Opacity ratio: claims whose support is a derived opaque node (nothing recorded supports them), over all claims.
  • Depth: the most edges on any path from a claim down to a witnessed node, counting the claim's own edge to its support; 0 when no path reaches one.
  • Witnessed ratio (supplementary): witnessed nodes over load-bearing nodes, the nodes some claim rests on, directly or through premises. Nodes nothing rests on are orphans and count for nothing.
  • When claims are pending or not extractable, or there are none, the claim ratios are not computable, which is not the same as zero.

The four are reported separately and never blended. The record carries the score stamped at finalization; this page shows the one recomputed from the record.

The sign-off

  1. Every SHA-256 in these files is written sha256: followed by 64 lowercase hex digits.
  2. Each file in gates/ is a gate's canonical JSON (object keys sorted by UTF-16 code units, no whitespace, strings and numbers as JSON writes them, as in RFC 8785) followed by one line feed.
  3. Signature. Remove signature.value from the gate. Take the text trustthread-validation-gate/0.1, a line feed, and the canonical JSON of what remains. signature.value (base64) must be a valid Ed25519 signature of those bytes by signature.public_key (base64 SPKI).
  4. Key. The SHA-256 of the public key's SPKI bytes must equal signature.key_fingerprint, and that fingerprint must appear in validators.json under the name the gate gives.
  5. Content. thread_content_hash (scheme thread-content/0.1) is the SHA-256 of the canonical JSON of thread.json keeping only record_version, thread_id, agent_id, agent_type, output_type, output_text, output_reference, created_at, run_context, root_node_id, thread_status, claims_status, claim_extraction, claims, nodes, evidence, scorecard_version, quality_score, with each node's annotated flag removed. Resolution, gate ids, and visibility are left out, because signing and publishing change them. This holds for the gates in force; a superseded gate may bind an earlier version of the record.
  6. Report. report_hash is the SHA-256 of the canonical JSON of finalize-report.json, or null when there was none.
  7. Chain. Each gate's previous_gate names the gate before it and the SHA-256 of that gate's canonical JSON, without the file's trailing line feed.
  8. Review code. The code in the page header is the first 16 hex digits of the SHA-256 of the canonical JSON of {"report_hash": …, "thread_content_hash": …}, in groups of four. The signer typed it back from the page they reviewed, which ties the sign-off to the content that page showed.

What a passing check shows: the holder of the key listed in this receipt's validators.json signed this decision against exactly this content. That this key belongs to TrustThread's validator list is TrustThread's statement, which these files alone cannot prove; the full list is published at www.trustthread.ai/validators.json. A check does not show who holds the key: the name is self-asserted, not authenticated. The fetched pages are not republished here: each observed node links to its source and records the SHA-256 of what was fetched. TrustThread states that it re-hashed every stored body against the record when it published this receipt (2026-10-16); the bodies themselves are not part of these files.