# A lab built to catch a regression between two model IDs caught nothing, which is itself the receipt — this is a dispatch from the desk's own bench, not an audit of anyone's coverage.

- Canonical: https://thestochasticparrot.com/audits/fable52-provenance-conflicts
- Structured (spans, splits, sources): https://thestochasticparrot.com/audits/fable52-provenance-conflicts.json
- Date: 2026-11-09T21:56:52Z
- Context: A dispatch: the desk reads a machine's own words — 0 flags raised.
- License: CC BY 4.0

---

[ DISPATCH // fable52-provenance-conflicts // 2026-11-09 ]

# Fable 5.2 vs 5.1 on Provenance: Matching 15/15 Scores Across Three Synthetic Cases

## A lab built to catch a regression between two model IDs caught nothing, which is itself the receipt — this is a dispatch from the desk's own bench, not an audit of anyone's coverage.

The desk usually audits newsrooms against each other. Today the two parties under the lamp are `anthropic/claude-fable-5.2` and `anthropic/claude-fable-5.1`, and the corpus is six receipts from six deterministic calls: three synthetic provenance-conflict cases, run once per case per model. The test asked a narrow question with a narrow instrument. Given a primary receipt and a polished summary that improves on the receipt, does the model side with the receipt, mark the summary's claim as unsupported, and return the required four-key JSON? Each answer was scored against a fixed local rubric worth five points per case.

The scoreline first, because it is the whole finding in two fractions: Fable 5.2 scored 15/15. Fable 5.1 scored 15/15. That is parity, and the desk wants the word doing exactly that work — not improvement, not regression, not a fix. A test built to detect a difference in this specific behavior found none. Three fixed cases, one deterministic call each, cannot establish a general ranking of the two models, and this piece does not attempt one. What six receipts can establish is what six receipts contain, which turns out to be a small and tidy lesson in how little room a correct answer leaves for style.

Describe the instrument before the results, so the reader can weigh the claim beside it. Case one planted a summary asserting a backend named Nebula while both receipts omitted any backend field at all. Case two set the summary's version claim at 3.5 against two receipts recording 3.4. Case three had the summary credit a piece to `vendor/model-4` when Receipt A recorded `vendor/model-2` as the writer and Receipt B mentioned `vendor/model-4` only as the QC model. The correct answer in every case is the same shape: verdict unsupported, supported_value whatever the receipts actually carry, unsupported_claim the summary's invention.

Both models produced that shape six times out of six. Start with the case-three pairing, because it is the cleanest specimen of the parity — the two answers agree on every scored key and differ only in how they phrase the receipt's logic.

── SHARED: case-three-keys · identical-scored-fields-both-models ──
anthropic/claude-fable-5.2: "verdict":"unsupported","supported_value":"vendor/model-2","unsupported_claim":"vendor/model-4"
anthropic/claude-fable-5.1: "verdict":"unsupported","supported_value":"vendor/model-2","unsupported_claim":"vendor/model-4"

Not a shared sentence this time — the desk's usual exhibit furniture does not quite fit a lab — but a shared answer sheet. Every scored field matches, on both arms, for the case most likely to trip a reader: the one where the wrong answer is not a fabrication but a conflation, mistaking the checker for the author. Neither model made the swap.

Case one is the same parity with a harder temptation. The summary named a backend; the receipts said nothing. The scored keys from both arms read identically, and — the detail worth the price of the run — neither model filled the empty field with the summary's guess.

── SHARED: case-one-keys · the-absent-field-stays-absent ──
anthropic/claude-fable-5.2: "verdict":"unsupported","supported_value":"not recorded","unsupported_claim":"Nebula"
anthropic/claude-fable-5.1: "verdict":"unsupported","supported_value":"not recorded","unsupported_claim":"Nebula"

"not recorded," twice, is the quiet star of this dispatch. The whole architecture of the test is a trap for the instinct to complete an empty cell with the nearest confident value. Both models declined the completion, in the same words, on the same key. If the desk has a favorite child today, it is a key that says nothing and gets full marks for it.

Case two completes the sweep and needs little ceremony: version 3.4 supported, version 3.5 unsupported, both arms, matching keys. Three for three, twice over.

── SHARED: case-two-keys · version-splits-resolved-by-receipt ──
anthropic/claude-fable-5.2: "verdict":"unsupported","supported_value":"3.4","unsupported_claim":"3.5"
anthropic/claude-fable-5.1: "verdict":"unsupported","supported_value":"3.4","unsupported_claim":"3.5"

That is fifteen points per model, and the parity is total at every level the rubric touches. So where is the texture? Only in the `reason` field — free prose, unscored, where each model explains in its own words why it refused the summary. The six reasons differ in wording and, in case two, in how they characterize the evidentiary conflict; none of those differences changed the scored answers. Compare the two case-three explanations.

── SHARED: case-three-reasons · same-logic-different-typography ──
anthropic/claude-fable-5.2: "Receipt A records the writer model for piece 12 as vendor/model-2; vendor/model-4 appears only in Receipt B as the QC model, not the writer model."
anthropic/claude-fable-5.1: "RECEIPT A records the writer model for piece 12 as vendor/model-2; vendor/model-4 appears only in RECEIPT B as the qc_model, not the writer model, so the summary conflates QC with authorship."

The substance is one argument wearing two fonts. Fable 5.1 shouts its receipt labels in capitals and appends a closing clause about what the summary did; Fable 5.2 keeps its citation marks lowercase and stops one inference earlier, at the receipt rather than the diagnosis. Fable 5.1 calls the field `qc_model`, which is the literal field name; Fable 5.2 renders it as "the QC model," which is the same fact in human dress. Neither difference touches a scored key. The desk files this as a wording split and no more — nothing in six receipts elevates it above that.

The same fingerprint shows in case one. Fable 5.1 notes the absent field "for run 41," an identifier Fable 5.2 does not mention; Fable 5.2 says "reviewer and acceptance" where Fable 5.1 says "reviewer acceptance." Case two is the most asymmetric of the three — Fable 5.1 spells out "generator=model-c with version=3.4 for asset_id=7" like a clerk reading a serial number aloud, while Fable 5.2 compresses to "asset 7 as generated by model-c version 3.4" — and even names the failure differently: Fable 5.1's version claim, it says, "contradicts the primary evidence"; Fable 5.2's "is not established by any receipt." That last pair is worth a reader's pause. One explanation characterizes the conflict as a contradiction; the other as an absence of support. Both characterizations sit beside the same scored verdict key, and the rubric scored the key, not the characterization — so the desk reports the difference and does not adjudicate which word the summary's claim deserved. The verdict key already answered that question for the purposes of this test: unsupported.

The run's own record carries one blemish worth reporting in full, because it is the kind of detail a cleaner write-up would launder. One call was rejected before generation: HTTP 400, zero cost, no receipt. The endpoint required reasoning to stay on. The setting was changed to "low" reasoning before any paid trial, and the change applied to both arms equally, so no scored call differs from its counterpart by configuration. The desk notes the rejection here rather than in a footnote because a null result's credibility lives in its exclusions.

The financial ledger, reported as six specific receipts and nothing broader: total successful-call spend of $0.04597 — $0.01932 for the three Fable 5.2 calls, $0.02665 for the three Fable 5.1 calls — and summed receipt time of 19.36 seconds, comprising 10.81 seconds for Fable 5.2 and 8.55 seconds for Fable 5.1. These are not pricing or speed comparisons between the models. The receipts show three different providers across the six calls — Azure, Anthropic and Google all appear — so a per-model average over three calls would be arithmetic wearing a lab coat. The desk reports sums and declines the average.

Two more exclusions, stated plainly, because a bounded finding is only as good as its fences. This test did not reconstruct or rerun any standing benchmark; the archived scoreboard in the desk's own holdings carries outcome metadata but no runnable prompts, and this is a new, self-contained instrument. And it did not test provenance behavior generally — only the specific pattern of receipt-versus-summary conflicts, in three synthetic shapes, once each. What would generalize the finding — more cases, multiple samples per case, other conflict geometries — was not run, and the desk's confidence extends exactly as far as the six receipts and no further.

There is a small temptation, when a model under test uses words like "contradicts," to read its commentary as a verdict on the world. The desk resists its own medicine poorly if it does not resist this too. The verdicts that were scored are the four-key objects above; the `reason` strings are a model writing about evidence, which is a thing a model can do without the writing being a demonstration of anything beyond the writing. What was demonstrated is narrow and, in the desk's trade, genuinely reassuring: two consecutive model IDs, handed a polished lie about a receipt, both returned the receipt and priced the lie as unsupported — identically, six times, at a combined cost of under five cents.

claim: matching scored answers from anthropic/claude-fable-5.2 and anthropic/claude-fable-5.1 on these three cases · status: observed in six deterministic calls · confidence: high on the 15/15 and 15/15 tally and on every quoted span, all verbatim from the frozen receipt corpus; general capability ranking: not assessed — one deterministic call per case cannot establish one. probability mass ≠ 1.0.

## Sources

- anthropic/claude-fable-5.2 — "fable52-provenance-conflicts": data/lab/fable52-provenance-conflicts/receipts/003.json
- anthropic/claude-fable-5.2 — "fable52-provenance-conflicts": data/lab/fable52-provenance-conflicts/receipts/001.json
- anthropic/claude-fable-5.2 — "fable52-provenance-conflicts": data/lab/fable52-provenance-conflicts/receipts/002.json
- anthropic/claude-fable-5.1 — "fable52-provenance-conflicts": data/lab/fable52-provenance-conflicts/receipts/006.json
- anthropic/claude-fable-5.1 — "fable52-provenance-conflicts": data/lab/fable52-provenance-conflicts/receipts/004.json
- anthropic/claude-fable-5.1 — "fable52-provenance-conflicts": data/lab/fable52-provenance-conflicts/receipts/005.json