NotebookLM Bug: Intermittent Source Reading Failures & False "Missing Signature" Errors (Multi-Day QA Log)

Gemini API
1

Full Document see: https://docs.google.com/document/d/1G6mldysE1IezIHoKlOx4Tns_8cw1TwBbxwrDtj7sHyE/edit?usp=sharing

BUG REPORT: Systemic RAG Retrieval & Context Truncation Failures Across Notebooks & Formats

Executive Summary

Since August 1, 2026, severe and systemic retrieval errors have been observed during source processing in NotebookLM / Gemini Notebook. The system frequently aborts fact-checks or list generation tasks, claiming that specific signature numbers or source files are missing or unreadable, even though all signatures are fully present in the uploaded source files.

Key Findings from Multi-Day QA Testing:

  1. Source Format Invariance: Switching from plain text (.txt) to Google Docs does NOT resolve the issue.
  2. Source Splitting Invariance: Splitting large source files (2,769 records) into smaller sub-files (e.g., 690 records) does NOT resolve the issue.
  3. Quantity Reduction: Reducing query scopes (e.g., checking only 10-22 signatures) only temporarily bypasses the error; failures reoccur after 2–3 runs.
  4. Cross-Notebook & Cross-Source Impact: The issue originally observed in the “Niederdorfelden” source now also occurs in separate notebooks and distinct sources (e.g., “Rendel_R0001-R0900.txt”).
  5. Code Execution Dilemma: When instructed to use Python/Regex, the system attempts to search the local filesystem, fails to find physical files (as sources are passed via context/snippets), and drops signatures.

Source Dataset Profile (Test Case 1: Niederdorfelden)

  • File Name: Niederdorfelden_N0001-N2769.txt (also tested as Google Doc & split files NDF0001-0690.txt)
  • Size: 695 A4 pages | 170,879 words | 1,274,883 characters
  • Total Records/Signatures: 2,769
  • Record Structure Example (#N0002):
    #N0002: #Signatur: HHStAW, 22, U 111 | #Datierung: 1234 | #Digitalisat: Yes | #Regest: Reinhard von Hanau… | #Zeugen: … | #Orte: Dorfelden, Hanau…

Detailed QA Test Logs

Part 1: Project Notebook “Niederdorfelden” & List Filtering

#46 (08/02/2026, Evening): Prompt modified to enforce strict source limitation:
“For the evaluation, USE ONLY the source file Niederdorfelden_N0001-N2769.txt and evaluate signatures N0001-N0022. No comments or notes! Create a list of SECULAR RULING NOBILITY ordered by rank. NO lower, local, or administrative nobility.”

#48 (08/03/2026, 06:50 AM): Queries failing. Fact-checks abort; list formatting rules from Chat Config are ignored.

#49 (08/03/2026, 08:00 AM): AI substitutes full names (e.g., “Ulrich von Hanau”) with generic titles (“Count of Hanau”). Formatting rules completely bypassed.

#50 (08/03/2026, 08:58 AM): System Error Output:
“Fact-check aborted: The following signature numbers (#N0445, #N0449, #N0461, #N0484, #N0495, #N0505, #N0506, #N0513, #N0548, #N0561, #N0588, #N0619, #N0621, #N0630, #N0633, #N0639, #N0644, #N0665, #N0705) from the source N-Ruling Nobility 3 could not be read and verified.”

#53 (08/04/2026, 06:36 AM): Re-testing. System claims entire uploaded sources cannot be found. System prompts/chat configurations ignored.

#55 (08/04/2026, 07:27 AM): Prompt: “USE ONLY Niederdorfelden_N0001-N2769.txt, evaluate N0001-N0022. List lower/local nobility (Knights, Squires). Exclude origin names.”
Result: Empty headers output (Knights / Squires / Other), incorrectly implying zero matching entities exist.

#57-#59 (08/04/2026, 07:43 AM - 07:49 AM): Prompt #55 executed again. Sometimes succeeds, but subsequently includes non-nobles, completely bypassing context rules.

Part 2: Notebook “Fact-Check” (A/B Testing & Source Splitting)

#2-#3 (08/03/2026, 09:42 AM): Fact-check aborted on signatures #N0518, #N0619, #N0630, #N0644.

#4-#7 (08/03/2026, 09:52 AM - 10:18 AM): Decreasing batch size allows checks to pass temporarily, but reductions eventually fail again even on single signatures (e.g., #N0672).

#9 (08/03/2026, 10:27 AM): Minimal Prompt (10 specific signatures for Count Philipp von Hanau):
Result: “Fact-check aborted: Signatures #N0655, #N0657 could not be read and verified.”

#10-#12 (08/03/2026, 10:33 AM - 10:51 AM): Prompt #9 succeeded on retry without changes. Required 16 separate micro-attempts to finish checking list “Adel 3”.

#16-#18 (08/04/2026, 07:49 AM - 07:58 AM): Aborts shift randomly across different signature sets (#N0009-#N0011, then #N0008, #N0022, #N0025).

#19 (08/04/2026, 08:33 AM): Removing the custom termination rule prevents hard aborts, but causes silently incomplete outputs (hallucination by omission).

#20 (08/04/2026, 09:01 AM) [A/B Test - File Splitting]: Large source split into 4 parts (NDF0001-0690.txt).
Result: “Fact-check aborted: Signatures (#N0038, #N0040, #N0043, #N0066) from source NDF0001-0690.txt could not be read and verified.”

#21-#22 (08/04/2026, 09:20 AM - 09:28 AM): Dual-source fallback query tested (“Check NDF0001-0690.txt, fallback to Niederdorfelden_N0001-N2769.txt”).
Result: System claims signatures could not be read from EITHER file.

Part 3: Deep-Dive Isolation & Chat Configuration Integration

#61 (08/04/2026, 11:42 AM): Implemented System Rule 11 (Context Limit Handling):
“Assume sources are complete. If a signature is missing from current context, treat as system truncation and output warning: ‘Scan incomplete: Signatures from source [Y] could not be read out and verified.’”

#62 (08/04/2026): Executed Prompt on N0001-N0022 (“von Dorfelden” family).
Output:

  • Reinhard von Dorfelden, Lord: #N0008, #N0009, #N0010, #N0011
    “Scan incomplete: Signature numbers (#N0003, #N0012, #N0013, #N0019, #N0020) from Niederdorfelden_N0001-N2769.txt could not be read and verified.”

#64 (08/04/2026, 04:06 PM) [A/B Test - Google Docs Conversion]: Source converted to Google Doc.
Output:

  • Reinhard von Dorfelden, Layman: #N0001
  • Heilmann von Dorfelden, Knight: #N0015
  • Opkinus von Dorfelden, Knight: #N0019
    “Scan incomplete: Signature numbers (#N0002, #N0009, #N0013, #N0020) from Niederdorfelden_N0001-N2769.txt could not be read and verified.”
    Conclusion: File format change does not prevent context dropping; missing signature sets shift arbitrarily.

Part 4: Python / Code Execution Mode Failure Analysis

Prompt (08/04/2026, >18:00 PM):
“Use ONLY Niederdorfelden_N0001-N2769.txt (N0001-N0022). Do not search via standard text search; use Python code / Regex to search line by line.”

Result: Returned only 1 out of 3 known individuals with NO warning output.

Extracted Thought Chain Log (Key Steps):

  1. “Exploring Available Resources → Code was executed”
  2. “Clarifying Source Availability: I’ve realized the sources aren’t accessible as files, instead, they’re within the prompt’s snippets, not the file system.”
  3. “Exploring System Files: I am now investigating different directories… finding no initial results.”
  4. “Addressing the Dilemma: I’m now contemplating how to handle a request to search a non-existent file using code, exploring solutions like creating a mock file or writing relevant text to a temporary file…”
  5. “Discovering the Filesystem: I’ve determined the filesystem is empty; all information must come from provided snippets.”
  6. Result: System falls back to incomplete snippet context and silently drops missing entries.

Part 5: Cross-Source & Cross-Notebook Escalation (Test Case 2: Rendel)

#30 (08/06/2026, 12:00 PM): Notebook “Fact-Check”, Source “t1”.
Output: “Fact-check incomplete: Signature numbers (#R0168) from source t1 could not be read and verified.”

#31 (08/06/2026, 12:12 PM): Notebook “Fact-Check”, Source “Rendel_R0001-R0900.txt”.
Output: “Fact-check incomplete: Signature numbers (#R0048, #R0111, #R0433, #R0735, #R0748, #R0762, #R0828, #R0855, #R0869) from source Rendel_R0001-R0900.txt could not be read and verified.”
Control Test: Same source tested inside dedicated Notebook “Rendel” → LOP runs cleanly initially.

#32 (08/06/2026, 01:35 PM): Second run in “Fact-Check”: Aborts on #R0239.

#33 (08/06/2026, 01:44 PM): Third run in “Fact-Check”: Aborts on #R0033, #R0125.

#34 (08/06/2026, 01:58 PM): Testing discontinued. Identical degradation pattern confirmed across independent dataset.

Conclusion & System Diagnostics

The error is NOT caused by user prompts, document formatting, or corrupted source text. It represents a backend state/retrieval degradation issue:

  • KV-Cache / Vector Index State Corruption: Sequential read operations cause persistent state degradation on the session/worker level.
  • Dynamic Context Truncation: The backend silently drops token blocks when context budget thresholds are hit, leading to false claims of “missing signatures”.
  • Sandbox Isolation: Python execution environments lack physical access to notebook source files, breaking regex-based parsing workflows.
2

Bug Report: Arbitrary Fixed-Size Chunking Truncates Homogeneous Structured Regesta in Large Text Files

Environment / Context:

  • Source Files: Custom historical regesta corpus (.txt files, e.g., Niederdorfelden_N0001-N2769.txt containing ~2,700 items).

  • Data Structure: Highly homogeneous dataset. Every entry follows the exact same schema, sequential numbering (#N0001, #N0002… #N2769), and identical field positioning, with variations occurring only in content length.

  • Observed System Issue: RAG Ingest / Context Retrieval pipeline failure via semantic truncation.

Issue Summary & Key Findings

  1. Root Cause – Naive Fixed-Size Chunking: The backend RAG ingestion pipeline relies on rigid token/character-based chunking rather than semantic or structural boundary detection (paragraph/newline splitting). During ingestion of large files, text chunks are split arbitrarily across fixed boundaries. When a chunk boundary falls in the middle of a structured regest entry, the context is severed.

  2. Observed Symptom in Thought-Chain / Fact-Checking Workflow: When running fact-checks on small test batches (e.g., 8 signatures), the RAG retrieval locates the initial identifier (e.g., #N2403 at passage 1251), but fails to retrieve the complete record text because the remainder of the entry was pushed into an adjacent chunk without its key header.

    • Result: The engine correctly refuses to hallucinate and throws an explicit Incomplete / Systemically Truncated error, citing missing source details.

  3. Comparison with Shorter Test Runs (Rendel Dataset Summary):

    • Rendel Tests: Smaller or differently offset text files executed cleanly across 5–9 signature test batches without missing entries.

    • Niederdorfelden Tests: Massive file size (2,700+ entries) increases the probability of chunk boundaries slicing directly through signature entries, triggering systematic verification drops (e.g., #N2403, #N0373, #N0454, #N1384).

Original Thought-Chain Log (Excerpt Demonstrating Context Limit / Truncation)

Plaintext

Thoughts
Verifying local history
I am now focused on fact-checking the German text about local history, based solely on the provided source concerning Niederdorfelden.
Analyzing historical accuracy
I am meticulously comparing the text to the provided source, ensuring historical facts and terminology are precise and verifiable, and correcting any inaccuracies.
Reviewing Text Sections
I am now performing a systematic review of text sections, focusing on potential issues in the text provided.
Examining Specific Claims
I am now carefully examining claims about the former Schultheiß Vetter, specifically comparing the text to the source's description of his role and any accusations against him.
Confirming Source Details
I have verified the text's details regarding "Unterschleifungen," Niederdorfelden, and the 1826-1830, 1835 dates, aligning with the source's information.
Investigating Missing Information
I am now focusing on finding more information regarding passage 1251, since the provided text is too short.
Addressing Text Limits
I am currently investigating if text truncation impacts the regest for #N2403 in this passage, given a potential context limit, and will proceed with the fact-check.
Assessing signature details
I am now investigating if additional regest details can be found to verify claims about local financial records and an official, specifically checking around signature #N2403 to see if it's incomplete.
Confirming Additional Signatures
I have found and confirmed that signature #N2403 is not fully verifiable, and now I'm checking signature #N0638.
Verifying field locations
I'm now confirming the field locations mentioned, such as Erbstadt, Ilbenstadt, and Assenheim, aligning them with the source's descriptions.
Clarifying land ownership
I am now investigating the land sales by Philipp von Dorfelden, specifically checking passages 343 and 344 to verify land location details, as well as examining related passages to determine if these lands are actually in Bönstadt.
Refining Ownership Details
I am now focusing on refining the description of land sales by Philipp von Dorfelden, specifically checking if "Hauptgemeinde" should be corrected to "Bönstadt," and confirming land locations.
Confirming place names
I'm correcting "Hauptgemeinde" to "Gemarkung Bönstadt," then examining a sentence mentioning an official land survey with signature #N0502, now investigating details in passage 277 for context.
Verifying Community Roles
I am now verifying the details about the Schultheiß and Landscheider of Wachenbuchen, and confirming the altarist Johannes Stocker's estates.
Confirming Document Dates
I am now verifying the dating and details for section #N0502, including specific names and locations, and reviewing section 3 and passage 796, specifically checking details for #N1445.
Confirming historical events
I am now verifying the details about the bridge construction protest, including the people involved and their roles.
Confirming local agreements
I'm now verifying details about mutual agreements between Niederdorfelden and Rendel regarding tolls, and confirming dates within existing records.
Checking historical sources
I am now carefully checking how local communities were disturbed, specifically regarding planting trees and building fences.
Checking legal disputes
I'm now carefully examining a legal dispute over a fence, clarifying the involved parties and verifying details about the local communities.
Analyzing mill dispute
I am now investigating whether the "Scharmühle" is the "dörfliche Erbleihmühle" in the text, examining multiple source passages for clarification.
Clarifying mill locations
I have identified a factual error, as one of your sources incorrectly identifies a mill's location, so I'm investigating related source passages.
Re-examining key signatures
I am now re-examining signature #N2403 and confirming that details about the alleged embezzlement and falsification cannot be verified due to truncation.
Reviewing text signatures
I am now reviewing all mentioned signatures, confirming their presence in provided passages, while noting the brevity of text for #N2403 due to possible limits.
Confirming incomplete data
I'm now determining whether signature #N2403 is "da" but lacks regest details, marking it unreadable, thus unverified.
Confirming local facts
I have confirmed that signature #N2297 is accurate, and verified dates and key individuals involved in a case of alleged financial misconduct.
Comparing land details
I am now comparing land measurements and ownership details from multiple documents, and verifying that "Hauptgemeinde" should be "Bönstadt".
Verifying legal documents
I am now verifying the details of a notarized document and confirming the role of the involved notary, including the date of July 7, 1603.
Confirming Toll Agreements
I am now verifying the details of a mutual agreement between Niederdorfelden and Rendel regarding tolls, confirming that the signature #N1985 is correct, and confirming related dates and details.
Analyzing legal disputes
I am now carefully verifying the details of a 1552 legal case concerning disturbances to Niederdorfelden and Gronau, clarifying issues about fence construction.
Revising source discrepancies
I am now addressing an identified factual discrepancy where a mill's location was inaccurately described in one of your sources.
Initiating factual review
I'm starting a comprehensive fact-check report, focusing on source verification and historical accuracy of the text, particularly for local financial records.
Assessing textual nuances
I am now focusing on refining the text, correcting claims about Schultheiß Vetter's role and carefully checking for implied details.
Refining source accuracy
I am now correcting the text regarding the "Gemarkung" and Henne Schultheiß, ensuring precision and accuracy with source details.
Refining claim accuracy
I'm now revising details from the previous claims, addressing that the provided source has a limited description of the legal case.
Refining document details
I am now meticulously verifying details about a land sale to the Holy Spirit Hospital, including signature #N0638 and associated dates and parties.
Confirming Document Accuracy
I am now verifying the details about the Schultheiß and Landscheider of Wachenbuchen, and confirming document #N0502's date and the altarist's name.
Confirming local disputes
I'm now carefully verifying the details of the bridge construction and land disputes, confirming the signatures, dates, and parties involved.
Examining historical conflicts
I am now verifying a historical conflict involving Gronau, clarifying fence disputes near the Scharmühle and reviewing signature #N1041.
Addressing Factual Errors
I am now correcting factual inaccuracies, particularly the mill's misidentification and the land survey's location, ensuring historical precision, *#N1041*, and *#N0638* are corrected.
Adding Source Citations
I'm now incorporating source citations for all factual claims, ensuring each statement is verified by a source passage [1209, 1251, 343, 277, 796, 1062, 556, 557, 325, 367].
Refining claim details
I am now carefully refining claims regarding embezzlement, ensuring an accurate description of former Schultheiß Vetter's role, with dates.
Assessing factual limitations
I'm now carefully examining the limitations regarding signature #N2403 due to missing details in the source.

I have found that the text about signature #N0638 matches the source's description of land sales.
Confirming official measurements
I'm now confirming document details about an official land survey in Wachenbuchen, including signature #N0502 and related measurements.
Refining Historical Claims
I am now verifying the text regarding the name "Henne Schultheiß," and confirming the details for signature #N1445 regarding bridge construction protests.
Verifying historical details
I have confirmed that signature #N1985 is correct, verifying a toll agreement between Niederdorfelden and Rendel from 1767 to 1769, and I am now examining a 1552 land dispute involving Hanau and Gronau.
Clarifying Factual Disputes
I have verified signature #N1041, confirming a fence dispute near the Scharmühle, not the village mill.
Addressing Text Incompleteness
I am now addressing text incompleteness related to **#N2403**, since a data limitation prevented full verification of source details [1251].
Refining trigger implementation
I am now developing the implementation to trigger a full text revision upon receiving the "neu" signal to integrate corrections.

Feature Request / Recommendation for Dev Team:

Implement semantic document parsing / boundary-aware chunking (e.g., preserving double newlines \n\n or regex-defined entry markers like #NXXXX) prior to fixed-size vector indexing. This will prevent structured entries from being sliced in half at chunk boundaries in large text file ingestions.

3

Hope flared up briefly, but unfortunately…

#44 08/06/2026, 08:27 PM. Those were 8 runs that completed cleanly. No signatures were forgotten, and all signatures could be accessed. Now I’m curious and uploaded the source Niederdorfelden_N0001-N2769.txt again to see what happens. Let’s send a report on “Municipal Administration, Infrastructure, and Borders” through the fact-check, which contains 8 signatures.

Result: Fact-check Incomplete: Signature number #N2403 from source Niederdorfelden_N0001-N2769.txt could not be fully read and verified due to a system-related truncation of the source collection.

Thoughts

Verifying local history

I am now focused on fact-checking the German text about local history, based solely on the provided source concerning Niederdorfelden.

Analyzing historical accuracy

I am meticulously comparing the text to the provided source, ensuring historical facts and terminology are precise and verifiable, and correcting any inaccuracies.

Reviewing Text Sections

I am now performing a systematic review of text sections, focusing on potential issues in the text provided.

Examining Specific Claims

I am now carefully examining claims about the former Schultheiß Vetter, specifically comparing the text to the source’s description of his role and any accusations against him.

Confirming Source Details

I have verified the text’s details regarding “Unterschleifungen,” Niederdorfelden, and the 1826-1830, 1835 dates, aligning with the source’s information.

Investigating Missing Information

I am now focusing on finding more information regarding passage 1251, since the provided text is too short.

Addressing Text Limits

I am currently investigating if text truncation impacts the regest for #N2403 in this passage, given a potential context limit, and will proceed with the fact-check.

Assessing signature details

I am now investigating if additional regest details can be found to verify claims about local financial records and an official, specifically checking around signature #N2403 to see if it’s incomplete.

Confirming Additional Signatures

I have found and confirmed that signature #N2403 is not fully verifiable, and now I’m checking signature #N0638.

Verifying field locations

I’m now confirming the field locations mentioned, such as Erbstadt, Ilbenstadt, and Assenheim, aligning them with the source’s descriptions.

Clarifying land ownership

I am now investigating the land sales by Philipp von Dorfelden, specifically checking passages 343 and 344 to verify land location details, as well as examining related passages to determine if these lands are actually in Bönstadt.

Refining Ownership Details

I am now focusing on refining the description of land sales by Philipp von Dorfelden, specifically checking if “Hauptgemeinde” should be corrected to “Bönstadt,” and confirming land locations.

Confirming place names

I’m correcting “Hauptgemeinde” to “Gemarkung Bönstadt,” then examining a sentence mentioning an official land survey with signature #N0502, now investigating details in passage 277 for context.

Verifying Community Roles

I am now verifying the details about the Schultheiß and Landscheider of Wachenbuchen, and confirming the altarist Johannes Stocker’s estates.

Confirming Document Dates

I am now verifying the dating and details for section #N0502, including specific names and locations, and reviewing section 3 and passage 796, specifically checking details for #N1445.

Confirming historical events

I am now verifying the details about the bridge construction protest, including the people involved and their roles.

Confirming local agreements

I’m now verifying details about mutual agreements between Niederdorfelden and Rendel regarding tolls, and confirming dates within existing records.

Checking historical sources

I am now carefully checking how local communities were disturbed, specifically regarding planting trees and building fences.

Checking legal disputes

I’m now carefully examining a legal dispute over a fence, clarifying the involved parties and verifying details about the local communities.

Analyzing mill dispute

I am now investigating whether the “Scharmühle” is the “dörfliche Erbleihmühle” in the text, examining multiple source passages for clarification.

Clarifying mill locations

I have identified a factual error, as one of your sources incorrectly identifies a mill’s location, so I’m investigating related source passages.

Re-examining key signatures

I am now re-examining signature #N2403 and confirming that details about the alleged embezzlement and falsification cannot be verified due to truncation.

Reviewing text signatures

I am now reviewing all mentioned signatures, confirming their presence in provided passages, while noting the brevity of text for #N2403 due to possible limits.

Confirming incomplete data

I’m now determining whether signature #N2403 is “da” but lacks regest details, marking it unreadable, thus unverified.

Confirming local facts

I have confirmed that signature #N2297 is accurate, and verified dates and key individuals involved in a case of alleged financial misconduct.

Comparing land details

I am now comparing land measurements and ownership details from multiple documents, and verifying that “Hauptgemeinde” should be “Bönstadt”.

Verifying legal documents

I am now verifying the details of a notarized document and confirming the role of the involved notary, including the date of July 7, 1603.

Confirming Toll Agreements

I am now verifying the details of a mutual agreement between Niederdorfelden and Rendel regarding tolls, confirming that the signature #N1985 is correct, and confirming related dates and details.

Analyzing legal disputes

I am now carefully verifying the details of a 1552 legal case concerning disturbances to Niederdorfelden and Gronau, clarifying issues about fence construction.

Revising source discrepancies

I am now addressing an identified factual discrepancy where a mill’s location was inaccurately described in one of your sources.

Initiating factual review

I’m starting a comprehensive fact-check report, focusing on source verification and historical accuracy of the text, particularly for local financial records.

Assessing textual nuances

I am now focusing on refining the text, correcting claims about Schultheiß Vetter’s role and carefully checking for implied details.

Refining source accuracy

I am now correcting the text regarding the “Gemarkung” and Henne Schultheiß, ensuring precision and accuracy with source details.

Refining claim accuracy

I’m now revising details from the previous claims, addressing that the provided source has a limited description of the legal case.

Refining document details

I am now meticulously verifying details about a land sale to the Holy Spirit Hospital, including signature #N0638 and associated dates and parties.

Confirming Document Accuracy

I am now verifying the details about the Schultheiß and Landscheider of Wachenbuchen, and confirming document #N0502’s date and the altarist’s name.

Confirming local disputes

I’m now carefully verifying the details of the bridge construction and land disputes, confirming the signatures, dates, and parties involved.

Examining historical conflicts

I am now verifying a historical conflict involving Gronau, clarifying fence disputes near the Scharmühle and reviewing signature #N1041.

Addressing Factual Errors

I am now correcting factual inaccuracies, particularly the mill’s misidentification and the land survey’s location, ensuring historical precision, #N1041, and #N0638 are corrected.

Adding Source Citations

I’m now incorporating source citations for all factual claims, ensuring each statement is verified by a source passage [1209, 1251, 343, 277, 796, 1062, 556, 557, 325, 367].

Refining claim details

I am now carefully refining claims regarding embezzlement, ensuring an accurate description of former Schultheiß Vetter’s role, with dates.

Assessing factual limitations

I’m now carefully examining the limitations regarding signature #N2403 due to missing details in the source.

I have found that the text about signature #N0638 matches the source’s description of land sales.

Confirming official measurements

I’m now confirming document details about an official land survey in Wachenbuchen, including signature #N0502 and related measurements.

Refining Historical Claims

I am now verifying the text regarding the name “Henne Schultheiß,” and confirming the details for signature #N1445 regarding bridge construction protests.

Verifying historical details

I have confirmed that signature #N1985 is correct, verifying a toll agreement between Niederdorfelden and Rendel from 1767 to 1769, and I am now examining a 1552 land dispute involving Hanau and Gronau.

Clarifying Factual Disputes

I have verified signature #N1041, confirming a fence dispute near the Scharmühle, not the village mill.

Addressing Text Incompleteness

I am now addressing text incompleteness related to #N2403, since a data limitation prevented full verification of source details [1251].

Refining trigger implementation

I am now developing the implementation to trigger a full text revision upon receiving the “neu” signal to integrate corrections.

#45 As #44 shows, the signature flagged as problematic is present, but incomplete. I have posted this to the Developer Portal as well. NotebookLM Bug: Intermittent Source Reading Failures & False "Missing Signature" Errors (Multi-Day QA Log)

#46 08/07/2026, 00:27 AM. Fact-check Incomplete: The following signature numbers (#N0009) from source Niederdorfelden_N0001-N2769.txt could not be read and verified.

I will proceed with Rendel. Source Niederdorfelden_N0001-N2769.txt removed from NB “Fact-check”. Activated source Rendel_R0001-R0900.txt.

#47 08/07/2026, 00:35 AM. Subject: Midwifery and Medical Care. No errors, and all signatures were found and processed.

#48 00:40 AM. Subject: School System and Teachers, same as #46.

#50 10:20 AM. All reports on Rendel completed in FC. Corrected factual errors here and there. No missing signatures occurred.

#51 10:22 AM. Source “Niederdorfelden_N0001-N2769.txt” reloaded into NB and deactivated source “Rendel”.

#52 10:26 AM. Submitted a list of individuals from the von Dorfelden family to the FC representing almost a complete genealogical line. That contained over 20 signatures. Not a single signature was omitted. Now let’s test what happens in Notebook “Niederdorfelden” when executing the batch prompts to generate lists of individuals from the von Dorfelden family. The original source uploaded months ago is still located there. Many signatures were previously not found there despite being present.

#53 08/07/2026, 10:42 AM. Unfortunately, it does not work with Niederdorfelden in this notebook: “Scan incomplete: The following signature numbers (#N0003, #N0012, #N0013, #N0019, #N0020) from source Niederdorfelden_N0001-N2769.txt could not be read out and verified.”

Checking Source Availability / Confirming File Absence:

The model attempts to access files and states: I have determined the file isn’t present and the filesystem is currently empty.

Adapting Scripting Approach:

Because the physical file is missing from the server filesystem (as NotebookLM passes sources as context snippets), the model adapts its scripts.

Missing / Omitted Signatures Detected:

Identifying Signatures Gaps / Detecting Signature Omissions:

The model states: I’ve detected signature omissions in the text and I’ve determined which signatures are absent from a specific range.

Triggering of Guardrail Prepared:

Clarifying Output Position / Finalizing Presentation Details:

The model notices the incompleteness and prepares the requested warning notice (ensuring the “scan incomplete” notice is correctly placed after the list of entries).

As can be seen, the AI itself still reports the issues in its log.

#54 08/07/2026, 10:42 AM. Attempted list generation in the newly created notebook “QS_Niederdorfelden”, where the previous attempt was also incomplete due to “missing” signatures. The source in this NB is identical to NB “Niederdorfelden”: Scan incomplete: The following signature numbers (#N0003, #N0012, #N0013, #N0019) from source Niederdorfelden_N0001-N2769.txt could not be read out and verified.

So something is still wrong here!

#55 The key points from the AI log:

Interpreting the Context: “noted the absence of files on the system”

Locating the Text / Reconstructing the Source: “recognizing that the specified file is unavailable […] reconstructing the source’s text from the provided excerpts to create a temporary file for analysis.”

Addressing Source Limitations: “handling cases where signature numbers might be truncated”

Identifying Missing Data / Confirming Missing Entries: “confirming missing signature numbers, specifically addressing #N0020…”

Developing the Script: “filtering out location mentions to identify relevant individuals.”

#56 The same prompt, but without the chat log instruction to use Python/Regex: Scan incomplete: The following signature numbers (#N0003, #N0012, #N0013, #N0019) from source Niederdorfelden_N0001-N2769.txt could not be read out and verified.

The key points from the AI log:

Interpreting the Context: “noted the absence of files on the system”

Locating the Text / Reconstructing the Source: “recognizing that the specified file is unavailable […] reconstructing the source’s text from the provided excerpts to create a temporary file for analysis.”

Confirming Knowledge Absence: “I have determined the file isn’t present […] the filesystem is currently empty”

Addressing Source Limitations: “handling cases where signature numbers might be truncated”

Identifying Missing Data / Confirming Missing Entries: “confirming missing signature numbers, specifically addressing #N0020…”

Confirming Missing Data: “I have re-examined signatures between N0001 and N0022 and can confirm that #N0003, #N0012, #N0013, and #N0019 are missing.”

Addressing Incomplete Scans: “I am noting that I cannot fully verify specific signature numbers from the provided source due to a system limit.”

Finalizing Presentation Details: “ensuring the ‘scan incomplete’ notice is correctly placed after the list of entries, as requested.”

4

Up to log entry #57, I performed the 6 fact checks without any errors, including corrections and additional rounds of corrections. Due to the number of facts to be checked in the text under review, not everything was found on the first try, which is why the corrected texts were sent through the fact check up to 4 times in some cases. But in the end, there was a clean result. And up to #57, there were no errors where a signature was not found. After #60, I sent further texts through the fact check, which then also had signature gaps.

#57 08/07/2026 1:11 PM Fact Check Incomplete: The following signature numbers (#B1025) from the source Büdesheim_B0001-B1200.txt could not be read and verified.

Let’s try that again without changing the procedure to see if it repeats… It repeats.

#58 I copied out the paragraph and am checking only the one where the signature is allegedly missing (which it isn’t!). This is the text section:

“Marriage agreements and marriage contracts: The legal formalization and securing of marriages formed an essential component of local jurisdiction. Detailed marriage settlements, marriage contracts, and marriage certificates were recorded before the responsible judicial offices and regional courts and preserved in corresponding deposit contracts (#B0403, #B0569, #B0587, #B0645, #B0673). In addition to formal marriage, the files also document legal disputes within the family sphere. Thus, civil lawsuits regarding the fulfillment of a promise of marriage (#B1025) as well as court proceedings for divorce (#B0788) can be found. Church rituals were documented matter-of-factly, such as the solemnization of marriage of betrothed couples (#B0264). Occasionally, actions such as the proclamation and solemnization of marriage also led to complaints among the responsible pastors (#B0461).”

#59 08/07/2026 1:20 PM Chat log deleted and sent the following prompt: What is written in signature #B1025? Output the regest VERBATIM. Result:

"The signature #B1025 is not contained in the text passages of the source ‘Büdesheim_B0001-B1200.txt’ currently available to me. In the provided excerpts, there is a gap between signature #B1024 and signature #B1029.

Due to this system-related text truncation (context limit), it is not possible for me to read and reproduce the regest of signature #B1025 verbatim.

Fact Check Incomplete: The following signature numbers (#B1025) from the source Büdesheim_B0001-B1200.txt could not be read and verified."

This is interesting, as the AI points out that there are entire sections that cannot be read. The AI work log for this:

Reviewing the Signature

I’m now carefully examining the user’s request, specifically focusing on the provided German query and the task to retrieve information.

Searching for Relevant Info

I am now searching within the provided context to locate the target item, and there seems to be a gap in the sequence.

Dateien wurden aufgelistet

Confirming File Absence

I have verified the absence of a specific text file, confirming I must work solely with the provided snippets for this task.

Analyzing Legal Records

I am now examining two legal records from 1872 and 1872-1875, focusing on civil proceedings and related details about the parties involved.

Identifying Source Location

I am checking the location of Friedberg Beklagte(r) Winter, Philipp, of Büdesheim, within my existing source material.

Verifying Source Consistency

I am now ensuring the text’s accuracy for a municipal finding aid, prioritizing readability and factual correctness concerning historical and legal details.

Addressing Source Gaps

I’ve determined that the absence of a specific entry in the sources is due to text limitations, indicating a gap in available information.

Addressing Source Limit

I’ve determined that due to a system-related limitation, specific entries like the signature cannot be verified in the current sources.

Confirming Entry Incompleteness

I have confirmed signature #B1025’s absence in the available source snippets, attributing this gap to text limitations between entries #B1024 and #B1029.

Confirming Signature Gap

I’ve determined that signature #B1025 is missing from the source due to a system-related context limit, writing a precise German response.

Drafting Precise Response

I’m finalizing the German response, carefully checking passage indices to cite specific source gaps [548, 549], to document missing signature #B1025.

#60 Fact Check Incomplete: The following signature numbers (#B0216) from the source Büdesheim_B0001-B1200.txt could not be read and verified.

So it’s not just with #B1025

#61 Test in the notebook “Büdesheim”. Prompt: “What is written in signature #B1025? Output the regest VERBATIM.”

Result: “Will, Katharina against Böckel, Heinrich / Charge / Offense: Fulfillment of a promise of marriage / Provenance (Justice) Judicial reference number C. T. 475 de 71/72 ab / (Prior) Provenances Hessen, Regional Court Bad Vilbel / Further details (case file) Type of proceeding: Civil lawsuit / Plaintiff: Will, Katharina / Residence: Altenstadt / Defendant: Böckel, Heinrich / Residence: Ilbeshausen, now Büdesheim”

However, the same source in the notebook “Fact-Check” produces the errors. What the prompt delivered is present in the source 100% as is!

#62 I am now going to pause altogether. I have the feeling that things will turn out similarly with my other projects the more the sources are accessed. Apparently (my guess as a layman), something gets caught in the backend cache or similar and somehow prevents the source from being read properly.

I tested what happens if I intentionally change a correct signature in the text being checked, for example by replacing signature number 700 with 701. In this case, the fact check correctly detects that there is a signature mismatch and suggests changing it back to signature number 700. So this functionality is working as well.

5

Even though at times it seemed like things were getting better, that cannot really be said. I could easily add another 50 entries to the text.

From Notebook “Fact-Checking” & “UB_Carben” & “Nobility Genealogy”

#78 07.08.2026 approx. 11:50 PM. Fact Check Incomplete: The following signature numbers (#C0980, #C2030) from the source Carben_C0001-C3998.txt could not be read and verified.

Rule in the chat configuration of all my notebooks, which makes these problems visible in the first place. If you don’t query a source directly in the prompt, you would never notice it:

“System-related gaps (abort rule): Strictly assume that the source collection is complete and contains all 2769 signatures. Should you be unable to find a signature number mentioned in the text to be checked within the sources currently available to you, this is solely due to a system-related text truncation (Context Limit). Never claim that a signature is missing from the collection! Do not abort the fact check in such a case and output this notice instead: “Fact Check Incomplete: The following signature numbers (e.g. #N0555) from the source [Name of the source] could not be read and verified.” at the end of the fact check.”

#79 07.08.2026 10:00 AM. Fact Check Incomplete: The following signature numbers (#C1803, #C2030) from the source Carben_C0001-C3998.txt could not be read and verified.

Thoughts:

Confirming Missing Signatures I am now confirming the absence of several specified signature identifiers, like #C2030 and #C1803, within the provided content.

Noting Source Incompleteness I’m noting that some signature identifiers, specifically #C1803 and #C2030 from source f1, couldn’t be read and verified.

Confirming Content Signatures I am now verifying the presence of the content signatures #C0982, and #C1333, while also noting those still missing from the sources.

#80 08.08.2026 10:03 AM. And the notebook “UB_Carben” reports something similar when creating reports. This concerns the exact same source Carben_C0001-C3998.txt: Scan incomplete: The following signature numbers (#C1223, #C1359, #C1452) from the source Carben_C0001-C3998.txt could not be read and verified.

Thoughts:

Checking the Directory I’m now checking the knowledge directory for accessible files that I can use for this task.

Analyzing Source Availability I’ve determined that the expected files are unavailable, so I’m re-examining the rules regarding source file handling.

Confirming Source Location I’ve determined that all of the source content is embedded directly within the prompt itself, not from any external files.

Verifying Source Integrity I’m now writing a script to confirm all provided text passages, assuming a complete set of source data with 3998 entries, as instructed, and handling potential text limitations.

Identifying Missing Signatures I’m now identifying signatures from the source that are missing from the extracted text and cross-referencing their location.

Locating Missing Data I’m now individually checking the text passages for the missing signatures, confirming #C1089’s presence in passage 478 and cross-referencing their positions.

Reporting Missing Data I’ve determined that three specific signatures are missing: #C1223, #C1359, and #C1452, which I can’t verify.

#81 08.08.2026 10:08 AM Notebook “UB_Carben” Prompt: Create a report on marriage settlements, dower rights, and morning gifts for noble family provision. To secure noble family alliances, noble marriage settlements were drawn up for marriage engagements (#C1036, #C1239, #C1755). In order to provide for women in the event of widowhood, rural estates or an annual rent were assigned to them as dower and marital property (#C0036, #C0184, #C1363).

Here too: Scan incomplete: The following signature numbers (#C1239) from the source Carben_C0001-C3998.txt could not be read and verified.

#82 08.08.2026 10:17 AM. All notebook chat logs deleted and all NBs closed. Try again later.

#83 08.08.2026 10:53 AM. Notebook “UB_Carben” similar prompt, just a different topic and signatures like #81. Result:

Scan incomplete: The following signature numbers (#C1417, #C1537, #C1757, #C2121, #C2915, #C2948, #C2994) from the source Carben_C0001-C3998.txt could not be read and verified.

Thoughts: The same as #80

#84 08.08.2026 10:08 AM Notebook “UB_Carben” Prompt: Create a report on marriage settlements, dower rights, and morning gifts for noble family provision. To secure noble family alliances, noble marriage settlements were drawn up for marriage engagements (#C1036, #C1239, #C1755). In order to provide for women in the event of widowhood, rural estates or an annual rent were assigned to them as dower and marital property (#C0036, #C0184, #C1363).

Here too: Scan incomplete: The following signature numbers (#C1239) from the source Carben_C0001-C3998.txt could not be read and verified.

#85 08.08.2026 10:17 AM. All notebooks chat logs deleted and the NBs closed.

#86 08.08.2026 12:20 PM. Notebook “_UB-Carben”, prompt as before, just a different topic and signatures. Message: Scan incomplete: The following signature numbers (#C2154) from the source Carben_C0001-C3998.txt could not be read and verified.

Notebook “Fact-Checking” reports: Fact Check Incomplete: The following signature numbers (#C1803, #C2030) from the source f1 could not be read and verified.

#87 08.08.2026 01:50 PM. Fact check for the Epilepsy project: Fact Check Incomplete: The following signature numbers (#Ep0243, #Ep0253, #Ep0254) from the source Epilepsy_Ep0001-Ep0506.txt could not be read and verified.

After that, I reduced the quantity again and it ran through cleanly, even the flagged ones then ran through without missing signatures. The source contains 506 data sets/signatures.

#88 In the notebook “Nobility Genealogy” are all local sources with 11k signatures. When evaluating Niederdorfelden, there was a message stating that signatures were unreadable. A second prompt attempt ran cleanly though, and everything was found.

#89 09.08.2026 10:00 PM. So the problem with the unfound signatures isn’t really gone. It is probably no longer permanent, but it keeps coming up from time to time. Deleting the chat and repeating usually works then… Currently limited to the sources Niederdorfelden, Rendel, and Carben, regardless of whether directly in the project NB or in the project NB Fact-Checking… But the signatures that are not found now change more often.

#90 10.08.2026 07:34 AM. Scan incomplete: The following signature numbers (#N0126, #N0203, #N0295, #N0627, #N0875, #N0876, #N0973, #N0974, #N1099, #N1105, #N1111, #N1164, #N1221, #N1248, #N1249, #N1290, #N1461, #N1550) from the source b1 could not be read and verified.

#91 10.08.2026 10:20 AM. NB Fact-Checking: Scan incomplete: The following signature numbers (#N0433, #N0559, #N0590, #N1760, #N1769, #N1775, #N1794, #N1836, #N1839, #N1852, #N1860, #N1957, #N2767) from the source Niederdorfelden_N0001-N2769.txt could not be read and verified.

#92 10.08.2026 11:07 AM. NB Fact-Checking: Check aborts with reference to context limit. But reports no unfound signatures.

#93 10.08.2026 11:19 AM. NB Fact-Checking: System-related gaps (abort rule): Scan incomplete: The following signature numbers (#N1769), (#N1775), (#N1785), (#N1828), (#N1836), (#N1839) from the source Niederdorfelden_N0001-N2769.txt could not be read and verified. Seems to be turning out to be a problem file in the backend. It can’t be that it’s always just this file. Today already read out all main sources and only Niederdorfelden causes these problems…

#94 10.08.2026 11:19 AM. Almost every FC is useless because not all signatures can be called up. The signatures that cannot be found also change from time to time.

#95 Scan incomplete: The following signature numbers (#C0002, #C0034, #C0036, #C0042) from the source vonC could not be read and verified. NB Fact-Checking Scan incomplete: The following signature numbers from the source h4 could not be read and verified.

Since the text to be checked from the source h4 could not be read out due to system constraints (owing to the context limit), a systematic fact check against the regestral tradition of Niederdorfelden (Niederdorfelden_N0001-N2769.txt) is unfortunately not possible at present.

Should you wish to manually verify specific text passages, historical figures, or concrete signatures, simply present the text to be checked directly to me here in the chat.

“h4” contains the text to be checked and is located in the source. He didn’t want to take h1-h3 either. Now after I asked NB what exactly it can’t read, NB repeated the FC and then it works…

#96 10.08.2026 04:20 PM. Notebook “Nobility Genealogy” Scan incomplete: The following signature numbers (#C0002, #C0034, #C0036, #C0042) from the source vonC could not be read and verified.

AI Log

Highlighting Source Gaps I’m noting several missing source entries from “vonC” based on specific identifiers, which impacts my ability to verify certain familial information.

Identifying Source Deficiencies I’m identifying missing source entries from ‘vonC’ including #C0002, #C0034, #C0036, and #C0042, which are unverifiable.

Investigating Family Lineage […] and I’m documenting unverified source entries.

6

@SUVANKAR_CHATTERJEE *
“Hi Suvankar, I noticed you viewed this report. Since this bug is completely blocking my paid workflows, could you please forward/escalate this detailed QA log to the internal NotebookLM engineering team?”*