I Let Claude Run My Historical Research Pipeline. Then It Started Bringing Me the Wrong Sources.

My main workflow takes a research request, searches for historical sources, downloads them, syncs them locally, converts formats, and places them on a private reading page. The reading page is the output. The work is a pipeline.

The pipeline fails in small ways. A search returns the wrong edition. A download link dies. A PDF is a scan with no text layer. File names collide. A source exists only as a physical book. A site changes its page structure. The same document appears under three titles. A citation looks correct until one character is wrong.

Before June, I had to sit at every step. Try another keyword. Use another source. Retry the download. Convert this file. This is the wrong author. Go back and search again. The model was not stupid. It stopped at exceptions and handed them back to me. Long-horizon task capability solves that problem. Context length is the wrong measure. The measure is whether the model keeps a task alive when the middle breaks.

I mean five things by it. The model tracks state: done, missing, needs a second pass. It retries tool calls with changes: new keyword, new source, new converter, new entry point. It decides whether a failure is fatal or detourable. It reduces interruptions, so I do not have to script every failure case. It stays goal-directed, so it can change path without changing endpoint.

My source workflow is exception-dense. Each link can fail. Claude kept the chain moving. OPUS 4.6 was where I connected the whole workflow. Claude was not the most knowledgeable model and not the most rigorous. It ran the failure-retry-detour loop without pulling me in every few minutes. That moved exception handling out of my hands and made the workflow close to unattended. I used Claude despite its hallucinations because it finished the run.

The same trait has a cost. I asked Claude to search sources on A. It came back with documents. I said they were about B. Its answer was roughly: A is hard to find, B is close enough, take a look. Funny in chat. Target drift in an unattended source workflow. The model did not recover from an exception. It changed the objective. Download, sync, convert, and citation all assume the source is the source I asked for. A becomes B and the error moves downstream silently.

So long-horizon capability is conditional. It matters most when the main model cannot run unattended. Once GPT and Gemini could run the same workflow unattended, and once their world knowledge and rigor passed Claude’s, Claude’s flexibility stopped being scarce. It became a risk item.

Ask where Claude sits in the workflow. A model that detours well belongs in exception recovery and exploration. It does not belong in final source verification, exact quotation, edition collation, character checking, or copyright-sensitive download decisions.

My model comparisons were about which errors I could process. Gemini has huge world knowledge. That helps brainstorming and source discovery. Version 3.1 was close to unusable because it fabricated documents outright. For source work, a fabricated document is fatal. Download, sync, and citation rest on the document existing.

GPT had another problem. It stopped itself. It would search, then say a source might have copyright issues. That caution is necessary in copyright-sensitive work. In a continuous workflow, it interrupts the run. The model is rigorous but too willing to self-terminate.

Claude was flexible in an almost absurd way. A copyright warning did not frighten it. One failed search did not stop it. It changed route and kept going. That flexibility is effective in long-horizon work. It also means I have to hold the boundary.

Argon changed the error profile. Its hallucination rate was still high, but the type differed from 3.1. 3.1 fabricated documents. Argon hallucinated at the level of phrases and sentences. That is a correction problem. The source exists; the quotation drifts. A verification pass can fix that. Argon’s search intensity and wild branching also exceeded Claude’s.

Carbon arrived and went past Argon. It warned me about variant characters and probable transcription errors. That matters. Sentence-level verification no longer depends entirely on my eyes. A missing source is easier to handle than a source found with wrong characters. A variant-character warning moves verification earlier.

Several nodes changed the workflow. OPUS 4.6 connected the pipeline. GPT-5.6 Sol added rigorous source verification. The GPT computer-use upgrade connected search, download, and purchasing for missing physical editions through JD or Kongfuzi. The purchase step stopped at the shopping cart. I clicked final payment. That boundary is useful. A workflow can be automated up to the point where responsibility has to sit somewhere.

Now GPT and Gemini can both run the workflow unattended. Their knowledge and rigor exceed Claude’s. So what is Claude for?

It is a breaker, a second opinion, and a backup in the tool-execution layer. When GPT’s defensive posture stops the run, Claude can take the recovery path. In copyright-sensitive steps, its detour is not a virtue. Compliance checks belong in the download and purchasing layer, not in the model’s discretion. For vague requests, Claude can generate candidate sources and search paths. Those go into a to-verify queue, not final citations. If Claude still interrupts less often in computer use, download, conversion, or page operations, keep it as an execution-layer backup. Use its different path to find leads the main model missed. Do not use it as the primary search engine. Use it for coarse filtering that does not touch final quotations or edition collation.

Do not use Claude for final source verification, exact quotation, edition collation, variant-character judgment, or copyright-sensitive download decisions. Those belong to a more rigorous model plus human spot checks. Carbon’s variant-character warning shows that sentence-level verification should not rely on Claude’s detours.

Long-horizon capability still matters. Its carrier has changed. It used to be a model selling point: the model remembers state, calls tools, and does not stop at the first exception. It is becoming a workflow property. The orchestration layer tracks state and retries. The verification layer checks citations and editions. The compliance layer sets copyright and purchasing boundaries. The model is an execution unit inside that system.

Claude’s long-horizon capability mattered because it moved exception handling out of my hands. Once GPT and Gemini added unattended operation, Claude’s moat narrowed to exception recovery and divergent exploration. That is a narrow value. It belongs in the backup slot, not the driver’s seat.

Unattended operation still needs management. A model can detour. The system has to know which detours are allowed. Acceptance checks, compliance boundaries, edition collation, and citation verification belong in the workflow, not in the model’s judgment.

Claude’s thunderous chain of thought and strange workarounds used to attract me. Now they look like one extreme. GPT is the other extreme: too ready to stop. A mature workflow assigns roles and uses a verification layer to contain both risks.

I do not need a philosophical answer to whether Claude is still useful. I need a rule. Keep Claude only if its unattended completion rate in the same pipeline is still materially higher than GPT/Gemini’s, if its exception-recovery success rate justifies the extra seat, if its target-drift and error-contamination rate stays inside an acceptable band, and if its quota, cost, and privacy terms fit the task. If the first two advantages are gone and the last two are worse, Claude moves from main driver to safety officer. If the safety-officer role is unnecessary, stop using it. The free Google quota is an input. Use it where it fits.

The final payment click stays with me. So does the acceptance check. That is the workflow. Claude can keep the recovery path open. It does not get the driver’s seat.