In this briefing
- 01Actual tool returns enter reviewable execution trajectories
- 02A single production task begins to become a training sample
- 03Expert standards determine whether the training signal works
- 04A platform launch and financing move the approach into the supplier market
- →What to watch next
- ↗Sources and verification
- DeepSeek Harness v0.2.2-alpha.3 makes execution trajectories preserve the actual text returned by failed tools, and the version makes automatic-review requests carry the Session identity and request purpose. The version remains a pre-release.
- Trajectory's vendor research reports that a single rollout for each task can provide a training signal on some knowledge-work benchmarks. The results depend on the task distribution, grading method and training conditions, and independent researchers have not reproduced the results.
- Another Trajectory study shows that a general-purpose model grader does not necessarily provide an effective training signal, while a small set of expert-graded samples can improve a grader. The related legal-task results still have overlapping confidence intervals and limits on grading consistency.
Actual tool returns enter reviewable execution trajectories
On 3 November, DeepSeek released DeepSeek Harness v0.2.2-alpha.3 as a pre-release. The pre-release makes Trajectory lists, details and copied content preserve the actual text returned by failed tools, and the same rule covers native tools and nested calls. The version also makes an automatic-review request carry the corresponding Session identity and request purpose, and the version prevents dispatch when a dynamic request header is invalid.
The version also places local diffs and pull-request diffs in the experimental change-request and worktree entry, and the version adds the apply_patch tool with the tool disabled by default. The release notes also state that read-before-write and external-change protections are now disabled by default, while sandbox limits and Review remain. This combination creates a more complete execution record, but this combination also requires an enterprise to verify every default separately.
For enterprise adoption, a training system can reconstruct a failure as an auditable learning sample only when the system preserves the actual tool return, session identity, request purpose and code diff. Synthetic error text can hide the actual cause, but a complete trajectory can also contain sensitive data, so an enterprise still needs access control, field redaction, retention periods and human review. A pre-release and changes to defaults cannot establish production security by themselves.
A single production task begins to become a training sample
Trajectory's study, ‘You Only Rollout Once’, published on 2 November, tests four knowledge-work benchmarks covering legal analysis, professional legal and financial reasoning, business-workflow automation and scientific research. The study uses Inkling-Small, which has 276 billion total parameters and 12 billion active parameters. Group relative policy optimisation generates 8 rollouts for each task, so a training step containing 32 rollouts covers 4 tasks; the single-rollout methods make the same step cover 32 tasks.
The study reports that FlashREINFORCE, using one rollout for each task, achieves an improvement similar to group relative policy optimisation on the Harvey Legal Agent Benchmark. The study also reports that FlashREINFORCE improves the held-out task score on PRBench by 2.0 points, while group relative policy optimisation improves the score by 0.8 points. The task-specific baseline method that Trajectory calls GS1 raises the held-out pass rate on LABBench2 from 63% to 68%, but the figures come from vendor research and depend on the specified model, task distribution, learning rate and grader.
For enterprise adoption, if a single-rollout method works reliably, an enterprise may be able to use non-repeatable real tasks without constructing several identical environments for the same business request. The enterprise still needs to bind the task type, input permissions, tool version, grading criteria and model version to every trajectory; otherwise, the training system cannot distinguish variation in business difficulty from variation in model behaviour. The current results support further validation on the specified benchmarks, and the current results do not support sending all production logs directly into training.
Expert standards determine whether the training signal works
Trajectory's study, ‘Continual Grading’, published on the same day, first tests a general-purpose model grader. The research team makes GPT-6 Luna use a one-line prompt to grade 40000 rollouts, but the model trained against those scores shows almost no improvement. The research team then uses 30 expert-graded tasks to construct candidate graders and another 30 expert-graded tasks to select a grader. The best scalar grader correctly ranks 80.3% of response pairs, while the pairwise grader correctly ranks 71.8% of response pairs.
Legal tasks show a stricter boundary. Trajectory trains Qwen3.6-35B-A3B on the Harvey Legal Agent Benchmark with generated rubrics, and the method achieves a 6.6% all-criteria pass rate during the specified evaluation interval. The method using expert rubrics achieves 4.4%, and both methods start at 1.4%. The research team states explicitly that the two results have overlapping confidence intervals, so the research team cannot conclude that generated rubrics are better. Harvey's background results published in May (month 5) also show a highest strict all-criteria pass rate of only 7.1%, while the highest configuration costs about $50.90 for each task and takes about 22 minutes.
For enterprise adoption, a grader determines which behaviour a model reinforces, so the grader itself needs version control, expert review and validation on held-out tasks. An enterprise can ask experts to define facts, conclusions, citations, structure and validation actions, but the enterprise cannot treat one general-purpose score as a complete quality standard. The legal benchmark also shows that model performance varies across specialist fields, and quality, cost and latency need to enter the same acceptance record.
A platform launch and financing move the approach into the supplier market
On 2 November, TechCrunch reported that Trajectory had completed a $45 million Series A financing at a $300 million valuation. Sequoia led the round, and NVIDIA, Bessemer Venture Partners and Bain Capital Ventures participated. TechCrunch also reported that Trajectory had 18 employees and that Trajectory had made its post-training platform publicly available.
TechCrunch reported statements that Harvey, Rogo, Decagon, Mercor and NVIDIA already use Trajectory's platform or are customers. Trajectory's research articles connect production trajectories, expert grading and open-model post-training into one technical approach. The information shows that a more complete post-training supply market is emerging, but a media report and vendor research cannot establish long-term reliability, cross-industry applicability or independently verified customer outcomes.
For enterprise adoption, the object of procurement is expanding from a one-off fine-tuning service into an operating system that continually collects trajectories, maintains grading criteria and updates a model. An enterprise needs to place data ownership, model weights, training frequency, rollback, audit, cost and supplier-exit arrangements within the same governance boundary. The financing amount and public availability show increasing market investment, but the financing amount cannot replace validation of security, quality and business outcomes.
What to watch next
- Can DeepSeek Harness publish boundaries for redacting, retaining and controlling access to actual tool returns in a final release?
- Can independent teams reproduce Trajectory's single-rollout improvements with different models and continually changing production tasks?
- Can enterprise post-training platforms create a durable and auditable relationship between expert grading criteria, model updates, rollback records and actual business outcomes?
Sources and verification
Golden Data has edited this briefing from the public materials listed above. The original sources govern facts and figures. The enterprise relevance sections are Golden Data editorial analysis and do not constitute an endorsement of any third-party product.
← Back to AI Daily Briefing