Post-training tuning and continuous learning

Speaker notes

This deck is about what happens after a model ships. Training made it, and then it froze: the tools, schemas and failures it meets afterwards never reach its weights. Here the system is built to keep learning. Learners turn execution reports into memory, path statistics and routing advice, and small open models are tuned on checked examples and kept only when they beat their base on a test they never saw, under a rule fixed before scoring. Every result in this deck comes from our evaluation tasks, and nothing has been promoted or admitted.

Speaker notes

Training made the model, and then it froze. The tools, schemas and failures it meets after its release never reach its weights, so the same mistakes repeat. A bigger model isn’t the fix: on the same frozen test, Llama 3.2 3B beat the 1B, 52.0% to 12.0% exact match, but it wrote typed values as strings even more often, in 17.5% of its calls against 8.3%, and strict servers reject those calls. Even Muse Glimmer 30B, unchanged, answered 93.3% of a fresh test exactly right, and 96.7% of its replies were valid JSON. The fix is a loop built to keep learning after deployment, which keeps only what beats the base.

• Measured Llama 3.2 1B and 3B, unchanged, on the same frozen 150-row test: exact match 12.0% and 52.0%; typed values written as strings in 8.3% and 17.5% of calls. Tessera’s continuous-tuning record: the unchanged base models on the frozen, tool-call evaluation. A typed value written as a string, such as a quoted number, makes a strict server reject the call. Evidence
• Measured Muse Glimmer 30B, unchanged, on a fresh 150-row test: exact match 93.3%, valid JSON 96.7%. The base model’s scores in the published run record, on a fresh, tool-call evaluation whose hashes were registered before scoring. Evidence

Speaker notes

Read the loop clockwise from the top left. Owned templates generate the examples, so no model writes a row or a label. Curation rejects repeats and anything that looks like a test row. The test set is frozen, and its hashes are recorded before any model sees it. A small adapter trains on top of the base, which never changes, and it stays a candidate. Evaluation scores the candidate, the current champion and the unchanged base on the same rows, and the promotion rule, written before any training, decides. The saved adapter must then reload onto a fresh copy of the base and reply exactly the same. Only then does the champion pointer move. Keeping the champion is a valid outcome, and because every version is kept, rolling back is moving the pointer back. The next round starts from the champion, with new examples aimed at its failures.

Speaker notes

Everything that decides was written down before the results it judges. The frozen test is 150 rows over whole tool families that training never sees, pinned by its hash. Each round also draws fresh rows, and their hashes are recorded before any model sees them. The promotion rule was committed before the first training round, and a test pins it. Fingerprints keep look-alikes of test rows out of training. And the rule has said no to real-looking gains. In Llama 3.2 1B’s round 4, a new strategy lifted the development score from 61.25% to 76.25%, but the frozen test stayed at 50.7%. Llama 3.2 3B’s round 2 scored 79.3% but cut schema validity by 0.025, past its 0.02 limit. Muse Glimmer 30B trained on 63 examples gained 5.3 points on the frozen test, but its 95% interval started at 0.0. All three stayed out.

• Measured Frozen test: 150 rows over tool families held out of training. Tessera’s continuous-tuning record: the test split holds the buckets and feeds tool families, pinned by digest, and training prompts only ever show other families. Each round also freezes a fresh holdout from the test families. Evidence
• Measured Llama 3.2 1B, round 4: schema-augmented training lifted the development score from 61.25% to 76.25%, but the frozen test stayed at 50.7%; kept out. The development set and the frozen test disagreed. The augmented adapter tied the champion on the test, while its schema validity fell to 61.7% and its false refusals rose to 17.5%, so the rule kept the champion. Evidence
• Measured Llama 3.2 3B, round 2: 79.3% exact match, but schema validity fell 0.025 against a 0.02 limit; kept out. The adapter scored above the champion, but the promotion rule, committed before any training round, allows schema validity to drop by at most 0.02, so the champion stayed. Evidence
• Measured Muse Glimmer 30B with 63 training examples: +5.3 points on the frozen test, 95% CI 0.0 to +11.3; kept out. The 63-row run on the frozen development test: exact match 87.3% to 92.7%, with 14 rows better and 6 worse. The paired bootstrap lower bound was not above zero, so the rule kept the base. Evidence

Speaker notes

Two small Llama models, the same frozen, 150-task evaluation and the same loop on one local GPU. The 1B went from 12.0% to 43.3% in round 1 and to 50.7% in round 3. Rounds 2 and 4 did not beat the champion, so the rule kept it. Round 1 gained 31.3 points, with a 95% interval of 22.0 to 40.7, and round 3 gained 7.3, with 0.7 to 14.7. The 3B went from 52.0% to 72.7% in round 1, a gain of 20.7 points with 12.0 to 29.3. Its round 2 scored 79.3% but was kept out by a schema guard. The first 63 owned examples were enough to stop the 1B writing typed values as strings, and its type correctness rose from 37.5% to 94.2%. Each configuration ran once, with one seed.

• Measured Llama 3.2 1B, rounds 0 to 4 on the frozen, 150-task evaluation: exact match 12.0%, 43.3%, 43.3%, 50.7% and 50.7%; promoted in rounds 1 and 3. Tessera’s continuous-tuning record: one local GPU, one seed per configuration. Rounds 2 and 4 did not beat the champion, so the pre-registered rule kept it. Evidence
• Measured Llama 3.2 3B, rounds 0 to 2 on the same test: exact match 52.0%, 72.7% and 79.3%; promoted in round 1, round 2 kept out by the schema guard, rounds 3 and 4 not run. The same loop, test and rule. Round 2’s schema validity fell 0.025 against a 0.02 limit. Training then stopped until there is more checked data. Evidence
• Measured Promoted gains, with paired bootstrap 95% CIs: 1B round 1 +31.3 points (+22.0 to +40.7), round 3 +7.3 (+0.7 to +14.7); 3B round 1 +20.7 (+12.0 to +29.3). The promotion decisions in Tessera’s continuous-tuning record: 55 rows better and 8 worse in the 1B’s round 1, 20 and 9 in its round 3, and 40 and 9 in the 3B’s round 1. Evidence
• Measured 63 owned examples: typed values written as strings fell from 8.3% to 0% of calls, and type correctness rose from 37.5% to 94.2% (1B, round 1). The frozen-test diagnostics for the unchanged 1B and its round 1 adapter, in Tessera’s continuous-tuning record. Evidence
• Measured Llama 3.2 3B (Hybrid LoRA + Compact Pointer): exact recall 64.2% to 91.7% on a 120-row test with a 38.5-token pointer, cutting shim tokens by 91.0% vs a 427.6-token BM25 shim (8 of 8 reload check). Local RTX 3060 LoRA-as-Memory paired with a 1-line compact pointer for live post-training delta task objects; example data only, nothing admitted or deployed. Evidence
• Measured Llama 3.2 1B (Herma Shim Router): full route exact match 26.7% to 98.3% on a 120-row 5-action test (vs 10.0% zero-shot 1B), with 100.0% contributor-privacy refusal recall and 0.0% false refusals. Local RTX 3060 1B edge router adapter classifying direct reply, task-object injection, task decomposition, frontier synthesis, and contributor-tier privacy refusal. Evidence

Speaker notes

The largest model we tuned is Muse Glimmer 30B, with QLoRA on free cloud GPUs and 511 training rows. It was scored on a fresh 150-row test built after training, with its hashes registered before scoring. Exact match rose from 93.3% to 99.3%, a gain of 6.0 points with a 95% interval of 2.0 to 10.7: 10 rows got better and 1 got worse. Valid JSON went from 96.7% to 100.0%, and no test row matched a training row. The promotion rule, fixed before scoring, passed all 12 checks, and a fresh reload of the adapter reproduced its replies on 8 of 8 rows. On the frozen development test it went from 87.3% to 98.7%. The rule said promote, but nothing has been promoted or admitted, and the adapter weights aren’t published.

• Measured Muse Glimmer 30B: exact match 93.3% to 99.3% on a fresh 150-row test, +6.0 points, 95% CI +2.0 to +10.7; 10 rows improved, 1 regressed. QLoRA on free cloud GPUs, October 29, 2026. The fresh test was built after training, and its hashes were registered before scoring. Evidence
• Measured Muse Glimmer 30B: promotion rule 12 of 12 checks passed; a fresh reload matched 8 of 8 regenerated rows. The rule, tool-call-specialist.promotion version 1, was fixed before scoring and returned promote_candidate. Nothing has been promoted or admitted, and the adapter weights are not published. Evidence
• Measured Muse Glimmer 30B: valid JSON 96.7% to 100.0%, correct refusals 97.3% to 99.3%; 0 test rows match the 511 training rows. No exact, request-level or held-out-family match against the training rows. Against the older test set, 54 rows share only a request shape. Evidence
• Measured Muse Glimmer 30B on the frozen development test: 87.3% to 98.7%. The test used while developing the runs. It served more than one run, so the fresh test is the one that decides. Evidence
• Measured Muse Glimmer 30B (LoRA-as-Memory): closed-book exact match 0.0% to 62.5% on a pre-registered 120-row test with 0 shim tokens (+33.3 points vs 222.3-token BM25 shim, 2.28 times faster, 8 of 8 reload check). QLoRA on two free Kaggle T4 GPUs, October 31, 2026: closed-book task-object recall and multi-step workflow synthesis without runtime context injection. Promoted candidate only; nothing has been admitted or deployed. Evidence

Speaker notes

Not all learning changes a model. Learners are ordinary verbs that read execution reports and write versioned nouns. Memory fingerprints each run, so before new work an agent can recall similar past runs and how they went, failures included. Path statistics group runs into scorecards of what succeeds; they show associations, not causes. Routing weights suggest which shim, method, harness or model to try, keep exploring the alternatives, and move only with enough evidence. Their outputs are advice, never a permission. Changing model weights from real runs is the third layer: it needs opted-in, verified, rights-cleared runs, and that export is designed but not built, so today’s tuning uses example examples. The learners pass their tests, but they haven’t yet been shown to improve real work.

• Measured Memory, path-statistics and routing-weights learners: implemented and tested. Tessera’s execution-learning record. Local tests are not evidence that any learner improves real work; that needs the evaluations each learner names, on real permitted runs. Evidence
• Measured SIRA Self-Learning RAG + Herma Shims: Recall@1 23.3% to 100.0% across 60 queries over 79 seed packs (MRR 0.341 to 1.000 in 1.0 round, 82.8% token reduction, 0.30 ms retrieval, 10 compiled shims at 0.026 ms, 10 of 10 independent reviews). Offline standard-library evaluation over the 79-pack library: RARTF noun enrichment, analogy-aware cross-domain indexing, MR-Search adaptive reflection, and DGM-H independent verification (409 self-approval rejected). Evidence

Speaker notes

You don’t need to rent a GPU to try this. Muse Glimmer 30B was tuned with QLoRA on Kaggle’s free notebook GPUs, two T4s: its 511-row run took 64 training steps in 83 minutes, and the run record lists every file’s hash. Llama 3.2 1B and 3B ran every round on one RTX 3060 with 12 GB: about 46 and 51 GPU-minutes, with no provider calls and no spending. Without a GPU, a local CPU backend runs every verb, receipt, report and decision on a CPU, so you can check the plumbing before you spend anything. One tick command plans and runs a round, and Tessera’s continuous-tuning guide has the rest.

• Measured Muse Glimmer 30B QLoRA on Kaggle’s two free T4 GPUs: 64 training steps in 83 minutes, on 511 rows. The 511-row run record: 64 steps and 4,997.5 seconds of training, with each stage, library version and the peak memory of both GPUs recorded. The whole job took 8,789.5 seconds. Evidence
• Measured Llama 3.2 1B and 3B on one RTX 3060 12 GB: about 46 and 51 GPU-minutes, with no provider calls and no spending. Tessera’s continuous-tuning record: rounds 0 to 4 and a serving check for the 1B, and rounds 0 to 2 for the 3B. The GPU is owned, and its electricity is not metered. Evidence

Speaker notes

These rules keep the loop honest. Tests are frozen, held out by whole tool families and pinned by hash. The promotion rule and each fresh test are recorded before scoring. Fingerprints keep look-alikes of test rows out of training. A saved adapter must reload onto a fresh copy of the base and reply exactly the same. The rule asks for a real gain, with a lower bound above zero, and no drop in valid output, refusals or general skills. Because adapters are never overwritten and the champion is one pointer that keeps every version, rolling back is moving that pointer back. Promotion here changes no deployment, publication, admission or serving. Privacy is checked as well: after the 1B rounds, a second key tried to read all 303 records, blob chunks and runs they wrote, and every read was denied.

• Measured Another key tried to read everything the 1B rounds wrote: all 303 reads were denied. A final sweep after the 1B rounds: another principal tried to read 115 private records, 133 blob chunks and 55 execution runs. After the 3B rounds, all 193 such reads were denied too. Evidence
• Measured Meta ARE 4D Scenario Matrix (1,000 Episodes): 32.5% stakeholder-blocked vs 67.5% safely approved (15.1% direct plus 52.4% human-partitioned) across 1,000 4D scenarios spanning 25 domains. Deterministic Latin-hypercube evaluation over 10^12 addressable 4D tuples (persona, environment, method, stakeholder constraint), separating direct agent work, human-partitioned drafts, and stakeholder-blocked escalations. Evidence

Speaker notes

These are the limits. Every result is made up and comes from one task, typed tool calls for Tessera’s own tools, so it says nothing about other tasks or real traffic. Each configuration ran once, with one seed. The tests are small, so a real gain of a few points can be missed. Nothing learns from live traffic yet: rounds start on demand, and exporting real runs for training is designed but not built. The adapter weights aren’t published, and nothing has been promoted or admitted; a promotion here moved only a private pointer. The learners pass their tests but haven’t yet been shown to improve real work. Every kept-out round and every failure stays on the record.

• Measured The 150-row tests are small. Tessera’s continuous-tuning record: with the disagreement these runs showed, a 150-row test detects a true 5-point gain only about 31% of the time, and 80% power needs about 560 rows. Evidence