Learning

Two speeds

One store cannot be both quick and stable, so the genome has two.

• Fast: observations. When a customer edits an output or comments on a thumbs-down, the platform records an observation. It shows up in the next model call, marked as tentative and ranked below every gene.
• Slow: learned genes. An observation becomes a gene only after the same preference has shown up in more than one run and a fixed set of checks has passed. This happens at night.

No model is called in this learning. Every step is a fixed rule applied to the log, so the same log at the same time gives the same result.

Built, flag off by default G08, G10, G12 Recording observations and nightly consolidation have code and tests. They run only when a feature flag is on.

Nightly consolidation

The nightly job looks at observations from the last 14 days, one agent in one workspace at a time.

1. Normalise. Lower-case the text, strip the platform’s own prefix, drop punctuation and 32 very common words.
2. Group. Two observations are similar when their sets of words overlap by at least half (token Jaccard similarity of 0.5 or more).
3. Qualify. A group needs at least 2 observations from at least 2 different runs. One run repeated twice is not a pattern. If the immune check is on, the minimum is 3 runs plus an evidence floor described on Governance.
4. Check. The candidate gene is refused if it contains an injection marker, if it conflicts with the agent’s identity, if it contains a fact (an email address, a URL, an amount, a phone number or a run of four or more digits), or if it nearly duplicates a customer or kernel gene.
5. Write. A passing candidate becomes a learned style gene at strength 0.5. If a similar learned gene already exists, that gene gets more evidence instead of a copy. If the customer rejected a similar proposal less than 30 days ago, the candidate is dropped. If a similar gene retired earlier, it returns only with more evidence than it had, or after 30 days.

At most 5 consolidations and refreshes are written per workspace per pass, so one noisy workspace cannot rewrite its agents overnight. Consolidation consumes the observations it used.

How strength moves

Each time the customer accepts or rejects an output, the platform records one outcome. The outcome credits the genes that were in that call’s prompt, according to the platform’s own record. It does not use the model’s account of which genes it followed, because a model can be wrong about that.

For a learned gene with strength s:

• Confirmed: s becomes s + step × (1 − s).
• Contradicted: s becomes s − step × s.

The step is 0.5 × plasticity. Plasticity is 0.25 + 0.75 × 20 / (20 + n), where n is the number of outcomes the genome has already seen. So plasticity is 1.0 for a new genome, 0.625 after 20 outcomes, and it approaches 0.25. A new customer shapes the agent quickly. One odd edit cannot knock over an established genome.

• A learned gene below strength 0.15 retires. It leaves the genome, and a record of it is kept.
• A learned gene below 0.35 stays in the genome but is not compiled. It is still collecting evidence.
• A customer gene does not learn. Two contradictions since its last confirmation flag it as a hint for the customer to review. A confirmation clears the count.
• Genes in other strata are not changed by outcomes.

These are the default constants. They belong to tuning configuration version 0, the only version that exists. Built and tested G04

A worked example

This example comes from the formula, not from a measurement. A new learned gene starts at 0.5 in a genome that has seen no outcomes. Each of the next five outcomes includes it.

Strength of one learned gene over five outcomes

Outcomes seen	Result	Plasticity	Step	Strength before	Strength after
0	confirmed	1.0000	0.5000	0.5000	0.7500
1	confirmed	0.9643	0.4821	0.7500	0.8705
2	contradicted	0.9318	0.4659	0.8705	0.4649
3	confirmed	0.9022	0.4511	0.4649	0.7063
4	contradicted	0.8750	0.4375	0.7063	0.3973

The gene ends at 0.3973, above 0.35, so it is still compiled. A fresh gene at 0.5 that is contradicted once at the start falls to 0.25 and stops being compiled. A second contradiction takes it to about 0.13, and it retires. In an old genome where plasticity has reached its floor, the same gene needs ten contradictions in a row to fall below 0.15.

The nightly job

Consolidation is one phase of a nightly job that runs at 02:30. The job has ten ordered phases. Each phase is isolated, so a failure in one does not stop the others. Phases include a re-screen of stored genes for injected text, the verdicts of the leave-one-out experiment below, gene family proposals, competence counts, expiry of open threads and episodes, and the live quality scorecards. One phase, governance alerts, is a placeholder that does nothing yet.

Built and tested G13 The runner has code and tests, and it checks no genome flag. Each phase checks its own flags.

Testing a gene by leaving it out

An optional experiment tests whether a gene helps. On a small share of eligible runs (a 10% keyed draw), one eligible gene is left out of the prompt, and the outcomes with and without it are counted. Each night the platform issues a verdict per gene: keep, retire, no effect, or stop because it did harm. The experiment never runs for agents that can act on the world, and a workspace can opt out.

Built, flag off by default G38, G39 It has code and tests. It is off unless a feature flag turns it on. No result file reports its verdicts.

Limits and what is untested

• The benchmark uses its own simpler consolidation: a similarity of 0.6 against the first member of a group, at least 2 members, and a starting strength of 0.45 plus 0.1 for each extra member, capped at 0.9. These are heuristic choices, not tuned or validated values.
• The benchmark’s personas never change their preferences. Strength decay, retirement and relearning are therefore never triggered in any measured result. They are implemented and have unit tests, but their effect on real agents is not measured.
• Matching is by shared words. A paraphrase such as “keep it short” and “under 40 words” is not matched.
• A learned gene’s strength says how often it was confirmed. It does not say the gene is right.