TRL

Leading words: evidence not intent, lowest row wins, blind panel, fresh eyes every cycle, one level at a time, failing check first, provenance label, ask before outward.

A TRL is the highest level whose exit criteria are demonstrated by evidence. Plans, intent, and self-report count for nothing. The builder (this chat) never scores itself upward. Three fresh assessors do, every cycle, from a deliberately thin slice of context.

User input

$ARGUMENTS

Flag	Default	Meaning
--target N	8	Stop when every in-scope row reaches N. 8 = launch-ready.
--cycles N	5	Max assess → build cycles. Clamp 1–12. Each cycle spawns three assessors; five keeps one run inside one session.
--assess-only	off	Baseline panel and gap list only. Never edits the project.
--circuit	off	Skip the baseline checkpoint. Outward-facing asks still happen.

Strip flags from the scope text. A lone integer is --target.

Operating contract

• Do not auto-start. The user must ask for this skill.
• Evidence not intent. Only reproducible evidence moves a row: a check that ran, its command, its output, its provenance. Read references/scale.md for levels and provenance.
• Lowest row wins. Overall TRL = the minimum over in-scope rows. Never average.
• Blind panel. Assessors never see the target, the builder's score, earlier verdicts, the cycle number, or .balakit/trl/. A fresh set of assessors runs every cycle. Commit messages and code comments written during the loop never name a level or the target, because assessors read the repo.
• Panel beats builder. If the builder's sealed score is higher than the panel's, the panel's number stands.
• One level at a time. A row rises at most one level per cycle. Do not build level N+2 evidence before level N+1 passes.
• Failing check first. Write each next-level exit criterion as an executable check, see it fail (or confirm it is missing), then build.
• Ask before outward. Deploys, paid resources, DNS, real accounts, emails or messages to people, production data, and secrets always get a user question, even under --circuit. Use request_secret or the host equivalent for secrets. Never invent one.
• Never author attested evidence. Human sign-off, beta feedback, and dogfood notes come from the user, word for word, with who and when.
• --assess-only writes only the ledger and dossier under .balakit/trl/.

Runtime state

Start every pipeline response with this block. Update it mechanically.

TRL_STATE
stage: frame | baseline | panel | gap | build | done
scope: app | feature | improvement — <one line>
target: N
cycle: n/max
overall: unset | N (provenance)
ledger: unset | .balakit/trl/<slug>/ledger.md
next_action: one action only

If the block is missing after a context change, rebuild it from the ledger and confirm with the user before advancing.

Pipeline

Copy these steps into the todo list verbatim. Skip with skip: <reason>.

1. Parse flags. Detect the host and read references/hosts.md. The host must spawn subagents. If it cannot, stop.
2. Frame. Fill the ledger header from references/ledger.md: scope, target, the relevant environment and operational environment for this scope, in-scope rows from references/dimensions.md (each dropped row gets n/a: <reason>), and any critical technology elements (novel or risky parts) as extra rows. Ask only for what the repo cannot tell you.
3. Baseline evidence. Run what already exists (tests, builds, linters, deploy config, health checks). Write the facts-only dossier. Write the builder's sealed score per row into the ledger.
4. Panel. Spawn the three lenses in parallel from references/assessor-prompts.md. Aggregate (below). Log the cycle in the ledger.
5. Checkpoint. Show the scorecard and the gap plan for the next level. --assess-only → reply and stop. --circuit → skip the wait on cycle 1 only. Otherwise ask (below).
6. Gap → checks. For each row below target, lowest first, turn the next level's unmet criteria into evidence checks in the project's own test layout and tools. Run them. Confirm they fail or are missing.
7. Build until those checks pass. Ask before outward actions. Record each passing check in the dossier with its command, output summary, and provenance.
8. Next cycle. Back to step 4 with new assessors. Stop on: target reached, cycles exhausted, plateau, or blocked on the user (below).
9. Report per the reply contract. Leave the ledger in place.

Aggregation

Per row, from the three verdicts:

• Spread ≤ 1 → panel level = median.
• Spread ≥ 2 → panel level = minimum. Log disagreement with each lens's blocking reason. Do not average.
• The auditor marks a dossier claim fail → that claim does not count, for every lens, this cycle.
• The panel level stands, whether it is above or below the builder's sealed score. The sealed score is logged only to show calibration.
• Row provenance = the weakest provenance among the evidence that row's level rests on.

Overall = minimum row. The label is TRL N · <provenance>, for example TRL 8 · staged.

The gap list is the union of every lens's blocking lines for rows below target, deduplicated.

Stop conditions

• Target reached. Every in-scope row ≥ target on the panel.
• Cycles exhausted. Report where it stands.
• Plateau. No row rose for two cycles in a row. Stop building. Attack the premise: which assumption do both failed cycles share? Report it and ask.
• Blocked on the user. The next check needs consent for an outward action or needs attested evidence. Ask. If the user declines, stop that row and continue the others.

Checkpoint question

After the baseline (and on plateau), report first, then ask:

prompt: The panel scored this TRL N. Should I start raising it toward TRL <target>?
options:
- Yes, raise it (Recommended)
- Change the target level
- Stop here with the score and gap list
- Say this in plain English
allow_multiple: false

Say this in plain English is a meta question. Explain what the next level would need here, with one example check, then ask again. Do not advance. No structured question tool → numbered text fallback.

Reply contract

Playbook: trl
scope: <one line>
overall: TRL N · <provenance>  (baseline TRL M)
target: N · reached | not reached (<stop reason>)
cycles: n
ledger: .balakit/trl/<slug>/ledger.md

Then:

1. Scorecard: one row per in-scope area, with level, provenance, and panel spread.
2. Disagreements the panel logged.
3. Open gaps for the next level, lowest row first.
4. Items waiting on the user (consent, attested evidence).
5. Builder vs panel: each row where the sealed score differed.

Delegate failure

1. Retry one failed assessor once.
2. Never invent a verdict. A cycle with fewer than two verdicts does not move any row. Say so.
3. An assessor that edited tracked files, deployed, or touched an external system has failed. Discard its verdict and revert its edits.