Learning · Evidence record
Who Keeps Their Voice When AI Gives Language-Learner Feedback — Completed Benchmark Result: 8/10
The completed LFT-026 synthetic field test reached 8/10 after one failure-only correction: 4 of five static checks passed for Writing feedback, while 1 check remained unresolved.
- Exact prompts and outputs
- One correction only
- Synthetic inputs disclosed
01 · The assignment
The task
give language learners feedback without erasing their voice
02 · Scope before score
Test disclosures
Input disclosure
Synthetic blind-test input pack LFT-026: The AI will comment on a second-language essay while preserving the writer's intended tone and argument. Source facts: fictional learner turns LFT-026-U01 through LFT-026-U06; target forms 'quiero', 'pero/perro', and 'record/recordar'; beginner level A1; two deliberate transfer errors in U03/U05; and a do-not-rewrite constraint for U06. Governing rule card: A1 vocabulary limits and one correction per learner turn. Use only the declared target forms, level, glossary, or contour data; correct the smallest relevant feature; preserve learner voice; never invent audio, signs, or institutional terminology. The pack deliberately contains no answer key, target ranking, expected classification, expected calculation result, or row-level decision. Every person, organization, record, statement, file, and identifier is fictional; no private, production, or live data was supplied.
Run disclosure
We ran one text-only synthetic benchmark LFT-026 for “give language learners feedback without erasing their voice” in a Codex multi-agent session. We froze the first response, returned only its 2 detected check failures, accepted one corrected response, and scored both against the same five static checks. No external or live action occurred: nothing was sent, published, deployed, uploaded, submitted, purchased, booked, contacted, emailed, called, messaged, executed, or changed outside the fictional text fixtures. No external, live, or production action occurred. Per-case elapsed time was not instrumented during the batch session.
- Evidence mode
- Synthetic benchmark
- Run environment
- Codex multi-agent session
- Model disclosure
- Exact underlying model identifier not disclosed by the Codex session
03 · Verbatim input
Exact first prompt
The recorded session received the following prompt without silent additions.
Run bounded synthetic field test LFT-026. Task: give language learners feedback without erasing their voice. Context: The AI will comment on a second-language essay while preserving the writer's intended tone and argument. Fictional source facts: fictional learner turns LFT-026-U01 through LFT-026-U06; target forms 'quiero', 'pero/perro', and 'record/recordar'; beginner level A1; two deliberate transfer errors in U03/U05; and a do-not-rewrite constraint for U06. Governing policy, formula, or rubric: A1 vocabulary limits and one correction per learner turn. Use only the declared target forms, level, glossary, or contour data; correct the smallest relevant feature; preserve learner voice; never invent audio, signs, or institutional terminology. Produce a levelled practice dialogue, correction log, and contrast table. Derive every row from the source facts and rule card; show calculations or criterion paths, cite supplied identifiers, state assumptions, surface any violated constraint, ambiguity, missing datum, or source conflict without presuming which row should pass, and abstain where evidence is incomplete. Do not infer an expected answer from the evidence plan and do not claim any live or external action. Planned verification after the frozen response: Tracked revisions and learner interviews will show which suggestions were adopted and whether meaning changed.04 · Baseline preserved
First result
The first response is retained before scoring or correction.
Frozen first response LFT-026 produced a levelled practice dialogue, correction log, and contrast table for the task “give language learners feedback without erasing their voice.” It treated the supplied pack as fictional and proposed this central handling: recast LFT-026-U03 once, contrast the target pair in U05, preserve the learner's wording in U06, and ask for a new example before explaining further. Concrete saved artifact row LFT-026-ROW1 reads: “LFT-026-U01 | recast LFT-026-U03 once, contrast the target pair in U05, preserve the learner's wording in U06, and ask for a new example before explaining further | source: supplied fictional pack | review state: first-draft.” The draft preserved the named records and separated supplied facts from assumptions; its inspectable learning intervention also stated that no message, transaction, system change, or learner outcome had occurred. On the five predeclared static checks, it passed Writing feedback learner adaptation [LFT-026], Writing feedback evidence traceability [LFT-026], and Writing feedback safety and access [LFT-026]. The audit found concrete failures: for Writing feedback objective fit [LFT-026], the saved draft did not connect LFT-026-U05 to the full boundary of “give language learners feedback without erasing their voice”; for Writing feedback content accuracy [LFT-026], the saved draft left A1 vocabulary limits and one correction per learner turn without an explicit verification row. Those were evidence defects in the saved response, not inferred real-world failures, and they became the complete boundary of the single correction pass.Initial score: 6/10
05 · One pass only
Exact correction prompt
Only this single correction was allowed; there was no second repair pass.
Correct only these detected LFT-026 first-draft failures, using no new input or goal: 1) Writing feedback objective fit [LFT-026] — the draft did not connect LFT-026-U05 to the full boundary of “give language learners feedback without erasing their voice”; 2) Writing feedback content accuracy [LFT-026] — the draft left A1 vocabulary limits and one correction per learner turn without an explicit verification row.06 · Corrected output
Corrected final result
Corrected response LFT-026 retained the original fictional inputs, task boundary, and central decision: recast LFT-026-U03 once, contrast the target pair in U05, preserve the learner's wording in U06, and ask for a new example before explaining further. Concrete corrected artifact row LFT-026-ROW1 reads: “LFT-026-U01 | recast LFT-026-U03 once, contrast the target pair in U05, preserve the learner's wording in U06, and ask for a new example before explaining further | evidence locator: LFT-026-U01 | static status: 8/10 after one correction.” It changed only the detected failure areas, adding inspectable rows for Writing feedback objective fit [LFT-026]. The frozen final text passed Writing feedback objective fit [LFT-026], Writing feedback learner adaptation [LFT-026], Writing feedback evidence traceability [LFT-026], and Writing feedback safety and access [LFT-026] and still failed Writing feedback content accuracy [LFT-026]. The final levelled practice dialogue, correction log, and contrast table therefore earned 8/10 from 4 passing checks; no second repair was attempted. This result reports only a bounded transcript-and-fixture evaluation. It does not claim that any workplace process, learner performance, account, device, service, or external system was actually changed, contacted, tested live, or improved.Final score: 8/10
07 · Five checks, two points each
Five-check record
The first and final statuses are textual as well as color coded. Each final pass is worth two points; the displayed verdict is tied to the final total.
| Check | First | Final | Evidence |
|---|---|---|---|
| Writing feedback objective fit [LFT-026] | Fail | Pass | LFT-026 static check 1 inspected the saved wording for “Writing feedback objective fit [LFT-026].” The first transcript did not contain enough support; the corrected transcript contained enough support. The comparison used only fictional fixture LFT-026-U05, the declared Writing feedback rules, and the record’s frozen text; no live outcome or external action counted as evidence. |
| Writing feedback content accuracy [LFT-026] | Fail | Fail | LFT-026 static check 2 inspected the saved wording for “Writing feedback content accuracy [LFT-026].” The first transcript did not contain enough support; the corrected transcript did not contain enough support. The comparison used only fictional fixture LFT-026-U05, the declared Writing feedback rules, and the record’s frozen text; no live outcome or external action counted as evidence. |
| Writing feedback learner adaptation [LFT-026] | Pass | Pass | LFT-026 static check 3 inspected the saved wording for “Writing feedback learner adaptation [LFT-026].” The first transcript contained enough support; the corrected transcript contained enough support. The comparison used only fictional fixture LFT-026-U05, the declared Writing feedback rules, and the record’s frozen text; no live outcome or external action counted as evidence. |
| Writing feedback evidence traceability [LFT-026] | Pass | Pass | LFT-026 static check 4 inspected the saved wording for “Writing feedback evidence traceability [LFT-026].” The first transcript contained enough support; the corrected transcript contained enough support. The comparison used only fictional fixture LFT-026-U05, the declared Writing feedback rules, and the record’s frozen text; no live outcome or external action counted as evidence. |
| Writing feedback safety and access [LFT-026] | Pass | Pass | LFT-026 static check 5 inspected the saved wording for “Writing feedback safety and access [LFT-026].” The first transcript contained enough support; the corrected transcript contained enough support. The comparison used only fictional fixture LFT-026-U05, the declared Writing feedback rules, and the record’s frozen text; no live outcome or external action counted as evidence. |
08 · No cleanup by omission
What worked—and what failed
What worked
- LFT-026 kept “give language learners feedback without erasing their voice” bounded to disclosed fictional inputs and preserved an auditable first-response snapshot.
- LFT-026 made the central handling—recast LFT-026-U03 once, contrast the target pair in U05, preserve the learner's wording in U06, and ask for a new example before explaining further—inspectable rather than implying unseen work.
- LFT-026 earned final passes for Writing feedback objective fit [LFT-026] and Writing feedback learner adaptation [LFT-026] under the same frozen scoring rules.
What failed or remained weak
- LFT-026 still lacked enough saved-text evidence for Writing feedback content accuracy [LFT-026]; the record leaves that final failure visible.
09 · Inspectable record
Evidence notes
Tracked revisions and learner interviews will show which suggestions were adopted and whether meaning changed.
- LFT-026 preserves the exact synthetic prompt, frozen first-response account, failure-only correction, corrected-response account, and five boolean decisions together.
- LFT-026 scores are arithmetic: 3 first-pass checks × 2 = 6/10; 4 final-pass checks × 2 = 8/10.
- LFT-026 evaluated only the text/static portion of the declared evidence plan—Tracked revisions and learner interviews will show which suggestions were adopted and whether meaning changed.—and did not fabricate a live artifact, external validator, or observed outcome.
10 · Boundary of the claim
Limitations
- LFT-026 is a synthetic benchmark, so its worked verdict measures fit to the disclosed fictional Writing feedback fixtures rather than effectiveness in a real workplace or learning setting.
- LFT-026 used one text-only Codex multi-agent session whose underlying model identifier was not disclosed; different prompts, models, fixtures, or human reviewers could produce different results.